ruby : Add ruby binding for max_len (#3365)

* add ruby binding for max_len

* add test, update param numbers

.devops/main-musa.Dockerfile

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc4.0.1
+ARG MUSA_VERSION=rc4.2.0
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
 # Target the MUSA runtime image
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 WORKDIR /app
@@ -32,8 +32,9 @@ RUN apt-get update && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/* /tmp/* /var/tmp/*
 
-COPY --from=build /app /app
-RUN du -sh /app/*
-RUN find /app -type f -size +100M
+COPY --from=build /app/build/bin /app/build/bin
+COPY --from=build /app/samples /app/samples
+COPY --from=build /app/models /app/models
+
 ENV PATH=/app/build/bin:$PATH
 ENTRYPOINT [ "bash", "-c" ]

.github/workflows/build.yml

@@ -4,6 +4,27 @@ on:
   push:
     branches:
       - master
+    tags:
+      - 'v*'
+    paths: ['.github/workflows/build.yml',
+            '**/CMakeLists.txt',
+            '**/Makefile',
+            '**/*.mk',
+            '**/*.cmake',
+            '**/*.in',
+            '**/*.h',
+            '**/*.hpp',
+            '**/*.c',
+            '**/*.cpp',
+            '**/*.cu',
+            '**/*.cuh',
+            '**/*.cl',
+            '**/*.swift',
+            '**/*.m',
+            '**/*.mm',
+            '**/*.metal',
+            '**/*.comp',
+            '**/*.java']
   pull_request:
     types: [opened, synchronize, reopened]
   workflow_dispatch:
@@ -41,6 +62,7 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       tag_name: ${{ steps.tag.outputs.name }}
+      should_release: ${{ steps.tag.outputs.should_release }}
 
     steps:
       - name: Checkout with full history
@@ -55,6 +77,7 @@ jobs:
           BUILD_NUMBER=$(git rev-list --count HEAD)
           SHORT_HASH=$(git rev-parse --short=7 HEAD)
           CUSTOM_TAG="${{ github.event.inputs.pre_release_tag }}"
+          SHOULD_RELEASE="false"
 
           echo "Raw values:"
           echo "BUILD_NUMBER: $BUILD_NUMBER"
@@ -62,21 +85,34 @@ jobs:
           echo "BRANCH_NAME: ${{ env.BRANCH_NAME }}"
           echo "CUSTOM_TAG: $CUSTOM_TAG"
 
-          # Use custom tag if provided
-          if [[ -n "$CUSTOM_TAG" ]]; then
+          if [[ "${{ github.ref_type }}" == "tag" ]]; then
+            echo "Using pushed tag name"
+            TAG_NAME="${{ github.ref_name }}"
+            SHOULD_RELEASE="true"
+          elif [[ -n "$CUSTOM_TAG" ]]; then
             echo "Using custom tag"
             TAG_NAME="${CUSTOM_TAG}"
+            SHOULD_RELEASE="true"
+          elif [[ "${{ github.event.inputs.create_release }}" == "true" ]]; then
+            echo "Manual release requested"
+            SHOULD_RELEASE="true"
+            TAG_NAME="b${BUILD_NUMBER}"
           elif [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
             echo "Using master branch format"
             TAG_NAME="b${BUILD_NUMBER}"
+            SHOULD_RELEASE="false"
           else
             echo "Using non-master branch format"
             SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
             TAG_NAME="${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}"
+            SHOULD_RELEASE="false"
           fi
 
           echo "Final tag name: $TAG_NAME"
+          echo "Should release: $SHOULD_RELEASE"
           echo "name=$TAG_NAME" >> $GITHUB_OUTPUT
+          echo "should_release=$SHOULD_RELEASE" >> $GITHUB_OUTPUT
+
 
   ubuntu-22:
     if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
@@ -579,6 +615,7 @@ jobs:
     if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
             github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-latest
+    needs: determine-tag
 
     strategy:
       matrix:
@@ -662,9 +699,7 @@ jobs:
               Compress-Archive -Path "build/bin/${{ matrix.build }}" -DestinationPath "whisper-bin-${{ matrix.arch }}.zip"
 
       - name: Upload binaries
-        if: matrix.sdl2 == 'ON' && ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
-                github.event.inputs.create_release == 'true' ||
-                github.event.inputs.pre_release_tag != '' }}
+        if: matrix.sdl2 == 'ON' && ${{ needs.determine-tag.outputs.should_release }}
         uses: actions/upload-artifact@v4
         with:
           name: whisper-bin-${{ matrix.arch }}.zip
@@ -750,9 +785,7 @@ jobs:
               Compress-Archive -Path "build/bin/${{ matrix.build }}" -DestinationPath "whisper-blas-bin-${{ matrix.arch }}.zip"
 
       - name: Upload binaries
-        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON' && ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
-                github.event.inputs.create_release == 'true' ||
-                github.event.inputs.pre_release_tag != '' }}
+        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON' && ${{ needs.determine-tag.outputs.should_release }}
         uses: actions/upload-artifact@v4
         with:
           name: whisper-blas-bin-${{ matrix.arch }}.zip
@@ -762,6 +795,7 @@ jobs:
     if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
             github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-2022
+    needs: determine-tag
     strategy:
       fail-fast: false
       matrix:
@@ -960,9 +994,7 @@ jobs:
               Compress-Archive -Path "build/bin/${{ matrix.build }}" -DestinationPath "whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}.zip"
 
       - name: Upload binaries
-        if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
-                github.event.inputs.create_release == 'true' ||
-                github.event.inputs.pre_release_tag != '' }}
+        if: ${{ needs.determine-tag.outputs.should_release }}
         uses: actions/upload-artifact@v4
         with:
           name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}.zip
@@ -1039,16 +1071,11 @@ jobs:
 
       - name: Pack artifacts
         id: pack_artifacts
-        if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
-                github.event.inputs.create_release == 'true' ||
-                github.event.inputs.pre_release_tag != '' }}
         run: |
           zip --symlinks -r whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip build-apple/whisper.xcframework
 
       - name: Upload artifacts
-        if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
-                github.event.inputs.create_release == 'true' ||
-                github.event.inputs.pre_release_tag != '' }}
+        if: ${{ needs.determine-tag.outputs.should_release }}
         uses: actions/upload-artifact@v4
         with:
           path: whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip
@@ -1226,7 +1253,7 @@ jobs:
           ./build/bin/quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0
 
   release:
-    if: ${{ github.event.inputs.create_release == 'true' || github.event.inputs.pre_release_tag != '' }}
+    if: ${{ github.event.inputs.create_release == 'true' || github.event.inputs.pre_release_tag != '' || startsWith(github.ref, 'refs/tags/v') }}
 
     runs-on: ubuntu-latest
 
@@ -1269,6 +1296,7 @@ jobs:
         with:
           tag_name: ${{ needs.determine-tag.outputs.tag_name }}
           prerelease: ${{ github.event.inputs.pre_release_tag != '' }}
+          draft: true
 
       - name: Upload release
         id: upload_release
@@ -1295,7 +1323,8 @@ jobs:
   coreml-base-en:
     if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
             github.event.inputs.create_release == 'true' ||
-            github.event.inputs.pre_release_tag != '' }}
+            github.event.inputs.pre_release_tag != '' ||
+            startsWith(github.ref, 'refs/tags/v') }}
     runs-on: macos-latest
     needs: determine-tag
 

.github/workflows/docker.yml

@@ -15,6 +15,7 @@ jobs:
     env:
       COMMIT_SHA: ${{ github.sha }}
     strategy:
+      fail-fast: false
       matrix:
         config:
           - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64" }
@@ -41,21 +42,35 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      - name: Build and push Docker image (versioned)
-        if: github.event_name == 'push'
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platform }}
-          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
-          file: ${{ matrix.config.dockerfile }}
+      - name: Free up disk space
+        run: |
+          sudo apt-get remove -y '^dotnet-.*' '^llvm-.*' '^mysql-.*' '^postgresql-.*'
+          sudo apt-get autoremove -y
+          sudo apt-get autoclean
+
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /opt/hostedtoolcache/CodeQL
+
+          docker system prune -af
+
+          df -h
+
+      - name: Generate tags
+        id: tags
+        run: |
+          TAGS="ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
+          if [ "${{ github.event_name }}" == "push" ]; then
+            TAGS="$TAGS,ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          fi
+          echo "tags=$TAGS" >> $GITHUB_OUTPUT
 
       - name: Build and push Docker image (tagged)
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platform }}
-          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
+          tags: ${{ steps.tags.outputs.tags }}
           file: ${{ matrix.config.dockerfile }}

CMakeLists.txt

@@ -178,6 +178,10 @@ get_directory_property(WHISPER_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
 set_target_properties(whisper PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/whisper.h)
 install(TARGETS whisper LIBRARY PUBLIC_HEADER)
 
+target_compile_definitions(whisper PRIVATE
+    WHISPER_VERSION="${PROJECT_VERSION}"
+)
+
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/whisper-config.cmake.in
         ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake

README.md

@@ -80,7 +80,7 @@ Now build the [whisper-cli](examples/cli) example and transcribe an audio file l
 ```bash
 # build the project
 cmake -B build
-cmake --build build --config Release
+cmake --build build -j --config Release
 
 # transcribe an audio file
 ./build/bin/whisper-cli -f samples/jfk.wav
@@ -149,7 +149,7 @@ standard cmake setup with:
 ```bash
 # build with GGML_BLAS defined
 cmake -B build -DGGML_BLAS=1
-cmake --build build --config Release
+cmake --build build -j --config Release
 ./build/bin/whisper-cli [ .. etc .. ]
 ```
 
@@ -163,7 +163,7 @@ Here are the steps for creating and using a quantized model:
 ```bash
 # quantize a model with Q5_0 method
 cmake -B build
-cmake --build build --config Release
+cmake --build build -j --config Release
 ./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
 
 # run the examples as usual, specifying the quantized model file
@@ -386,7 +386,7 @@ Run the inference examples as usual, for example:
 ## Moore Threads GPU support
 
 With Moore Threads cards the processing of the models is done efficiently on the GPU via muBLAS and custom MUSA kernels.
-First, make sure you have installed `MUSA SDK rc4.0.1`: https://developer.mthreads.com/sdk/download/musa?equipment=&os=&driverVersion=&version=4.0.1
+First, make sure you have installed `MUSA SDK rc4.2.0`: https://developer.mthreads.com/sdk/download/musa?equipment=&os=&driverVersion=&version=4.2.0
 
 Now build `whisper.cpp` with MUSA support:
 
@@ -489,7 +489,7 @@ You will need to have [sdl2](https://wiki.libsdl.org/SDL2/Installation) installe
 
 ```bash
 cmake -B build -DWHISPER_SDL2=ON
-cmake --build build --config Release
+cmake --build build -j --config Release
 ./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 

bindings/go/Makefile

@@ -15,7 +15,7 @@ BUILD_DIR := build_go
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
 INCLUDE_PATH := $(abspath ../../include):$(abspath ../../ggml/include)
-LIBRARY_PATH := $(abspath ../../${BUILD_DIR}/src:$(abspath ../../${BUILD_DIR}/ggml/src))
+LIBRARY_PATH := $(abspath ../../${BUILD_DIR}/src):$(abspath ../../${BUILD_DIR}/ggml/src)
 
 ifeq ($(GGML_CUDA),1)
 	LIBRARY_PATH := $(LIBRARY_PATH):$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib/
@@ -23,7 +23,8 @@ ifeq ($(GGML_CUDA),1)
 endif
 
 ifeq ($(UNAME_S),Darwin)
-	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit
+	LIBRARY_PATH := $(LIBRARY_PATH):$(abspath ../../${BUILD_DIR}/ggml/src/ggml-blas):$(abspath ../../${BUILD_DIR}/ggml/src/ggml-metal)
+	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit -lggml-metal -lggml-blas
 endif
 
 all: clean whisper examples

bindings/go/whisper.go

@@ -9,7 +9,9 @@ import (
 // CGO
 
 /*
-#cgo LDFLAGS: -lwhisper -lggml -lggml-base -lggml-cpu  -lm -lstdc++ -fopenmp
+#cgo LDFLAGS: -lwhisper -lggml -lggml-base -lggml-cpu -lm -lstdc++
+#cgo linux LDFLAGS: -fopenmp
+#cgo darwin LDFLAGS: -lggml-metal -lggml-blas
 #cgo darwin LDFLAGS: -framework Accelerate -framework Metal -framework Foundation -framework CoreGraphics
 #include <whisper.h>
 #include <stdlib.h>

bindings/ruby/ext/options.rb

@@ -64,7 +64,7 @@ class Options
   def configure_coreml
     if enabled?("WHISPER_COREML")
       $LDFLAGS << " -framework Foundation -framework CoreML"
-      $CPPFLAGS << " -DRUBY_WHISPER_USE_COREML"
+      $defs << "-DRUBY_WHISPER_USE_COREML"
     end
   end
 
@@ -73,10 +73,13 @@ class Options
   end
 
   def enabled?(option)
-    if @options[option][1].nil?
+    op = @options[option]
+    raise "Option not exist: #{option}" unless op
+    raise "Option not boolean: #{option}(#{op[0]})" unless op[0] == "BOOL"
+    if op[1].nil?
       cmake_options[option][1]
     else
-      @options[option][1]
+      op[1]
     end
   end
 end

bindings/ruby/ext/ruby_whisper.c

@@ -148,6 +148,7 @@ void Init_whisper() {
   mWhisper = rb_define_module("Whisper");
   mVAD = rb_define_module_under(mWhisper, "VAD");
 
+  rb_define_const(mWhisper, "VERSION", rb_str_new2(whisper_version()));
   rb_define_const(mWhisper, "LOG_LEVEL_NONE", INT2NUM(GGML_LOG_LEVEL_NONE));
   rb_define_const(mWhisper, "LOG_LEVEL_INFO", INT2NUM(GGML_LOG_LEVEL_INFO));
   rb_define_const(mWhisper, "LOG_LEVEL_WARN", INT2NUM(GGML_LOG_LEVEL_WARN));

bindings/ruby/ext/ruby_whisper_params.c

@@ -26,7 +26,7 @@
   rb_define_method(cParams, #param_name, ruby_whisper_params_get_ ## param_name, 0); \
   rb_define_method(cParams, #param_name "=", ruby_whisper_params_set_ ## param_name, 1);
 
-#define RUBY_WHISPER_PARAMS_PARAM_NAMES_COUNT 35
+#define RUBY_WHISPER_PARAMS_PARAM_NAMES_COUNT 36
 
 extern VALUE cParams;
 extern VALUE cVADParams;
@@ -49,6 +49,7 @@ static ID id_print_timestamps;
 static ID id_suppress_blank;
 static ID id_suppress_nst;
 static ID id_token_timestamps;
+static ID id_max_len;
 static ID id_split_on_word;
 static ID id_initial_prompt;
 static ID id_diarize;
@@ -514,6 +515,33 @@ ruby_whisper_params_set_token_timestamps(VALUE self, VALUE value)
 {
   BOOL_PARAMS_SETTER(self, token_timestamps, value)
 }
+
+/*
+ * max segment length in characters.
+ *
+ * call-seq:
+ *   max_len -> Integer
+ */
+static VALUE
+ruby_whisper_params_get_max_len(VALUE self)
+{
+  ruby_whisper_params *rwp;
+  TypedData_Get_Struct(self, ruby_whisper_params, &ruby_whisper_params_type, rwp);
+  return INT2NUM(rwp->params.max_len);
+}
+/*
+ * call-seq:
+ *   max_len = length -> length
+ */
+static VALUE
+ruby_whisper_params_set_max_len(VALUE self, VALUE value)
+{
+  ruby_whisper_params *rwp;
+  TypedData_Get_Struct(self, ruby_whisper_params, &ruby_whisper_params_type, rwp);
+  rwp->params.max_len = NUM2INT(value);
+  return value;
+}
+
 /*
  * If true, split on word rather than on token (when used with max_len).
  *
@@ -1137,6 +1165,7 @@ ruby_whisper_params_initialize(int argc, VALUE *argv, VALUE self)
       SET_PARAM_IF_SAME(suppress_blank)
       SET_PARAM_IF_SAME(suppress_nst)
       SET_PARAM_IF_SAME(token_timestamps)
+      SET_PARAM_IF_SAME(max_len)
       SET_PARAM_IF_SAME(split_on_word)
       SET_PARAM_IF_SAME(initial_prompt)
       SET_PARAM_IF_SAME(offset)
@@ -1271,30 +1300,31 @@ init_ruby_whisper_params(VALUE *mWhisper)
   DEFINE_PARAM(suppress_blank, 8)
   DEFINE_PARAM(suppress_nst, 9)
   DEFINE_PARAM(token_timestamps, 10)
-  DEFINE_PARAM(split_on_word, 11)
-  DEFINE_PARAM(initial_prompt, 12)
-  DEFINE_PARAM(diarize, 13)
-  DEFINE_PARAM(offset, 14)
-  DEFINE_PARAM(duration, 15)
-  DEFINE_PARAM(max_text_tokens, 16)
-  DEFINE_PARAM(temperature, 17)
-  DEFINE_PARAM(max_initial_ts, 18)
-  DEFINE_PARAM(length_penalty, 19)
-  DEFINE_PARAM(temperature_inc, 20)
-  DEFINE_PARAM(entropy_thold, 21)
-  DEFINE_PARAM(logprob_thold, 22)
-  DEFINE_PARAM(no_speech_thold, 23)
-  DEFINE_PARAM(new_segment_callback, 24)
-  DEFINE_PARAM(new_segment_callback_user_data, 25)
-  DEFINE_PARAM(progress_callback, 26)
-  DEFINE_PARAM(progress_callback_user_data, 27)
-  DEFINE_PARAM(encoder_begin_callback, 28)
-  DEFINE_PARAM(encoder_begin_callback_user_data, 29)
-  DEFINE_PARAM(abort_callback, 30)
-  DEFINE_PARAM(abort_callback_user_data, 31)
-  DEFINE_PARAM(vad, 32)
-  DEFINE_PARAM(vad_model_path, 33)
-  DEFINE_PARAM(vad_params, 34)
+  DEFINE_PARAM(max_len, 11)
+  DEFINE_PARAM(split_on_word, 12)
+  DEFINE_PARAM(initial_prompt, 13)
+  DEFINE_PARAM(diarize, 14)
+  DEFINE_PARAM(offset, 15)
+  DEFINE_PARAM(duration, 16)
+  DEFINE_PARAM(max_text_tokens, 17)
+  DEFINE_PARAM(temperature, 18)
+  DEFINE_PARAM(max_initial_ts, 19)
+  DEFINE_PARAM(length_penalty, 20)
+  DEFINE_PARAM(temperature_inc, 21)
+  DEFINE_PARAM(entropy_thold, 22)
+  DEFINE_PARAM(logprob_thold, 23)
+  DEFINE_PARAM(no_speech_thold, 24)
+  DEFINE_PARAM(new_segment_callback, 25)
+  DEFINE_PARAM(new_segment_callback_user_data, 26)
+  DEFINE_PARAM(progress_callback, 27)
+  DEFINE_PARAM(progress_callback_user_data, 28)
+  DEFINE_PARAM(encoder_begin_callback, 29)
+  DEFINE_PARAM(encoder_begin_callback_user_data, 30)
+  DEFINE_PARAM(abort_callback, 31)
+  DEFINE_PARAM(abort_callback_user_data, 32)
+  DEFINE_PARAM(vad, 33)
+  DEFINE_PARAM(vad_model_path, 34)
+  DEFINE_PARAM(vad_params, 35)
 
   rb_define_method(cParams, "on_new_segment", ruby_whisper_params_on_new_segment, 0);
   rb_define_method(cParams, "on_progress", ruby_whisper_params_on_progress, 0);

bindings/ruby/sig/whisper.rbs

@@ -10,6 +10,7 @@ module Whisper
   type encoder_begin_callback = ^(Whisper::Context, void, Object user_data) -> void
   type abort_callback = ^(Whisper::Context, void, Object user_data) -> boolish
 
+  VERSION: String
   LOG_LEVEL_NONE: Integer
   LOG_LEVEL_INFO: Integer
   LOG_LEVEL_WARN: Integer
@@ -134,6 +135,7 @@ module Whisper
       ?suppress_blank: boolish,
       ?suppress_nst: boolish,
       ?token_timestamps: boolish,
+      ?max_len: Integer,
       ?split_on_word: boolish,
       ?initial_prompt: string | nil,
       ?diarize: boolish,
@@ -221,6 +223,12 @@ module Whisper
     #
     def token_timestamps: () -> (true | false)
 
+    def max_len=: (Integer) -> Integer
+
+    # max segment length in characters.
+    #
+    def max_len: () -> Integer
+
     def split_on_word=: (boolish) -> boolish
 
     # If true, split on word rather than on token (when used with max_len).

bindings/ruby/test/test_package.rb

@@ -31,10 +31,11 @@ class TestPackage < TestBase
         Dir.mktmpdir do |dir|
           system "gem", "install", "--install-dir", dir.shellescape, "--no-document", "pkg/#{gemspec.file_name.shellescape}", "--", "--enable-whisper-coreml", exception: true
           assert_installed dir, gemspec.version
+          libdir = File.join(dir, "gems", "#{gemspec.name}-#{gemspec.version}", "lib")
           assert_nothing_raised do
-            libdir = File.join(dir, "gems", "#{gemspec.name}-#{gemspec.version}", "lib")
             system "ruby", "-I", libdir, "-r", "whisper", "-e", "Whisper::Context.new('tiny')", exception: true
           end
+          assert_match(/COREML = 1/, `ruby -I #{libdir.shellescape} -r whisper -e 'puts Whisper.system_info_str'`)
         end
       end
     end

bindings/ruby/test/test_params.rb

@@ -13,6 +13,7 @@ class TestParams < TestBase
     :suppress_blank,
     :suppress_nst,
     :token_timestamps,
+    :max_len,
     :split_on_word,
     :initial_prompt,
     :diarize,
@@ -139,6 +140,13 @@ class TestParams < TestBase
     assert !@params.token_timestamps
   end
 
+  def test_max_len
+    @params.max_len = 42
+    assert_equal @params.max_len, 42
+    @params.max_len = 0
+    assert_equal @params.max_len, 0
+  end
+
   def test_split_on_word
     @params.split_on_word = true
     assert @params.split_on_word

bindings/ruby/test/test_segment.rb

@@ -72,6 +72,16 @@ class TestSegment < TestBase
     whisper.transcribe(AUDIO, params)
   end
 
+  def test_transcription_after_segment_retrieved
+    params = Whisper::Params.new
+    segment = whisper.each_segment.first
+    assert_match(/ask not what your country can do for you, ask what you can do for your country/, segment.text)
+
+    whisper.transcribe(AUDIO, Whisper::Params.new(offset: 5000))
+    assert_not_match(/ask not what your country can do for you, ask what you can do for your country/, segment.text)
+    assert_match(/what you can do for your country/i, segment.text)
+  end
+
   def test_pattern_matching
     segment = whisper.each_segment.first
     segment => {start_time:, end_time:, text:, no_speech_prob:, speaker_turn_next:}

bindings/ruby/test/test_whisper.rb

@@ -116,6 +116,10 @@ class TestWhisper < TestBase
     assert_match(/\AWHISPER : COREML = \d | OPENVINO = \d |/, Whisper.system_info_str)
   end
 
+  def test_version
+    assert_kind_of String, Whisper::VERSION
+  end
+
   def test_log_set
     user_data = Object.new
     logs = []

build-xcframework.sh

@@ -15,6 +15,7 @@ GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
 GGML_METAL_USE_BF16=ON
 GGML_OPENMP=OFF
+BUILD_STATIC_XCFRAMEWORK=${BUILD_STATIC_XCFRAMEWORK:-OFF}
 
 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
 COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
@@ -327,6 +328,15 @@ combine_static_libraries() {
         arch_flags+=" -arch $arch"
     done
 
+
+    if [[ "${BUILD_STATIC_XCFRAMEWORK}" == "ON" ]]; then
+        echo "Packaging static framework for ${platform}."
+        mkdir -p "$(dirname "${base_dir}/${output_lib}")"
+        cp "${temp_dir}/combined.a" "${base_dir}/${output_lib}"
+        rm -rf "${temp_dir}"
+        return
+    fi
+
     # Create dynamic library
     echo "Creating dynamic library for ${platform}."
     xcrun -sdk $sdk clang++ -dynamiclib \
@@ -529,6 +539,20 @@ combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
+
+if [[ "${BUILD_STATIC_XCFRAMEWORK}" == "ON" ]]; then
+    xcodebuild -create-xcframework \
+        -framework $(pwd)/build-ios-sim/framework/whisper.framework \
+        -framework $(pwd)/build-ios-device/framework/whisper.framework \
+        -framework $(pwd)/build-macos/framework/whisper.framework \
+        -framework $(pwd)/build-visionos/framework/whisper.framework \
+        -framework $(pwd)/build-visionos-sim/framework/whisper.framework \
+        -framework $(pwd)/build-tvos-device/framework/whisper.framework \
+        -framework $(pwd)/build-tvos-sim/framework/whisper.framework \
+        -output $(pwd)/build-apple/whisper.xcframework
+    exit 0
+fi
+
 xcodebuild -create-xcframework \
     -framework $(pwd)/build-ios-sim/framework/whisper.framework \
     -debug-symbols $(pwd)/build-ios-sim/dSYMs/whisper.dSYM \

examples/addon.node/README.md

@@ -1,8 +1,10 @@
-# addon
+# whisper.cpp Node.js addon
 
 This is an addon demo that can **perform whisper model reasoning in `node` and `electron` environments**, based on [cmake-js](https://github.com/cmake-js/cmake-js).
 It can be used as a reference for using the whisper.cpp project in other node projects.
 
+This addon now supports **Voice Activity Detection (VAD)** for improved transcription performance.
+
 ## Install
 
 ```shell
@@ -26,12 +28,88 @@ For Electron addon and cmake-js options, you can see [cmake-js](https://github.c
 
 ## Run
 
+### Basic Usage
+
 ```shell
 cd examples/addon.node
 
 node index.js --language='language' --model='model-path' --fname_inp='file-path'
 ```
 
-Because this is a simple Demo, only the above parameters are set in the node environment.
+### VAD (Voice Activity Detection) Usage
 
-Other parameters can also be specified in the node environment.
+Run the VAD example with performance comparison:
+
+```shell
+node vad-example.js
+```
+
+## Voice Activity Detection (VAD) Support
+
+VAD can significantly improve transcription performance by only processing speech segments, which is especially beneficial for audio files with long periods of silence.
+
+### VAD Model Setup
+
+Before using VAD, download a VAD model:
+
+```shell
+# From the whisper.cpp root directory
+./models/download-vad-model.sh silero-v5.1.2
+```
+
+### VAD Parameters
+
+All VAD parameters are optional and have sensible defaults:
+
+- `vad`: Enable VAD (default: false)
+- `vad_model`: Path to VAD model file (required when VAD enabled)
+- `vad_threshold`: Speech detection threshold 0.0-1.0 (default: 0.5)
+- `vad_min_speech_duration_ms`: Min speech duration in ms (default: 250)
+- `vad_min_silence_duration_ms`: Min silence duration in ms (default: 100)
+- `vad_max_speech_duration_s`: Max speech duration in seconds (default: FLT_MAX)
+- `vad_speech_pad_ms`: Speech padding in ms (default: 30)
+- `vad_samples_overlap`: Sample overlap 0.0-1.0 (default: 0.1)
+
+### JavaScript API Example
+
+```javascript
+const path = require("path");
+const { whisper } = require(path.join(__dirname, "../../build/Release/addon.node"));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
+
+// With VAD enabled
+const vadParams = {
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  vad: true,
+  vad_model: path.join(__dirname, "../../models/ggml-silero-v5.1.2.bin"),
+  vad_threshold: 0.5,
+  progress_callback: (progress) => console.log(`Progress: ${progress}%`)
+};
+
+whisperAsync(vadParams).then(result => console.log(result));
+```
+
+## Supported Parameters
+
+Both traditional whisper.cpp parameters and new VAD parameters are supported:
+
+- `language`: Language code (e.g., "en", "es", "fr")
+- `model`: Path to whisper model file
+- `fname_inp`: Path to input audio file
+- `use_gpu`: Enable GPU acceleration (default: true)
+- `flash_attn`: Enable flash attention (default: false)
+- `no_prints`: Disable console output (default: false)
+- `no_timestamps`: Disable timestamps (default: false)
+- `detect_language`: Auto-detect language (default: false)
+- `audio_ctx`: Audio context size (default: 0)
+- `max_len`: Maximum segment length (default: 0)
+- `max_context`: Maximum context size (default: -1)
+- `prompt`: Initial prompt for decoder
+- `comma_in_time`: Use comma in timestamps (default: true)
+- `print_progress`: Print progress info (default: false)
+- `progress_callback`: Progress callback function
+- VAD parameters (see above section)

examples/addon.node/__test__/whisper.spec.js

@@ -1,39 +1,133 @@
-const path = require("path");
-const { whisper } = require(path.join(
-  __dirname,
-  "../../../build/Release/addon.node"
-));
-const { promisify } = require("util");
+const { join } = require('path');
+const { whisper } = require('../../../build/Release/addon.node');
+const { promisify } = require('util');
 
 const whisperAsync = promisify(whisper);
 
-const whisperParamsMock = {
-  language: "en",
-  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
-  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+const commonParams = {
+  language: 'en',
+  model: join(__dirname, '../../../models/ggml-base.en.bin'),
+  fname_inp: join(__dirname, '../../../samples/jfk.wav'),
   use_gpu: true,
   flash_attn: false,
   no_prints: true,
-  comma_in_time: false,
-  translate: true,
   no_timestamps: false,
   detect_language: false,
   audio_ctx: 0,
-  max_len: 0,
-  prompt: "",
-  print_progress: false,
-  progress_callback: (progress) => {
-    console.log(`Progress: ${progress}`);
-  },
-  max_context: -1
+  max_len: 0
 };
 
-describe("Run whisper.node", () => {
-    test("it should receive a non-empty value", async () => {
-        let result = await whisperAsync(whisperParamsMock);
-      console.log(result);
+describe('Whisper.cpp Node.js addon with VAD support', () => {
+  test('Basic whisper transcription without VAD', async () => {
+    const params = {
+      ...commonParams,
+      vad: false
+    };
 
-        expect(result['transcription'].length).toBeGreaterThan(0);
-    }, 10000);
+    const result = await whisperAsync(params);
+    
+    expect(typeof result).toBe('object');
+    expect(Array.isArray(result.transcription)).toBe(true);
+    expect(result.transcription.length).toBeGreaterThan(0);
+    
+    // Check that we got some transcription text
+    const text = result.transcription.map(segment => segment[2]).join(' ');
+    expect(text.length).toBeGreaterThan(0);
+    expect(text.toLowerCase()).toContain('ask not');
+  }, 30000);
+
+  test('VAD parameters validation', async () => {
+    // Test with invalid VAD model - should return empty transcription
+    const invalidParams = {
+      ...commonParams,
+      vad: true,
+      vad_model: 'non-existent-model.bin',
+      vad_threshold: 0.5
+    };
+
+    // This should handle the error gracefully and return empty transcription
+    const result = await whisperAsync(invalidParams);
+    expect(typeof result).toBe('object');
+    expect(Array.isArray(result.transcription)).toBe(true);
+    // When VAD model doesn't exist, it should return empty transcription
+    expect(result.transcription.length).toBe(0);
+  }, 10000);
+
+  test('VAD parameter parsing', async () => {
+    // Test that VAD parameters are properly parsed (even if VAD model doesn't exist)
+    const vadParams = {
+      ...commonParams,
+      vad: false, // Disabled so no model required
+      vad_threshold: 0.7,
+      vad_min_speech_duration_ms: 300,
+      vad_min_silence_duration_ms: 150,
+      vad_max_speech_duration_s: 45.0,
+      vad_speech_pad_ms: 50,
+      vad_samples_overlap: 0.15
+    };
+
+    const result = await whisperAsync(vadParams);
+    
+    expect(typeof result).toBe('object');
+    expect(Array.isArray(result.transcription)).toBe(true);
+  }, 30000);
+
+  test('Progress callback with VAD disabled', async () => {
+    let progressCalled = false;
+    let lastProgress = 0;
+
+    const params = {
+      ...commonParams,
+      vad: false,
+      progress_callback: (progress) => {
+        progressCalled = true;
+        lastProgress = progress;
+        expect(progress).toBeGreaterThanOrEqual(0);
+        expect(progress).toBeLessThanOrEqual(100);
+      }
+    };
+
+    const result = await whisperAsync(params);
+    
+    expect(progressCalled).toBe(true);
+    expect(lastProgress).toBe(100);
+    expect(typeof result).toBe('object');
+  }, 30000);
+
+  test('Language detection without VAD', async () => {
+    const params = {
+      ...commonParams,
+      vad: false,
+      detect_language: true,
+      language: 'auto'
+    };
+
+    const result = await whisperAsync(params);
+    
+    expect(typeof result).toBe('object');
+    expect(typeof result.language).toBe('string');
+    expect(result.language.length).toBeGreaterThan(0);
+  }, 30000);
+
+  test('Basic transcription with all VAD parameters set', async () => {
+    // Test with VAD disabled but all parameters set to ensure no crashes
+    const params = {
+      ...commonParams,
+      vad: false, // Disabled so it works without VAD model
+      vad_model: '', // Empty model path
+      vad_threshold: 0.6,
+      vad_min_speech_duration_ms: 200,
+      vad_min_silence_duration_ms: 80,
+      vad_max_speech_duration_s: 25.0,
+      vad_speech_pad_ms: 40,
+      vad_samples_overlap: 0.08
+    };
+
+    const result = await whisperAsync(params);
+    
+    expect(typeof result).toBe('object');
+    expect(Array.isArray(result.transcription)).toBe(true);
+    expect(result.transcription.length).toBeGreaterThan(0);
+  }, 30000);
 });
 

examples/addon.node/addon.cpp

@@ -9,6 +9,7 @@
 #include <vector>
 #include <cmath>
 #include <cstdint>
+#include <cfloat>
 
 struct whisper_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -51,6 +52,16 @@ struct whisper_params {
     std::vector<std::string> fname_out = {};
 
     std::vector<float> pcmf32 = {}; // mono-channel F32 PCM
+
+    // Voice Activity Detection (VAD) parameters
+    bool        vad           = false;
+    std::string vad_model     = "";
+    float       vad_threshold = 0.5f;
+    int         vad_min_speech_duration_ms = 250;
+    int         vad_min_silence_duration_ms = 100;
+    float       vad_max_speech_duration_s = FLT_MAX;
+    int         vad_speech_pad_ms = 30;
+    float       vad_samples_overlap = 0.1f;
 };
 
 struct whisper_print_user_data {
@@ -333,16 +344,16 @@ class ProgressWorker : public Napi::AsyncWorker {
                 };
                 wparams.progress_callback_user_data = this;
 
-                // Abort mechanism example
-                {
-                    static bool is_aborted = false; // Note: this should be atomic to avoid data races
+                // Set VAD parameters
+                wparams.vad            = params.vad;
+                wparams.vad_model_path = params.vad_model.c_str();
 
-                    wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
-                        bool is_aborted = *(bool*)user_data;
-                        return !is_aborted;
-                    };
-                    wparams.encoder_begin_callback_user_data = &is_aborted;
-                }
+                wparams.vad_params.threshold               = params.vad_threshold;
+                wparams.vad_params.min_speech_duration_ms  = params.vad_min_speech_duration_ms;
+                wparams.vad_params.min_silence_duration_ms = params.vad_min_silence_duration_ms;
+                wparams.vad_params.max_speech_duration_s   = params.vad_max_speech_duration_s;
+                wparams.vad_params.speech_pad_ms           = params.vad_speech_pad_ms;
+                wparams.vad_params.samples_overlap         = params.vad_samples_overlap;
 
                 if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
                     fprintf(stderr, "failed to process audio\n");
@@ -385,14 +396,46 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   std::string language = whisper_params.Get("language").As<Napi::String>();
   std::string model = whisper_params.Get("model").As<Napi::String>();
   std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
-  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
-  bool flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
-  bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
-  bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
-  bool detect_language = whisper_params.Get("detect_language").As<Napi::Boolean>();
-  int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
-  bool comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
-  int32_t max_len = whisper_params.Get("max_len").As<Napi::Number>();
+  
+  bool use_gpu = true;
+  if (whisper_params.Has("use_gpu") && whisper_params.Get("use_gpu").IsBoolean()) {
+    use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
+  }
+  
+  bool flash_attn = false;
+  if (whisper_params.Has("flash_attn") && whisper_params.Get("flash_attn").IsBoolean()) {
+    flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
+  }
+  
+  bool no_prints = false;
+  if (whisper_params.Has("no_prints") && whisper_params.Get("no_prints").IsBoolean()) {
+    no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
+  }
+  
+  bool no_timestamps = false;
+  if (whisper_params.Has("no_timestamps") && whisper_params.Get("no_timestamps").IsBoolean()) {
+    no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
+  }
+  
+  bool detect_language = false;
+  if (whisper_params.Has("detect_language") && whisper_params.Get("detect_language").IsBoolean()) {
+    detect_language = whisper_params.Get("detect_language").As<Napi::Boolean>();
+  }
+  
+  int32_t audio_ctx = 0;
+  if (whisper_params.Has("audio_ctx") && whisper_params.Get("audio_ctx").IsNumber()) {
+    audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
+  }
+  
+  bool comma_in_time = true;
+  if (whisper_params.Has("comma_in_time") && whisper_params.Get("comma_in_time").IsBoolean()) {
+    comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
+  }
+  
+  int32_t max_len = 0;
+  if (whisper_params.Has("max_len") && whisper_params.Get("max_len").IsNumber()) {
+    max_len = whisper_params.Get("max_len").As<Napi::Number>();
+  }
   
   // Add support for max_context
   int32_t max_context = -1;
@@ -408,7 +451,7 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   
   // Add support for print_progress
   bool print_progress = false;
-  if (whisper_params.Has("print_progress")) {
+  if (whisper_params.Has("print_progress") && whisper_params.Get("print_progress").IsBoolean()) {
     print_progress = whisper_params.Get("print_progress").As<Napi::Boolean>();
   }
   // Add support for progress_callback
@@ -417,6 +460,47 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
     progress_callback = whisper_params.Get("progress_callback").As<Napi::Function>();
   }
 
+  // Add support for VAD parameters
+  bool vad = false;
+  if (whisper_params.Has("vad") && whisper_params.Get("vad").IsBoolean()) {
+    vad = whisper_params.Get("vad").As<Napi::Boolean>();
+  }
+  
+  std::string vad_model = "";
+  if (whisper_params.Has("vad_model") && whisper_params.Get("vad_model").IsString()) {
+    vad_model = whisper_params.Get("vad_model").As<Napi::String>();
+  }
+  
+  float vad_threshold = 0.5f;
+  if (whisper_params.Has("vad_threshold") && whisper_params.Get("vad_threshold").IsNumber()) {
+    vad_threshold = whisper_params.Get("vad_threshold").As<Napi::Number>();
+  }
+  
+  int vad_min_speech_duration_ms = 250;
+  if (whisper_params.Has("vad_min_speech_duration_ms") && whisper_params.Get("vad_min_speech_duration_ms").IsNumber()) {
+    vad_min_speech_duration_ms = whisper_params.Get("vad_min_speech_duration_ms").As<Napi::Number>();
+  }
+  
+  int vad_min_silence_duration_ms = 100;
+  if (whisper_params.Has("vad_min_silence_duration_ms") && whisper_params.Get("vad_min_silence_duration_ms").IsNumber()) {
+    vad_min_silence_duration_ms = whisper_params.Get("vad_min_silence_duration_ms").As<Napi::Number>();
+  }
+  
+  float vad_max_speech_duration_s = FLT_MAX;
+  if (whisper_params.Has("vad_max_speech_duration_s") && whisper_params.Get("vad_max_speech_duration_s").IsNumber()) {
+    vad_max_speech_duration_s = whisper_params.Get("vad_max_speech_duration_s").As<Napi::Number>();
+  }
+  
+  int vad_speech_pad_ms = 30;
+  if (whisper_params.Has("vad_speech_pad_ms") && whisper_params.Get("vad_speech_pad_ms").IsNumber()) {
+    vad_speech_pad_ms = whisper_params.Get("vad_speech_pad_ms").As<Napi::Number>();
+  }
+  
+  float vad_samples_overlap = 0.1f;
+  if (whisper_params.Has("vad_samples_overlap") && whisper_params.Get("vad_samples_overlap").IsNumber()) {
+    vad_samples_overlap = whisper_params.Get("vad_samples_overlap").As<Napi::Number>();
+  }
+
   Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
   std::vector<float> pcmf32_vec;
   if (pcmf32Value.IsTypedArray()) {
@@ -444,6 +528,16 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   params.prompt = prompt;
   params.detect_language = detect_language;
 
+  // Set VAD parameters
+  params.vad = vad;
+  params.vad_model = vad_model;
+  params.vad_threshold = vad_threshold;
+  params.vad_min_speech_duration_ms = vad_min_speech_duration_ms;
+  params.vad_min_silence_duration_ms = vad_min_silence_duration_ms;
+  params.vad_max_speech_duration_s = vad_max_speech_duration_s;
+  params.vad_speech_pad_ms = vad_speech_pad_ms;
+  params.vad_samples_overlap = vad_samples_overlap;
+
   Napi::Function callback = info[1].As<Napi::Function>();
   // Create a new Worker class with progress callback support
   ProgressWorker* worker = new ProgressWorker(callback, params, progress_callback, env);

examples/addon.node/vad-example.js

@@ -0,0 +1,132 @@
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../build/Release/addon.node"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
+
+// Example with VAD enabled
+const vadParams = {
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  use_gpu: true,
+  flash_attn: false,
+  no_prints: false,
+  comma_in_time: true,
+  translate: false,
+  no_timestamps: false,
+  detect_language: false,
+  audio_ctx: 0,
+  max_len: 0,
+  // VAD parameters
+  vad: true,
+  vad_model: path.join(__dirname, "../../models/ggml-silero-v5.1.2.bin"), // You need to download this model
+  vad_threshold: 0.5,
+  vad_min_speech_duration_ms: 250,
+  vad_min_silence_duration_ms: 100,
+  vad_max_speech_duration_s: 30.0,
+  vad_speech_pad_ms: 30,
+  vad_samples_overlap: 0.1,
+  progress_callback: (progress) => {
+    console.log(`VAD Transcription progress: ${progress}%`);
+  }
+};
+
+// Example without VAD (traditional approach)
+const traditionalParams = {
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  use_gpu: true,
+  flash_attn: false,
+  no_prints: false,
+  comma_in_time: true,
+  translate: false,
+  no_timestamps: false,
+  detect_language: false,
+  audio_ctx: 0,
+  max_len: 0,
+  vad: false, // Explicitly disable VAD
+  progress_callback: (progress) => {
+    console.log(`Traditional transcription progress: ${progress}%`);
+  }
+};
+
+async function runVADExample() {
+  try {
+    console.log("=== Whisper.cpp Node.js VAD Example ===\n");
+    
+    // Check if VAD model exists
+    const fs = require('fs');
+    if (!fs.existsSync(vadParams.vad_model)) {
+      console.log("⚠️  VAD model not found. Please download the VAD model first:");
+      console.log("   ./models/download-vad-model.sh silero-v5.1.2");
+      console.log("   Or run: python models/convert-silero-vad-to-ggml.py");
+      console.log("\n   Falling back to traditional transcription without VAD...\n");
+      
+      // Run without VAD
+      console.log("🎵 Running traditional transcription...");
+      const traditionalResult = await whisperAsync(traditionalParams);
+      console.log("\n📝 Traditional transcription result:");
+      console.log(traditionalResult);
+      return;
+    }
+    
+    console.log("🎵 Running transcription with VAD enabled...");
+    console.log("VAD Parameters:");
+    console.log(`  - Threshold: ${vadParams.vad_threshold}`);
+    console.log(`  - Min speech duration: ${vadParams.vad_min_speech_duration_ms}ms`);
+    console.log(`  - Min silence duration: ${vadParams.vad_min_silence_duration_ms}ms`);
+    console.log(`  - Max speech duration: ${vadParams.vad_max_speech_duration_s}s`);
+    console.log(`  - Speech padding: ${vadParams.vad_speech_pad_ms}ms`);
+    console.log(`  - Samples overlap: ${vadParams.vad_samples_overlap}\n`);
+    
+    const startTime = Date.now();
+    const vadResult = await whisperAsync(vadParams);
+    const vadDuration = Date.now() - startTime;
+    
+    console.log("\n✅ VAD transcription completed!");
+    console.log(`⏱️  Processing time: ${vadDuration}ms`);
+    console.log("\n📝 VAD transcription result:");
+    console.log(vadResult);
+    
+    // Compare with traditional approach
+    console.log("\n🔄 Running traditional transcription for comparison...");
+    const traditionalStartTime = Date.now();
+    const traditionalResult = await whisperAsync(traditionalParams);
+    const traditionalDuration = Date.now() - traditionalStartTime;
+    
+    console.log("\n✅ Traditional transcription completed!");
+    console.log(`⏱️  Processing time: ${traditionalDuration}ms`);
+    console.log("\n📝 Traditional transcription result:");
+    console.log(traditionalResult);
+    
+    // Performance comparison
+    console.log("\n📊 Performance Comparison:");
+    console.log(`VAD:         ${vadDuration}ms`);
+    console.log(`Traditional: ${traditionalDuration}ms`);
+    const speedup = traditionalDuration / vadDuration;
+    if (speedup > 1) {
+      console.log(`🚀 VAD is ${speedup.toFixed(2)}x faster!`);
+    } else {
+      console.log(`ℹ️  Traditional approach was ${(1/speedup).toFixed(2)}x faster in this case.`);
+    }
+    
+  } catch (error) {
+    console.error("❌ Error during transcription:", error);
+  }
+}
+
+// Run the example
+if (require.main === module) {
+  runVADExample();
+}
+
+module.exports = {
+  runVADExample,
+  vadParams,
+  traditionalParams
+}; 

examples/bench.wasm/README.md

@@ -2,7 +2,7 @@
 
 Benchmark the performance of whisper.cpp in the browser using WebAssembly
 
-Link: https://ggerganov.github.io/whisper.cpp/bench.wasm
+Link: https://ggml.ai/whisper.cpp/bench.wasm/
 
 Terminal version: [examples/bench](/examples/bench)
 
@@ -32,6 +32,16 @@ cp bin/libbench.js        /path/to/html/
 cp bin/libbench.worker.js /path/to/html/
 ```
 
+> 📝 **Note:** By default this example is built with `WHISPER_WASM_SINGLE_FILE=ON`
+> which means that that a separate .wasm file will not be generated. Instead, the
+> WASM module is embedded in the main JS file as a base64 encoded string. To
+> generate a separate .wasm file, you need to disable this option by passing
+> `-DWHISPER_WASM_SINGLE_FILE=OFF`:
+> ```console
+> emcmake cmake .. -DWHISPER_WASM_SINGLE_FILE=OFF
+> ```
+> This will generate a `libbench.wasm` file in the build/bin directory.
+
 > 📝 **Note:** As of Emscripten 3.1.58 (April 2024), separate worker.js files are no
 > longer generated and the worker is embedded in the main JS file. So the worker
 > file will not be geneated for versions later than `3.1.58`.

examples/command.wasm/README.md

@@ -3,7 +3,7 @@
 This is a basic Voice Assistant example that accepts voice commands from the microphone.
 It runs in fully in the browser via WebAseembly.
 
-Online demo: https://ggerganov.github.io/whisper.cpp/command.wasm
+Online demo: https://ggml.ai/whisper.cpp/command.wasm/
 
 Terminal version: [examples/command](/examples/command)
 
@@ -32,6 +32,16 @@ cp bin/libcommand.js        /path/to/html/
 cp bin/libcommand.worker.js /path/to/html/
 ```
 
+> 📝 **Note:** By default this example is built with `WHISPER_WASM_SINGLE_FILE=ON`
+> which means that that a separate .wasm file will not be generated. Instead, the
+> WASM module is embedded in the main JS file as a base64 encoded string. To
+> generate a separate .wasm file, you need to disable this option by passing
+> `-DWHISPER_WASM_SINGLE_FILE=OFF`:
+> ```console
+> emcmake cmake .. -DWHISPER_WASM_SINGLE_FILE=OFF
+> ```
+> This will generate a `libcommand.wasm` file in the build/bin directory.
+
 > 📝 **Note:** As of Emscripten 3.1.58 (April 2024), separate worker.js files are no
 > longer generated and the worker is embedded in the main JS file. So the worker
 > file will not be geneated for versions later than `3.1.58`.

examples/server/server.cpp

@@ -104,6 +104,7 @@ struct whisper_params {
     bool flash_attn      = false;
     bool suppress_nst    = false;
     bool no_context      = false;
+    bool no_language_probabilities = false;
 
     std::string language        = "en";
     std::string prompt          = "";
@@ -178,6 +179,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -nc,       --no-context        [%-7s] do not use previous audio context\n", params.no_context ? "true" : "false");
     fprintf(stderr, "  -ng,       --no-gpu            [%-7s] do not use gpu\n", params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n", params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nlp,      --no-language-probabilities [%-7s] exclude language probabilities from verbose_json output\n", params.no_language_probabilities ? "true" : "false");
     // Voice Activity Detection (VAD) parameters
     fprintf(stderr, "\nVoice Activity Detection (VAD) options:\n");
     fprintf(stderr, "             --vad                           [%-7s] enable Voice Activity Detection (VAD)\n",            params.vad ? "true" : "false");
@@ -237,6 +239,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-sns"  || arg == "--suppress-nst")    { params.suppress_nst    = true; }
         else if (arg == "-nth"  || arg == "--no-speech-thold") { params.no_speech_thold = std::stof(argv[++i]); }
         else if (arg == "-nc"   || arg == "--no-context")      { params.no_context      = true; }
+        else if (arg == "-nlp"  || arg == "--no-language-probabilities") { params.no_language_probabilities = true; }
 
         // server params
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
@@ -599,6 +602,10 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.vad_samples_overlap = std::stof(req.get_file_value("vad_samples_overlap").content);
     }
+    if (req.has_file("no_language_probabilities"))
+    {
+        params.no_language_probabilities = parse_str_to_bool(req.get_file_value("no_language_probabilities").content);
+    }
 }
 
 }  // namespace
@@ -674,7 +681,10 @@ int main(int argc, char ** argv) {
         if (params.dtw == "large.v3") {
             cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
         }
-
+        if (params.dtw == "large.v3.turbo") { 
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3_TURBO;
+        }
+        
         if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
             fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
             return 3;
@@ -1021,23 +1031,25 @@ int main(int argc, char ** argv) {
         } else if (params.response_format == vjson_format) {
             /* try to match openai/whisper's Python format */
             std::string results = output_str(ctx, params, pcmf32s); 
-            // Get language probabilities
-            std::vector<float> lang_probs(whisper_lang_max_id() + 1, 0.0f);
-            const auto detected_lang_id = whisper_lang_auto_detect(ctx, 0, params.n_threads, lang_probs.data());
             json jres = json{
                 {"task", params.translate ? "translate" : "transcribe"},
                 {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
                 {"duration", float(pcmf32.size())/WHISPER_SAMPLE_RATE},
                 {"text", results},
-                {"segments", json::array()},
-                {"detected_language", whisper_lang_str_full(detected_lang_id)},
-                {"detected_language_probability", lang_probs[detected_lang_id]},
-                {"language_probabilities", json::object()}
+                {"segments", json::array()}
             };
-            // Add all language probabilities
-            for (int i = 0; i <= whisper_lang_max_id(); ++i) {
-                if (lang_probs[i] > 0.001f) { // Only include non-negligible probabilities
-                    jres["language_probabilities"][whisper_lang_str(i)] = lang_probs[i];
+            // Only compute language probabilities if requested (expensive operation)
+            if (!params.no_language_probabilities) {
+                std::vector<float> lang_probs(whisper_lang_max_id() + 1, 0.0f);
+                const auto detected_lang_id = whisper_lang_auto_detect(ctx, 0, params.n_threads, lang_probs.data());
+                jres["detected_language"] = whisper_lang_str_full(detected_lang_id);
+                jres["detected_language_probability"] = lang_probs[detected_lang_id];
+                jres["language_probabilities"] = json::object();
+                // Add all language probabilities
+                for (int i = 0; i <= whisper_lang_max_id(); ++i) {
+                    if (lang_probs[i] > 0.001f) { // Only include non-negligible probabilities
+                        jres["language_probabilities"][whisper_lang_str(i)] = lang_probs[i];
+                    }
                 }
             }
             const int n_segments = whisper_full_n_segments(ctx);

examples/stream.wasm/README.md

@@ -2,7 +2,7 @@
 
 Real-time transcription in the browser using WebAssembly
 
-Online demo: https://whisper.ggerganov.com/stream/
+Online demo: https://ggml.ai/whisper.cpp/stream.wasm/
 
 ## Build instructions
 
@@ -30,6 +30,16 @@ cp bin/libstream.js        /path/to/html/
 cp bin/libstream.worker.js /path/to/html/
 ```
 
+> 📝 **Note:** By default this example is built with `WHISPER_WASM_SINGLE_FILE=ON`
+> which means that that a separate .wasm file will not be generated. Instead, the
+> WASM module is embedded in the main JS file as a base64 encoded string. To
+> generate a separate .wasm file, you need to disable this option by passing
+> `-DWHISPER_WASM_SINGLE_FILE=OFF`:
+> ```console
+> emcmake cmake .. -DWHISPER_WASM_SINGLE_FILE=OFF
+> ```
+> This will generate a `libstream.wasm` file in the build/bin directory.
+
 > 📝 **Note:** As of Emscripten 3.1.58 (April 2024), separate worker.js files are no
 > longer generated and the worker is embedded in the main JS file. So the worker
 > file will not be geneated for versions later than `3.1.58`.

examples/stream.wasm/emscripten.cpp

@@ -31,10 +31,11 @@ void stream_set_status(const std::string & status) {
     g_status = status;
 }
 
-void stream_main(size_t index) {
+void stream_main(size_t index, const std::string & lang) {
     stream_set_status("loading data ...");
 
     struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
+    bool is_multilingual = whisper_is_multilingual(g_contexts[index]);
 
     wparams.n_threads        = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
     wparams.offset_ms        = 0;
@@ -52,7 +53,7 @@ void stream_main(size_t index) {
     // disable temperature fallback
     wparams.temperature_inc  = -1.0f;
 
-    wparams.language         = "en";
+    wparams.language         = is_multilingual ? lang.c_str() : "en";
 
     printf("stream: using %d threads\n", wparams.n_threads);
 
@@ -127,9 +128,8 @@ void stream_main(size_t index) {
         g_contexts[index] = nullptr;
     }
 }
-
 EMSCRIPTEN_BINDINGS(stream) {
-    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
+    emscripten::function("init", emscripten::optional_override([](const std::string & path_model, const std::string & lang) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
                 g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
@@ -138,8 +138,8 @@ EMSCRIPTEN_BINDINGS(stream) {
                     if (g_worker.joinable()) {
                         g_worker.join();
                     }
-                    g_worker = std::thread([i]() {
-                        stream_main(i);
+                    g_worker = std::thread([i, lang]() {
+                        stream_main(i, lang);
                     });
 
                     return i + 1;

examples/stream.wasm/index-tmpl.html

@@ -55,6 +55,7 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <button id="fetch-whisper-base" onclick="loadWhisper('base')">base (142 MB)</button>
                 <br><br>
                 Quantized models:<br><br>
                 <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
@@ -66,6 +67,77 @@
                 -->
             </div>
 
+            <table>
+                <tr>
+                    <td>
+                        Language:
+                        <select id="language" name="language">
+                            <option value="en">English</option>
+                            <option value="ar">Arabic</option>
+                            <option value="hy">Armenian</option>
+                            <option value="az">Azerbaijani</option>
+                            <option value="eu">Basque</option>
+                            <option value="be">Belarusian</option>
+                            <option value="bn">Bengali</option>
+                            <option value="bg">Bulgarian</option>
+                            <option value="ca">Catalan</option>
+                            <option value="zh">Chinese</option>
+                            <option value="hr">Croatian</option>
+                            <option value="cs">Czech</option>
+                            <option value="da">Danish</option>
+                            <option value="nl">Dutch</option>
+                            <option value="en">English</option>
+                            <option value="et">Estonian</option>
+                            <option value="tl">Filipino</option>
+                            <option value="fi">Finnish</option>
+                            <option value="fr">French</option>
+                            <option value="gl">Galician</option>
+                            <option value="ka">Georgian</option>
+                            <option value="de">German</option>
+                            <option value="el">Greek</option>
+                            <option value="gu">Gujarati</option>
+                            <option value="iw">Hebrew</option>
+                            <option value="hi">Hindi</option>
+                            <option value="hu">Hungarian</option>
+                            <option value="is">Icelandic</option>
+                            <option value="id">Indonesian</option>
+                            <option value="ga">Irish</option>
+                            <option value="it">Italian</option>
+                            <option value="ja">Japanese</option>
+                            <option value="kn">Kannada</option>
+                            <option value="ko">Korean</option>
+                            <option value="la">Latin</option>
+                            <option value="lv">Latvian</option>
+                            <option value="lt">Lithuanian</option>
+                            <option value="mk">Macedonian</option>
+                            <option value="ms">Malay</option>
+                            <option value="mt">Maltese</option>
+                            <option value="no">Norwegian</option>
+                            <option value="fa">Persian</option>
+                            <option value="pl">Polish</option>
+                            <option value="pt">Portuguese</option>
+                            <option value="ro">Romanian</option>
+                            <option value="ru">Russian</option>
+                            <option value="sr">Serbian</option>
+                            <option value="sk">Slovak</option>
+                            <option value="sl">Slovenian</option>
+                            <option value="es">Spanish</option>
+                            <option value="sw">Swahili</option>
+                            <option value="sv">Swedish</option>
+                            <option value="ta">Tamil</option>
+                            <option value="te">Telugu</option>
+                            <option value="th">Thai</option>
+                            <option value="tr">Turkish</option>
+                            <option value="uk">Ukrainian</option>
+                            <option value="ur">Urdu</option>
+                            <option value="vi">Vietnamese</option>
+                            <option value="cy">Welsh</option>
+                            <option value="yi">Yiddish</option>
+                        </select>
+                    </td>
+                </tr>
+            </table>
+
             <br>
 
             <div id="input">
@@ -176,6 +248,7 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'base'   : 'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
 
                     'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
                     'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
@@ -184,6 +257,7 @@
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+                    'base':     142,
 
                     'tiny-en-q5_1':   31,
                     'base-en-q5_1':   57,
@@ -197,6 +271,7 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('fetch-whisper-base').style.display = 'none';
 
                 document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
@@ -212,6 +287,7 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base'); if (el) el.style.display = 'inline-block';
 
                     el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
@@ -368,7 +444,7 @@
 
             function onStart() {
                 if (!instance) {
-                    instance = Module.init('whisper.bin');
+                    instance = Module.init('whisper.bin', document.getElementById('language').value);
 
                     if (instance) {
                         printTextarea("js: whisper initialized, instance: " + instance);

examples/talk-llama/llama-arch.cpp

@@ -34,6 +34,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
     { LLM_ARCH_PLAMO,            "plamo"            },
+    { LLM_ARCH_PLAMO2,           "plamo2"           },
     { LLM_ARCH_CODESHELL,        "codeshell"        },
     { LLM_ARCH_ORION,            "orion"            },
     { LLM_ARCH_INTERNLM2,        "internlm2"        },
@@ -42,8 +43,12 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
     { LLM_ARCH_GEMMA3,           "gemma3"           },
+    { LLM_ARCH_GEMMA3N,          "gemma3n"          },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
+    { LLM_ARCH_MAMBA2,           "mamba2"           },
+    { LLM_ARCH_JAMBA,            "jamba"            },
+    { LLM_ARCH_FALCON_H1,        "falcon-h1"        },
     { LLM_ARCH_XVERSE,           "xverse"           },
     { LLM_ARCH_COMMAND_R,        "command-r"        },
     { LLM_ARCH_COHERE2,          "cohere2"          },
@@ -63,18 +68,26 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_EXAONE,           "exaone"           },
+    { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
     { LLM_ARCH_RWKV7,            "rwkv7"            },
     { LLM_ARCH_ARWKV7,           "arwkv7"           },
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
+    { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
+    { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
+    { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
+    { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
+    { LLM_ARCH_SMOLLM3,          "smollm3"          },
+    { LLM_ARCH_LFM2,             "lfm2"             },
+    { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -147,7 +160,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
-    { LLM_KV_ATTENTION_LAYER_INDICES,                "%s.attention.layer_indices"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -168,6 +180,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_SSM_INNER_SIZE,     "%s.ssm.inner_size"     },
     { LLM_KV_SSM_STATE_SIZE,     "%s.ssm.state_size"     },
     { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
+    { LLM_KV_SSM_GROUP_COUNT,    "%s.ssm.group_count"    },
     { LLM_KV_SSM_DT_B_C_RMS,     "%s.ssm.dt_b_c_rms"     },
 
     { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
@@ -180,6 +193,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
 
+    { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
+
     { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
     { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
     { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
@@ -773,6 +788,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_PLAMO2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_X,           "blk.%d.ssm_x" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+            { LLM_TENSOR_SSM_DT_NORM,     "blk.%d.ssm_dt_norm" },
+            { LLM_TENSOR_SSM_B_NORM,      "blk.%d.ssm_b_norm" },
+            { LLM_TENSOR_SSM_C_NORM,      "blk.%d.ssm_c_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_CODESHELL,
         {
@@ -932,6 +977,42 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA3N,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,           "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,          "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,            "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,               "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,          "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,               "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,          "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,               "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,             "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,       "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_NORM,             "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,             "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,             "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,               "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,        "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" },
+            { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" },
+            { LLM_TENSOR_PER_LAYER_PROJ_NORM,  "per_layer_proj_norm" },
+            { LLM_TENSOR_ALTUP_UNEMBD_PROJ,    "altup_unembd_proj" },
+            { LLM_TENSOR_ALTUP_PROJ,           "altup_proj" },
+            { LLM_TENSOR_PER_LAYER_INP_GATE,   "blk.%d.inp_gate" },
+            { LLM_TENSOR_PER_LAYER_PROJ,       "blk.%d.proj" },
+            { LLM_TENSOR_PER_LAYER_POST_NORM,  "blk.%d.post_norm" },
+            { LLM_TENSOR_ALTUP_CORRECT_COEF,   "blk.%d.altup_correct_coef" },
+            { LLM_TENSOR_ALTUP_CORRECT_SCALE,  "blk.%d.altup_correct_scale" },
+            { LLM_TENSOR_ALTUP_PREDICT_COEF,   "blk.%d.altup_predict_coef" },
+            { LLM_TENSOR_ALTUP_ROUTER,         "blk.%d.altup_router" },
+            { LLM_TENSOR_ALTUP_ROUTER_NORM,    "blk.%d.altup_router_norm" },
+            { LLM_TENSOR_LAUREL_L,             "blk.%d.laurel_l" },
+            { LLM_TENSOR_LAUREL_R,             "blk.%d.laurel_r" },
+            { LLM_TENSOR_LAUREL_POST_NORM,     "blk.%d.laurel_post_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {
@@ -966,6 +1047,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
         },
     },
+    {
+        LLM_ARCH_MAMBA2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,        "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+        },
+    },
+    {
+        LLM_ARCH_JAMBA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_X,           "blk.%d.ssm_x" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_DT_NORM,     "blk.%d.ssm_dt_norm" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_B_NORM,      "blk.%d.ssm_b_norm" },
+            { LLM_TENSOR_SSM_C_NORM,      "blk.%d.ssm_c_norm" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
+    {
+        LLM_ARCH_FALCON_H1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,        "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_XVERSE,
         {
@@ -1359,6 +1511,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_EXAONE4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        }
+    },
     {
         LLM_ARCH_RWKV6,
         {
@@ -1526,6 +1698,43 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_GRANITE_HYBRID,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,       "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // moe FFN
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,  "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            // shared expert
+            { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,   "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_CHAMELEON,
         {
@@ -1621,6 +1830,126 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
         }
     },
+    {
+        LLM_ARCH_ERNIE4_5,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_ERNIE4_5_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
+    {
+        LLM_ARCH_HUNYUAN_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
+    {
+        LLM_ARCH_SMOLLM3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,       "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_LFM2,
+        {
+            { LLM_TENSOR_ATTN_NORM,         "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,            "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,            "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,            "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,          "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_K_NORM,       "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,       "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_FFN_DOWN,          "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE,          "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_NORM,          "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,            "blk.%d.ffn_up" },
+            { LLM_TENSOR_SHORTCONV_CONV,    "blk.%d.shortconv.conv" },
+            { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
+            { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
+            { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+        }
+    },
+    {
+        LLM_ARCH_DREAM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1705,7 +2034,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_ACT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
+    {LLM_TENSOR_SSM_DT_NORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SSM_B_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SSM_C_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_D,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SSM_NORM,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_X,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_K,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -1749,6 +2082,23 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_GATE_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    // altup / laurel (gemma 3n)
+    {LLM_TENSOR_PER_LAYER_TOKEN_EMBD,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_PER_LAYER_MODEL_PROJ,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_PROJ_NORM,        {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_PROJ,                 {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_UNEMBD_PROJ,          {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_INP_GATE,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_PROJ,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_POST_NORM,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_CORRECT_COEF,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_CORRECT_SCALE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_PREDICT_COEF,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_ROUTER,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_ROUTER_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_LAUREL_L,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_LAUREL_R,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     // this tensor is loaded for T5, but never used
     {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
     {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_INPUT,     GGML_OP_IM2COL}},
@@ -1767,6 +2117,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_CONVNEXT_PW1,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CONVNEXT_PW2,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CONVNEXT_GAMMA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+    {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
@@ -1822,6 +2175,7 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
 bool llm_arch_is_recurrent(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
         case LLM_ARCH_RWKV6:
         case LLM_ARCH_RWKV6QWEN2:
         case LLM_ARCH_RWKV7:
@@ -1833,9 +2187,22 @@ bool llm_arch_is_recurrent(const llm_arch & arch) {
 }
 
 bool llm_arch_is_hybrid(const llm_arch & arch) {
-    // TODO: There are currently no hybrid models! Once there are, this will be
-    //  the place to identify them
     switch (arch) {
+        case LLM_ARCH_JAMBA:
+        case LLM_ARCH_FALCON_H1:
+        case LLM_ARCH_PLAMO2:
+        case LLM_ARCH_GRANITE_HYBRID:
+        case LLM_ARCH_LFM2:
+            return true;
+        default:
+            return false;
+    }
+}
+
+bool llm_arch_is_diffusion(const llm_arch & arch) {
+    switch (arch) {
+        case LLM_ARCH_DREAM:
+            return true;
         default:
             return false;
     }

examples/talk-llama/llama-arch.h

@@ -38,6 +38,7 @@ enum llm_arch {
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
+    LLM_ARCH_PLAMO2,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
@@ -46,8 +47,12 @@ enum llm_arch {
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
+    LLM_ARCH_GEMMA3N,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
+    LLM_ARCH_MAMBA2,
+    LLM_ARCH_JAMBA,
+    LLM_ARCH_FALCON_H1,
     LLM_ARCH_XVERSE,
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_COHERE2,
@@ -67,18 +72,26 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_EXAONE,
+    LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,
     LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_GRANITE_HYBRID,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
+    LLM_ARCH_ERNIE4_5,
+    LLM_ARCH_ERNIE4_5_MOE,
+    LLM_ARCH_HUNYUAN_MOE,
+    LLM_ARCH_SMOLLM3,
+    LLM_ARCH_LFM2,
+    LLM_ARCH_DREAM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -151,7 +164,6 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
-    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -172,6 +184,7 @@ enum llm_kv {
     LLM_KV_SSM_CONV_KERNEL,
     LLM_KV_SSM_STATE_SIZE,
     LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_GROUP_COUNT,
     LLM_KV_SSM_DT_B_C_RMS,
 
     LLM_KV_WKV_HEAD_SIZE,
@@ -219,6 +232,8 @@ enum llm_kv {
 
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
 
+    LLM_KV_SHORTCONV_L_CACHE,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -269,12 +284,32 @@ enum llm_tensor {
     LLM_TENSOR_LAYER_OUT_NORM,
     LLM_TENSOR_POST_ATTN_NORM,
     LLM_TENSOR_POST_MLP_NORM,
+    LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
+    LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n
+    LLM_TENSOR_PER_LAYER_INP_GATE,   // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ,       // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ_NORM,  // gemma3n
+    LLM_TENSOR_PER_LAYER_POST_NORM,  // gemma3n
+    LLM_TENSOR_ALTUP_PROJ,           // gemma3n
+    LLM_TENSOR_ALTUP_UNEMBD_PROJ,    // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_SCALE,  // gemma3n
+    LLM_TENSOR_ALTUP_PREDICT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER,         // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER_NORM,    // gemma3n
+    LLM_TENSOR_LAUREL_L,             // gemma3n
+    LLM_TENSOR_LAUREL_R,             // gemma3n
+    LLM_TENSOR_LAUREL_POST_NORM,     // gemma3n
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
     LLM_TENSOR_SSM_DT,
+    LLM_TENSOR_SSM_DT_NORM,
     LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_B_NORM,
+    LLM_TENSOR_SSM_C_NORM,
     LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
@@ -368,6 +403,9 @@ enum llm_tensor {
     LLM_TENSOR_POS_NET_ATTN_K,
     LLM_TENSOR_POS_NET_ATTN_V,
     LLM_TENSOR_POS_NET_ATTN_OUT,
+    LLM_TENSOR_SHORTCONV_CONV,
+    LLM_TENSOR_SHORTCONV_INPROJ,
+    LLM_TENSOR_SHORTCONV_OUTPROJ,
 };
 
 enum llm_tensor_layer {
@@ -444,3 +482,4 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
 
 bool llm_arch_is_recurrent(const llm_arch & arch);
 bool llm_arch_is_hybrid   (const llm_arch & arch);
+bool llm_arch_is_diffusion(const llm_arch & arch);

examples/talk-llama/llama-batch.cpp

@@ -27,6 +27,7 @@ bool llama_batch_allocr::init(
         const llama_vocab & vocab,
         const llama_memory_i * memory,
         uint32_t n_embd,
+        uint32_t n_seq_max,
         bool output_all) {
     clear();
 
@@ -40,6 +41,11 @@ bool llama_batch_allocr::init(
     // validate input batch
     //
 
+    if (n_seq_max > LLAMA_MAX_SEQ) {
+        LLAMA_LOG_ERROR("%s: n_seq_max = %d > %d\n", __func__, n_seq_max, LLAMA_MAX_SEQ);
+        return false;
+    }
+
     if (batch.token) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
@@ -52,8 +58,8 @@ bool llama_batch_allocr::init(
     if (batch.seq_id) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
-                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
-                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= (llama_seq_id) n_seq_max)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], (llama_seq_id) n_seq_max);
                     return false;
                 }
             }
@@ -86,7 +92,7 @@ bool llama_batch_allocr::init(
 
         // initialize the starting position for each sequence based on the positions in the memory
         llama_pos p0[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (!memory) {
                 // if no memory -> start from 0
                 p0[s] = 0;
@@ -143,13 +149,16 @@ bool llama_batch_allocr::init(
     // compute stats
     //
 
-    this->n_embd = n_embd;
+    this->n_embd    = n_embd;
+    this->n_seq_max = n_seq_max;
 
     // count the outputs in this batch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         n_outputs += batch.logits[i] != 0;
     }
 
+    has_cpl = false;
+
     // determine coupled sequences
     // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
@@ -166,6 +175,8 @@ bool llama_batch_allocr::init(
 
                 // note: tracking the other way around is not necessary for now
                 //seq_cpl[s0][s1] = true;
+
+                has_cpl = true;
             }
         }
     }
@@ -187,7 +198,7 @@ bool llama_batch_allocr::init(
             seq_set_map[cur].push_back(i);
         }
 
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (seq_set_unq.test(s)) {
                 seq_idx[s] = seq_id_unq.size();
                 seq_id_unq.push_back(s);
@@ -199,7 +210,7 @@ bool llama_batch_allocr::init(
         LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
 
         llama_ubatch ubatch {
-            /*.equal_seqs   =*/ false,
+            /*.b_equal_seqs =*/ false,
             /*.n_tokens     =*/ (uint32_t) batch.n_tokens,
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
@@ -212,6 +223,7 @@ bool llama_batch_allocr::init(
             /*.seq_id_unq   =*/ this->seq_id_unq.data(),
             /*.seq_idx      =*/ this->seq_idx.data(),
             /*.output       =*/ batch.logits,
+            /*.data         =*/ {},
         };
 
         ubatch_print(ubatch, debug);
@@ -239,27 +251,40 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_pos[s].empty()) {
             continue;
         }
 
-        if (memory) {
+        const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
+
+        if (p0 >= 0) {
+            bool ok = true;
+
             if (batch.token) {
-                if (seq_pos_min(s) != memory->seq_pos_max(s) + 1) {
-                    LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s);
-                    return false;
+                if (seq_pos_min(s) != p0 + 1) {
+                    ok = false;
                 }
             } else {
                 assert(batch.embd);
 
                 // for embeddings (typically used as vision input), we allow them to have repeating positions
                 // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
-                if (seq_pos_min(s) != memory->seq_pos_max(s) && seq_pos_min(s) != memory->seq_pos_max(s) + 1) {
-                    LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s);
-                    return false;
+                if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) {
+                    ok = false;
                 }
             }
+
+            if (!ok) {
+                LLAMA_LOG_ERROR(
+                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                        " it is required that the sequence positions remain consecutive: Y = X + 1\n",
+                        __func__, s, s, p0, s, seq_pos_min(s));
+
+                return false;
+            }
         }
 
         if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
@@ -269,8 +294,8 @@ bool llama_batch_allocr::init(
     }
 
     if (memory) {
-        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
-            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+        for (uint32_t s0 = 0; s0 < n_seq_max; ++s0) {
+            for (uint32_t s1 = 0; s1 < n_seq_max; ++s1) {
                 if (seq_cpl[s0][s1]) {
                     if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
                         memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
@@ -301,12 +326,12 @@ bool llama_batch_allocr::init(
     //
     {
         seq_set_t cur_seq_set[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_set[s].set();
         }
 
         llama_pos cur_seq_pos[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_pos[s] = -1;
         }
 
@@ -342,39 +367,38 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
     clear();
     split_reset();
 
-    ubatches.emplace_back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    auto & ubatch = ubatches.back();
-
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .clear();
-    ubatch.pos       .resize(n_tokens);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .clear();
+    udata->pos       .resize(n_tokens);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        ubatch.seq_idx[s] = s;
-        ubatch.seq_id_unq.push_back(s);
+        udata->seq_idx[s] = s;
+        udata->seq_id_unq.push_back(s);
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ true,
+        /*.b_equal_seqs =*/ true,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
 
-        /*.token        =*/ ubatch.token.data(),
+        /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     return res;
@@ -392,6 +416,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const {
     return n_outputs;
 }
 
+uint32_t llama_batch_allocr::get_n_used() const {
+    return n_used;
+}
+
 std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
     return out_ids;
 }
@@ -407,10 +435,10 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
 void llama_batch_allocr::split_reset() {
     out_ids.clear();
 
+    n_used = 0;
+
     used.clear();
     used.resize(get_n_tokens(), false);
-
-    ubatches.clear();
 }
 
 llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
@@ -431,6 +459,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         ++cur_idx;
 
@@ -446,9 +475,17 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
     return ubatch_add(idxs, idxs.size(), false);
 }
 
-llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+    if (sequential && has_cpl) {
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+
+        return {};
+    }
+
     std::vector<seq_set_t> cur_seq_set;
 
+    llama_seq_id last_seq_id = -1;
+
     // determine the non-overlapping sequence sets participating in this ubatch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         if (used[i]) {
@@ -465,9 +502,16 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             }
         }
 
+        // accept only increasing sequence ids
+        if (sequential) {
+            add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1);
+        }
+
         if (add) {
             cur_seq_set.push_back(seq_set[i]);
 
+            last_seq_id = batch.seq_id[i][0];
+
             if (cur_seq_set.size() > n_ubatch) {
                 break;
             }
@@ -516,6 +560,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             idxs_per_seq[s].push_back(idx);
 
             used[idx] = true;
+            ++n_used;
 
             ++cur_idx[s];
         }
@@ -557,6 +602,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         if (idxs.size() >= n_ubatch) {
             break;
@@ -607,78 +653,77 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     assert(n_tokens%n_seqs == 0);
 
-    ubatches.emplace_back();
-
-    auto & ubatch = ubatches.back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
     const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
 
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
     const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
 
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .resize(n_embd_all);
-    ubatch.pos       .resize(n_pos_all);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .resize(n_embd_all);
+    udata->pos       .resize(n_pos_all);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     seq_set_t seq_set_unq;
 
     for (size_t i = 0; i < idxs.size(); ++i) {
         if (batch.token) {
-            ubatch.token[i] = batch.token[idxs[i]];
+            udata->token[i] = batch.token[idxs[i]];
         }
 
         if (batch.embd) {
-            memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
+            memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
         for (int j = 0; j < n_pos_cur; ++j) {
-            ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
         }
 
-        ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]];
-        ubatch.seq_id[i]   = batch.seq_id[idxs[i]];
-        ubatch.output[i]   = batch.logits[idxs[i]];
+        udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
+        udata->seq_id[i]   = batch.seq_id[idxs[i]];
+        udata->output[i]   = batch.logits[idxs[i]];
 
-        for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
-            seq_set_unq.set(ubatch.seq_id[i][s]);
+        for (int s = 0; s < udata->n_seq_id[i]; ++s) {
+            seq_set_unq.set(udata->seq_id[i][s]);
         }
 
-        if (ubatch.output[i]) {
+        if (udata->output[i]) {
             out_ids.push_back(idxs[i]);
         }
     }
 
-    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_set_unq.test(s)) {
-            ubatch.seq_idx[s] = ubatch.seq_id_unq.size();
-            ubatch.seq_id_unq.push_back(s);
+            udata->seq_idx[s] = udata->seq_id_unq.size();
+            udata->seq_id_unq.push_back(s);
         }
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ equal_seqs,
+        /*.b_equal_seqs =*/ equal_seqs,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
-        /*.n_seqs_unq   =*/ (uint32_t) ubatch.seq_id_unq.size(),
+        /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
 
-        /*.token        =*/ batch.token ? ubatch.token.data() : nullptr,
-        /*.embd         =*/ batch.embd ? ubatch.embd.data() : nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.token        =*/ batch.token ? udata->token.data() : nullptr,
+        /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1);
+        LLAMA_LOG_DEBUG("%s: added ubatch to split:\n", __func__);
 
         ubatch_print(res, debug);
     }
@@ -688,7 +733,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
 void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs());
         LLAMA_LOG_DEBUG("%s:   n_tokens     = %d\n", __func__, ubatch.n_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seqs       = %d\n", __func__, ubatch.n_seqs);

examples/talk-llama/llama-batch.h

@@ -8,12 +8,17 @@
 #include <vector>
 #include <set>
 #include <bitset>
+#include <memory>
 #include <unordered_map>
 
 // keep this struct lightweight
-// it points to data in `llama_batch_allocr`
 struct llama_ubatch {
-    bool equal_seqs;
+    bool equal_seqs() const {
+        return b_equal_seqs != 0;
+    }
+
+    uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
+                           //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
 
     uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
@@ -34,6 +39,20 @@ struct llama_ubatch {
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
     int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
     int8_t       *  output;     // [n_tokens]         | i   | -
+
+    struct data_t {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id>   seq_id_unq;
+        std::vector<int32_t>        seq_idx;
+        std::vector<int8_t>         output;
+    };
+
+    // the llama_ubatch pointers above point to this data if set. otherwise - points to non-owning data
+    std::shared_ptr<data_t> data;
 };
 
 // a helper for sanitizing, fulfilling and splitting a batch
@@ -48,12 +67,14 @@ public:
             const llama_vocab & vocab,
             const llama_memory_i * memory,
             uint32_t n_embd,
+            uint32_t n_seq_max,
             bool output_all);
 
     const llama_batch & get_batch() const;
 
     uint32_t get_n_tokens()  const;
     uint32_t get_n_outputs() const;
+    uint32_t get_n_used()    const;
 
     // the array of output indices in the order they were encountered during the ubatch splitting
     std::vector<int32_t> & get_out_ids();
@@ -69,7 +90,8 @@ public:
     llama_ubatch split_simple(uint32_t n_ubatch);
 
     // make ubatches of equal-length sequences sets
-    llama_ubatch split_equal(uint32_t n_ubatch);
+    // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);
@@ -98,6 +120,7 @@ private:
     const uint32_t n_pos_per_embd;
 
     uint32_t n_embd;
+    uint32_t n_seq_max;
     uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
@@ -112,6 +135,9 @@ private:
     using pos_set_t = std::set<llama_pos>;
     using seq_cpl_t = std::vector<bool>;
 
+    // helper flag to quickly determine if there are any coupled sequences in the batch
+    bool has_cpl = false;
+
     std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
     std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
 
@@ -125,23 +151,10 @@ private:
     // batch indices of the output
     std::vector<int32_t> out_ids;
 
+    uint32_t n_used;
+
     // used[i] indicates if token i has already been used in a previous ubatch
     std::vector<bool> used;
 
-    // llama_ubatch points to this data:
-    struct ubatch {
-        std::vector<llama_token>    token;
-        std::vector<float>          embd;
-        std::vector<llama_pos>      pos;
-        std::vector<int32_t>        n_seq_id;
-        std::vector<llama_seq_id *> seq_id;
-        std::vector<llama_seq_id>   seq_id_unq;
-        std::vector<int32_t>        seq_idx;
-        std::vector<int8_t>         output;
-    };
-
-    // current splitting state:
-    std::vector<ubatch> ubatches;
-
     int debug;
 };

examples/talk-llama/llama-chat.cpp

@@ -56,6 +56,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
+    { "exaone4",           LLM_CHAT_TEMPLATE_EXAONE_4          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
@@ -64,6 +65,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
+    { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
+    { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -166,10 +169,13 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_3;
     } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
+        if (tmpl_contains("[|tool|]")) {
+            return LLM_CHAT_TEMPLATE_EXAONE_4;
+        }
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
         // EXAONE-3.0-7.8B-Instruct
         return LLM_CHAT_TEMPLATE_EXAONE_3;
-    } else if (tmpl_contains("rwkv-world")) {
+    } else if (tmpl_contains("rwkv-world") || tmpl_contains("{{- 'User: ' + message['content']|trim + '\\n\\n' -}}")) {
         return LLM_CHAT_TEMPLATE_RWKV_WORLD;
     } else if (tmpl_contains("<|start_of_role|>")) {
         return LLM_CHAT_TEMPLATE_GRANITE;
@@ -185,6 +191,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
     } else if (tmpl_contains("<|endofuserprompt|>")) {
         return LLM_CHAT_TEMPLATE_DOTS1;
+    } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
+        return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
+    } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
+        return LLM_CHAT_TEMPLATE_KIMI_K2;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -526,14 +536,35 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[|assistant|]";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
-        // this template requires the model to have "\n\n" as EOT token
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_4) {
         for (auto message : chat) {
             std::string role(message->role);
-            if (role == "user") {
-                ss << "User: " << message->content << "\n\nAssistant:";
-            } else {
-                ss << message->content << "\n\n";
+            if (role == "system") {
+                ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "user") {
+                ss << "[|user|]" << trim(message->content) << "\n";
+            } else if (role == "assistant") {
+                ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "tool") {
+                ss << "[|tool|]" << trim(message->content) << "[|endofturn|]\n";
+            }
+        }
+        if (add_ass) {
+            ss << "[|assistant|]";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
+        // this template requires the model to have "\n\n" as EOT token
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "system") {
+                ss << "System: " << trim(chat[i]->content) << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << trim(chat[i]->content) << "\n\n";
+                if (i == chat.size() - 1) {
+                    ss << "Assistant:";
+                }
+            } else if (role == "assistant") {
+                ss << "Assistant: " << trim(chat[i]->content) << "\n\n";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {
@@ -660,6 +691,37 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|response|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_MOE) {
+        // tencent/Hunyuan-A13B-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|startoftext|>" << message->content << "<|extra_4|>";
+            } else if (role == "assistant") {
+                ss << "<|startoftext|>" << message->content << "<|eos|>";
+            } else {
+                ss << "<|startoftext|>" << message->content << "<|extra_0|>";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
+        // moonshotai/Kimi-K2-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|im_system|>system<|im_middle|>";
+            } else if (role == "user") {
+                ss << "<|im_user|>user<|im_middle|>";
+            } else if (role == "assistant") {
+                ss << "<|im_assistant|>assistant<|im_middle|>";
+            } else if (role == "tool") {
+                ss << "<|im_system|>tool<|im_middle|>";
+            }
+
+            ss << message->content << "<|im_end|>";
+        }
+        if (add_ass) {
+            ss << "<|im_assistant|>assistant<|im_middle|>";
+        }
     } else {
         // template not supported
         return -1;

examples/talk-llama/llama-chat.h

@@ -35,6 +35,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
+    LLM_CHAT_TEMPLATE_EXAONE_4,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
@@ -44,6 +45,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_DOTS1,
+    LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
+    LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

examples/talk-llama/llama-context.cpp

@@ -98,10 +98,20 @@ llama_context::llama_context(
         LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
         cparams.n_batch = GGML_KQ_MASK_PAD;
     }
-
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
+    cparams.kv_unified = params.kv_unified;
+
+    {
+        const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
+
+        if (!supports_set_rows && !cparams.kv_unified) {
+            LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
+            cparams.kv_unified = true;
+        }
+    }
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -112,6 +122,7 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
     LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: kv_unified    = %s\n",   __func__, cparams.kv_unified ? "true" : "false");
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
@@ -227,8 +238,8 @@ llama_context::llama_context(
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
-        // buffer used to store the computation graph and the tensor meta data
-        buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
+        gf_res_prev.reset(new llm_graph_result(max_nodes));
+        gf_res_reserve.reset(new llm_graph_result(max_nodes));
 
         // TODO: move these checks to ggml_backend_sched
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
@@ -267,7 +278,7 @@ llama_context::llama_context(
 
     // reserve worst-case graph
     if (!hparams.vocab_only && memory) {
-        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -280,16 +291,16 @@ llama_context::llama_context(
 
         // simulate full KV cache
 
-        const auto mstate = memory->init_full();
-        if (!mstate) {
+        const auto mctx = memory->init_full();
+        if (!mctx) {
             throw std::runtime_error("failed to initialize KV cache");
         }
 
         cross.v_embd.clear();
 
-        // reserve pp graph first so that buffers are only allocated once
+        // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -298,9 +309,9 @@ llama_context::llama_context(
             n_nodes_pp  = ggml_graph_n_nodes(gf);
         }
 
-        // reserve with tg graph to get the number of splits and nodes
+        // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(1, 1, 1, mstate.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -311,7 +322,11 @@ llama_context::llama_context(
 
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
+            // TODO: not sure if the following graph would be worster case for multi-stream KV caches:
+            //
+            // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
+            //
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -388,10 +403,6 @@ ggml_backend_sched_t llama_context::get_sched() const {
     return sched.get();
 }
 
-ggml_context * llama_context::get_ctx_compute() const {
-    return ctx_compute.get();
-}
-
 uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
@@ -444,8 +455,8 @@ bool llama_context::kv_self_update(bool optimize) {
         optimize |= memory_force_optimize;
         memory_force_optimize = false;
 
-        const auto mstate = memory->init_update(this, optimize);
-        switch (mstate->get_status()) {
+        const auto mctx = memory->init_update(this, optimize);
+        switch (mctx->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
                 {
                     // noop
@@ -463,22 +474,27 @@ bool llama_context::kv_self_update(bool optimize) {
                 }
         }
 
-        if (!mstate->apply()) {
+        // reset the previous graph result to make sure that it won't be reused
+        // TODO: change the mctx->apply() to return information if a graph reserve is needed
+        //       reset the graph result only if the memory module did reset the scheduler
+        gf_res_prev->reset();
+
+        if (!mctx->apply()) {
             LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__);
         }
     }
 
     // if the memory module did any computation, we have to reserve a new worst-case graph
     {
-        const auto mstate = memory->init_full();
-        if (!mstate) {
-            throw std::runtime_error("failed to initialize memory state");
+        const auto mctx = memory->init_full();
+        if (!mctx) {
+            throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs   = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
-        auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
+        auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
         if (!gf) {
             LLAMA_LOG_ERROR("%s: failed to reserve graph after the memory update\n", __func__);
         }
@@ -492,12 +508,16 @@ enum llama_pooling_type llama_context::pooling_type() const {
 }
 
 float * llama_context::get_logits() {
+    output_reorder();
+
     return logits;
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
     int64_t j = -1;
 
+    output_reorder();
+
     try {
         if (logits == nullptr) {
             throw std::runtime_error("no logits");
@@ -534,12 +554,16 @@ float * llama_context::get_logits_ith(int32_t i) {
 }
 
 float * llama_context::get_embeddings() {
+    output_reorder();
+
     return embd;
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
     int64_t j = -1;
 
+    output_reorder();
+
     try {
         if (embd == nullptr) {
             throw std::runtime_error("no embeddings");
@@ -678,38 +702,59 @@ bool llama_context::apply_adapter_cvec(
     return cvec.apply(model, data, len, n_embd, il_start, il_end);
 }
 
-llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_state_i * mstate, ggml_status & ret) {
-    if (mstate && !mstate->apply()) {
-        LLAMA_LOG_ERROR("%s: failed to apply memory state\n", __func__);
+llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) {
+    if (mctx && !mctx->apply()) {
+        LLAMA_LOG_ERROR("%s: failed to apply memory context\n", __func__);
         ret = GGML_STATUS_FAILED;
         return nullptr;
     }
 
-    auto * gf = graph_init();
-    if (!gf) {
-        LLAMA_LOG_ERROR("%s: failed to initialize graph\n", __func__);
-        ret = GGML_STATUS_FAILED;
-        return nullptr;
+    auto * res = gf_res_prev.get();
+    auto * gf  = res->get_gf();
+
+    // the new graph parameters
+    // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
+    const auto gparams = graph_params(res, ubatch, mctx, gtype);
+
+    if (res->can_reuse(gparams)) {
+        //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
+
+        n_reused++;
+    } else {
+        res->reset();
+
+        ggml_backend_sched_reset(sched.get());
+        ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
+
+        //const auto t_start_us = ggml_time_us();
+
+        gf = model.build_graph(gparams);
+
+        //LLAMA_LOG_INFO("graph build time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0);
+
+        if (!gf) {
+            LLAMA_LOG_ERROR("%s: failed to initialize graph\n", __func__);
+            ret = GGML_STATUS_FAILED;
+            return nullptr;
+        }
+
+        if (!ggml_backend_sched_alloc_graph(sched.get(), gf)) {
+            LLAMA_LOG_ERROR("%s: failed to allocate graph\n", __func__);
+            ret = GGML_STATUS_ALLOC_FAILED;
+            return nullptr;
+        }
     }
 
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mstate);
-    if (!res) {
-        LLAMA_LOG_ERROR("%s: failed to build graph\n", __func__);
-        ret = GGML_STATUS_FAILED;
-        return nullptr;
+    // set the input data for the input tensors
+    {
+        //const auto t_start_us = ggml_time_us();
+
+        res->set_inputs(&ubatch);
+
+        //LLAMA_LOG_INFO("graph set inputs time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0);
     }
 
-    // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
-
-    if (!ggml_backend_sched_alloc_graph(sched.get(), gf)) {
-        LLAMA_LOG_ERROR("%s: failed to allocate graph\n", __func__);
-        ret = GGML_STATUS_ALLOC_FAILED;
-        return nullptr;
-    }
-
-    res->set_inputs(&ubatch);
-
-    const auto status = graph_compute(gf, ubatch.n_tokens > 1);
+    const auto status = graph_compute(res->get_gf(), ubatch.n_tokens > 1);
     if (status != GGML_STATUS_SUCCESS) {
         LLAMA_LOG_ERROR("%s: failed to compute graph, compute status: %d\n", __func__, status);
         ret = status;
@@ -731,16 +776,19 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd;
+    const int32_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
-    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) {
+    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
 
     const uint32_t n_tokens = balloc->get_n_tokens();
 
+    // [TAG_NO_CACHE_PAD]
+    // TODO: add new split mode where we pad the input sequences so that ubatch.equal_seqs == true
     const llama_ubatch ubatch = balloc->split_simple(n_tokens);
 
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
@@ -767,9 +815,6 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     n_outputs = n_tokens;
 
-    ggml_backend_sched_reset(sched.get());
-    ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
-
     const auto causal_attn_org = cparams.causal_attn;
 
     // always use non-causal attention for encoder graphs
@@ -778,7 +823,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     cparams.causal_attn = false;
 
     ggml_status status;
-    const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_ENCODER, nullptr, status);
+    const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_ENCODER, nullptr, status);
 
     cparams.causal_attn = causal_attn_org;
 
@@ -791,10 +836,20 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
+    auto * t_logits = res->get_logits();
     auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd();
 
+    // extract logits
+   if (logits && t_logits) {
+        ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
+        GGML_ASSERT(backend_res != nullptr);
+        GGML_ASSERT(logits != nullptr);
+
+        ggml_backend_tensor_get_async(backend_res, t_logits, logits, 0, n_tokens*n_vocab*sizeof(float));
+    }
+
     // extract embeddings
-    if (t_embd) {
+    if (embd && t_embd) {
         ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
         GGML_ASSERT(backend_embd != nullptr);
 
@@ -844,9 +899,11 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(sched.get());
+    if (!supports_set_rows) {
+        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+        // overlap with device computation.
+        ggml_backend_sched_reset(sched.get());
+    }
 
     // TODO: hacky solution
     if (model.arch == LLM_ARCH_T5 && t_embd) {
@@ -899,7 +956,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // when computing embeddings, all tokens are output
     const bool output_all = cparams.embeddings;
 
-    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) {
+    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -927,27 +984,28 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
     // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
+    output_swaps.clear();
 
     bool did_optimize = false;
 
     // handle any pending defrags/shifts
     kv_self_update(false);
 
-    llama_memory_state_ptr mstate;
+    llama_memory_context_ptr mctx;
 
     while (true) {
-        mstate = memory->init_batch(*balloc, cparams.n_ubatch, output_all);
-        if (!mstate) {
+        mctx = memory->init_batch(*balloc, cparams.n_ubatch, output_all);
+        if (!mctx) {
             return -2;
         }
 
-        switch (mstate->get_status()) {
+        switch (mctx->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
                 {
                 } break;
             case LLAMA_MEMORY_STATUS_NO_UPDATE:
                 {
-                    LLAMA_LOG_ERROR("%s: unexpected memory state status: %d\n", __func__, mstate->get_status());
+                    LLAMA_LOG_ERROR("%s: unexpected memory context status: %d\n", __func__, mctx->get_status());
 
                     return -2;
                 }
@@ -987,7 +1045,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     int64_t n_outputs_prev = 0;
 
     do {
-        const auto & ubatch = mstate->get_ubatch();
+        const auto & ubatch = mctx->get_ubatch();
 
         // count the outputs in this ubatch
         {
@@ -1005,11 +1063,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
             n_outputs = n_outputs_new;
         }
 
-        ggml_backend_sched_reset(sched.get());
-        ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
-
         ggml_status status;
-        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mstate.get(), status);
+        const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
@@ -1018,7 +1073,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
                 pos_min[s] = std::numeric_limits<llama_pos>::max();
             }
 
-            // TODO: fix sequence indexing
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
 
@@ -1126,7 +1180,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         n_outputs_prev += n_outputs;
-    } while (mstate->next());
+    } while (mctx->next());
 
     // set to total number of outputs in the batch, for use in llama_get_logits_ith
     n_outputs = n_outputs_all;
@@ -1150,9 +1204,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
         // make the outputs have the same order they had in the user-provided batch
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
-            const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint64_t n_embd  = model.hparams.n_embd;
-
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
@@ -1168,16 +1219,9 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     continue;
                 }
                 std::swap(out_ids[i], out_ids[j_min]);
-                if (logits_size > 0) {
-                    for (uint32_t k = 0; k < n_vocab; k++) {
-                        std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]);
-                    }
-                }
-                if (embd_size > 0) {
-                    for (uint32_t k = 0; k < n_embd; k++) {
-                        std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]);
-                    }
-                }
+
+                // remember the swaps and apply them lazily upon logits/embeddings access
+                output_swaps.push_back({ i, j_min });
             }
 
             std::fill(output_ids.begin(), output_ids.end(), -1);
@@ -1191,9 +1235,11 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(sched.get());
+    if (!supports_set_rows) {
+        // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+        // overlap with device computation.
+        ggml_backend_sched_reset(sched.get());
+    }
 
     return 0;
 }
@@ -1272,27 +1318,43 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     return n_outputs_max;
 }
 
+void llama_context::output_reorder() {
+    const uint32_t n_vocab = model.vocab.n_tokens();
+    const uint64_t n_embd  = model.hparams.n_embd;
+
+    for (uint32_t s = 0; s < output_swaps.size(); ++s) {
+        const uint32_t i0 = output_swaps[s].i0;
+        const uint32_t i1 = output_swaps[s].i1;
+
+        if (logits_size > 0) {
+            for (uint32_t k = 0; k < n_vocab; k++) {
+                std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]);
+            }
+        }
+
+        if (embd_size > 0) {
+            for (uint32_t k = 0; k < n_embd; k++) {
+                std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
+            }
+        }
+    }
+
+    output_swaps.clear();
+}
+
 //
 // graph
 //
 
-int32_t llama_context::graph_max_nodes() const {
-    return std::max<int32_t>(65536, 5*model.n_tensors());
+uint32_t llama_context::graph_max_nodes() const {
+    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
-ggml_cgraph * llama_context::graph_init() {
-    ggml_init_params params = {
-        /*.mem_size   =*/ buf_compute_meta.size(),
-        /*.mem_buffer =*/ buf_compute_meta.data(),
-        /*.no_alloc   =*/ true,
-    };
-
-    ctx_compute.reset(ggml_init(params));
-
-    return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false);
+llm_graph_result * llama_context::get_gf_res_reserve() const {
+    return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate) {
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
 
     if (n_tokens % n_seqs != 0) {
@@ -1302,6 +1364,11 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
     }
 
+    ggml_backend_sched_reset(sched.get());
+
+    // when the scheduler is reset, we cannnot reuse the old graph, so we reset the previous graph result to prevent that
+    gf_res_prev->reset();
+
     // store the n_outputs as it is, and restore it afterwards
     // TODO: not sure if needed, might simplify in the future by removing this
     const auto save_n_outputs = this->n_outputs;
@@ -1311,18 +1378,16 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
     llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
 
-    auto * gf = graph_init();
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate);
+    auto * res = gf_res_reserve.get();
+
+    const auto gparams = graph_params(res, ubatch, mctx, LLM_GRAPH_TYPE_DEFAULT);
+
+    res->reset();
+
+    auto * gf = model.build_graph(gparams);
 
     this->n_outputs = save_n_outputs;
 
-    if (!res) {
-        LLAMA_LOG_ERROR("%s: failed to build worst-case graph\n", __func__);
-        return nullptr;
-    }
-
-    ggml_backend_sched_reset(sched.get());
-
     // initialize scheduler with the specified graph
     if (!ggml_backend_sched_reserve(sched.get(), gf)) {
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
@@ -1332,28 +1397,27 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     return gf;
 }
 
-llm_graph_result_ptr llama_context::graph_build(
-                    ggml_context * ctx,
-                     ggml_cgraph * gf,
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-      const llama_memory_state_i * mstate) {
-    return model.build_graph(
-            {
-                /*.ctx         =*/ ctx,
-                /*.arch        =*/ model.arch,
-                /*.hparams     =*/ model.hparams,
-                /*.cparams     =*/ cparams,
-                /*.ubatch      =*/ ubatch,
-                /*.sched       =*/ sched.get(),
-                /*.backend_cpu =*/ backend_cpu,
-                /*.cvec        =*/ &cvec,
-                /*.loras       =*/ &loras,
-                /*.mstate      =*/ mstate,
-                /*.cross       =*/ &cross,
-                /*.n_outputs   =*/ n_outputs,
-                /*.cb          =*/ graph_get_cb(),
-            }, gf, gtype);
+llm_graph_params llama_context::graph_params(
+                        llm_graph_result * res,
+                      const llama_ubatch & ubatch,
+            const llama_memory_context_i * mctx,
+            llm_graph_type   gtype) const {
+    return {
+        /*.arch        =*/ model.arch,
+        /*.hparams     =*/ model.hparams,
+        /*.cparams     =*/ cparams,
+        /*.ubatch      =*/ ubatch,
+        /*.gtype       =*/ gtype,
+        /*.sched       =*/ sched.get(),
+        /*.backend_cpu =*/ backend_cpu,
+        /*.cvec        =*/ &cvec,
+        /*.loras       =*/ &loras,
+        /*.mctx        =*/ mctx,
+        /*.cross       =*/ &cross,
+        /*.n_outputs   =*/ n_outputs,
+        /*.cb          =*/ graph_get_cb(),
+        /*.res         =*/ res,
+    };
 }
 
 ggml_status llama_context::graph_compute(
@@ -1931,6 +1995,7 @@ llama_perf_context_data llama_context::perf_get_data() const {
     data.t_eval_ms   = 1e-3 * t_eval_us;
     data.n_p_eval    = std::max(1, n_p_eval);
     data.n_eval      = std::max(1, n_eval);
+    data.n_reused    = std::max(0, n_reused);
 
     return data;
 }
@@ -1939,6 +2004,7 @@ void llama_context::perf_reset() {
     t_start_us  = ggml_time_us();
     t_eval_us   = n_eval = 0;
     t_p_eval_us = n_p_eval = 0;
+    n_reused    = 0;
 }
 
 //
@@ -2029,7 +2095,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2042,8 +2108,8 @@ void llama_context::opt_epoch_iter(
 
         uint32_t n_outputs_all = n_tokens_all;
 
-        auto mstate = memory->init_batch(*balloc, cparams.n_ubatch, true);
-        if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
+        auto mctx = memory->init_batch(*balloc, cparams.n_ubatch, true);
+        if (!mctx || mctx->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
         }
@@ -2056,17 +2122,22 @@ void llama_context::opt_epoch_iter(
 
         uint32_t pos_batch = 0;
         do {
-            const auto & ubatch = mstate->get_ubatch();
+            const auto & ubatch = mctx->get_ubatch();
 
             n_outputs = ubatch.n_tokens;
 
-            if (!mstate->apply()) {
-                LLAMA_LOG_ERROR("%s: failed to update the memory state\n", __func__);
+            if (!mctx->apply()) {
+                LLAMA_LOG_ERROR("%s: failed to update the memory context\n", __func__);
                 break;
             }
 
-            auto * gf = graph_init();
-            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate.get());
+            auto * res = gf_res_prev.get();
+
+            const auto gparams = graph_params(res, ubatch, mctx.get(), LLM_GRAPH_TYPE_DEFAULT);
+
+            res->reset();
+
+            auto * gf = model.build_graph(gparams);
 
             struct ggml_context * ctx_compute_opt;
             {
@@ -2101,7 +2172,7 @@ void llama_context::opt_epoch_iter(
             ggml_free(ctx_compute_opt);
 
             pos_batch += ubatch.n_tokens;
-        } while (mstate->next());
+        } while (mctx->next());
     }
 }
 
@@ -2188,6 +2259,7 @@ llama_context_params llama_context_default_params() {
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
+        /*.kv_unified                  =*/ false,
     };
 
     return result;
@@ -2808,6 +2880,7 @@ void llama_perf_context_print(const llama_context * ctx) {
     LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
     LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+    LLAMA_LOG_INFO("%s:    graphs reused = %10d\n", __func__, data.n_reused);
 }
 
 void llama_perf_context_reset(llama_context * ctx) {

examples/talk-llama/llama-context.h

@@ -18,7 +18,7 @@ class llama_io_read_i;
 class llama_io_write_i;
 
 struct llama_memory_i;
-struct llama_memory_state_i;
+struct llama_memory_context_i;
 
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
@@ -35,8 +35,6 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    ggml_context * get_ctx_compute() const;
-
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
     uint32_t n_batch()       const;
@@ -93,14 +91,14 @@ struct llama_context {
                 int32_t   il_end);
 
     // process a single ubatch with a specific graph type
-    // if memory_state is provided, it will be applied first to the context's memory
+    // if memory_context is provided, it will be applied first to the context's memory
     // ret contains the status of the graph computation
     // returns nullptr only if ret != GGML_STATUS_SUCCESS
-    llm_graph_result_ptr process_ubatch(
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-            llama_memory_state_i * mstate,
-                     ggml_status & ret);
+    llm_graph_result * process_ubatch(
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+            llama_memory_context_i * mctx,
+                       ggml_status & ret);
 
     int encode(const llama_batch & batch_inp);
     int decode(const llama_batch & batch_inp);
@@ -183,29 +181,30 @@ private:
     // Returns max number of outputs for which space was reserved.
     uint32_t output_reserve(int32_t n_outputs);
 
+    void output_reorder();
+
     //
     // graph
     //
 
 public:
-    int32_t graph_max_nodes() const;
+    uint32_t graph_max_nodes() const;
 
-    // zero-out inputs and create the ctx_compute for the compute graph
-    ggml_cgraph * graph_init();
+    // can reuse the llm_graph_result instance of the context (for example to update a memory module)
+    llm_graph_result * get_gf_res_reserve() const;
 
     // returns the result of ggml_backend_sched_graph_compute_async execution
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
 
 private:
-    llm_graph_result_ptr graph_build(
-                    ggml_context * ctx,
-                     ggml_cgraph * gf,
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-      const llama_memory_state_i * mstate);
+    llm_graph_params graph_params(
+                        llm_graph_result * res,
+                      const llama_ubatch & ubatch,
+            const llama_memory_context_i * mctx,
+                          llm_graph_type   gtype) const;
 
     llm_graph_cb graph_get_cb() const;
 
@@ -253,13 +252,18 @@ private:
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 
+    struct swap_info {
+        uint32_t i0;
+        uint32_t i1;
+    };
+
+    std::vector<swap_info> output_swaps;
+
     ggml_backend_sched_ptr sched;
 
     ggml_backend_t backend_cpu = nullptr;
     std::vector<ggml_backend_ptr> backends;
 
-    ggml_context_ptr ctx_compute;
-
     // training
     ggml_opt_context_t opt_ctx = nullptr;
 
@@ -275,14 +279,18 @@ private:
     std::vector<ggml_backend_t>             backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    // memory buffers used to evaluate the model
-    std::vector<uint8_t> buf_compute_meta;
+    llm_graph_result_ptr gf_res_prev;
+    llm_graph_result_ptr gf_res_reserve;
 
     // host buffer for the model output (logits and embeddings)
     ggml_backend_buffer_ptr buf_output;
 
     bool has_evaluated_once = false;
 
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    bool supports_set_rows = false;
+
     // perf
     mutable int64_t t_start_us  = 0;
     mutable int64_t t_load_us   = 0;
@@ -294,4 +302,6 @@ private:
 
     mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
     mutable int32_t n_eval   = 0; // number of eval calls
+
+    mutable int32_t n_reused = 0; // number of times the previous graph was reused
 };

examples/talk-llama/llama-cparams.h

@@ -11,8 +11,8 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
-    int      n_threads;       // number of threads to use for generation
-    int      n_threads_batch; // number of threads to use for batch processing
+    int32_t  n_threads;       // number of threads to use for generation
+    int32_t  n_threads_batch; // number of threads to use for batch processing
 
     float rope_freq_base;
     float rope_freq_scale;
@@ -33,6 +33,7 @@ struct llama_cparams {
     bool no_perf;
     bool warmup;
     bool op_offload;
+    bool kv_unified;
 
     enum llama_pooling_type pooling_type;
 

examples/talk-llama/llama-graph.cpp

@@ -28,6 +28,15 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
+    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[0] == params.ubatch.n_tokens);
+
+    return res;
+}
+
 void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -50,6 +59,14 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= pos->ne[0] == params.ubatch.n_tokens;
+
+    return res;
+}
+
 void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && attn_scale) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -71,7 +88,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
         const int64_t n_tokens = ubatch->n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer));
-        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+        GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
         int32_t * data = (int32_t *) pos_bucket->data;
 
@@ -87,7 +104,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
 
 void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
     if (pos_bucket) {
-        kv_state->set_input_pos_bucket(pos_bucket, ubatch);
+        mctx->set_input_pos_bucket(pos_bucket, ubatch);
     }
 }
 
@@ -118,6 +135,14 @@ void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= n_outputs == params.n_outputs;
+
+    return res;
+}
+
 void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
         const int64_t n_tokens     = ubatch->n_tokens;
@@ -221,7 +246,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
 void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_rs = mem_state->get_n_rs();
+    const int64_t n_rs = mctx->get_n_rs();
 
     if (s_copy) {
         GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
@@ -229,7 +254,7 @@ void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
 
         // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
         for (uint32_t i = 0; i < n_rs; ++i) {
-            data[i] = mem_state->s_copy(i);
+            data[i] = mctx->s_copy(i);
         }
     }
 }
@@ -281,19 +306,64 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        kv_state->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->set_input_v_idxs(self_v_idxs, ubatch);
+
+    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+}
+
+bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_unified_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
+    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= mctx->get_supports_set_rows(); // TODO: tmp
+
+    return res;
 }
 
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        kv_state->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
 
-    if (self_kq_mask_swa) {
-        kv_state->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+
+    mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
+    mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);
+
+    mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
+}
+
+bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_unified_iswa_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
+    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
+    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= mctx->get_base()->get_supports_set_rows(); // TODO: tmp
+
+    return res;
 }
 
 void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
@@ -303,7 +373,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     const int64_t n_tokens = ubatch->n_tokens;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
-    GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+    GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
     float * data = (float *) cross_kq_mask->data;
 
@@ -333,21 +403,93 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mem_state->get_state_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    inp_attn->set_input(ubatch);
+    inp_rs->set_input(ubatch);
+}
+
+//
+// llm_graph_result
+//
+
+llm_graph_result::llm_graph_result(int64_t max_nodes) : max_nodes(max_nodes) {
+    reset();
+
+    const char * LLAMA_GRAPH_RESULT_DEBUG = getenv("LLAMA_GRAPH_RESULT_DEBUG");
+    debug = LLAMA_GRAPH_RESULT_DEBUG ? atoi(LLAMA_GRAPH_RESULT_DEBUG) : 0;
+}
+
+int64_t llm_graph_result::get_max_nodes() const {
+    return max_nodes;
+}
+
+void llm_graph_result::reset() {
+    t_tokens      = nullptr;
+    t_logits      = nullptr;
+    t_embd        = nullptr;
+    t_embd_pooled = nullptr;
+
+    params = {};
+
+    inputs.clear();
+
+    buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
+
+    ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute_meta.size(),
+        /*.mem_buffer =*/ buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    ctx_compute.reset(ggml_init(params));
+
+    gf = ggml_new_graph_custom(ctx_compute.get(), max_nodes, false);
+}
+
+void llm_graph_result::set_inputs(const llama_ubatch * ubatch) {
+    for (auto & input : inputs) {
+        input->set_input(ubatch);
     }
+}
 
-    const int64_t n_rs = mem_state->get_state_recr()->get_n_rs();
-
-    if (s_copy) {
-        GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
-        int32_t * data = (int32_t *) s_copy->data;
-
-        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
-        for (uint32_t i = 0; i < n_rs; ++i) {
-            data[i] = mem_state->get_state_recr()->s_copy(i);
+bool llm_graph_result::can_reuse(const llm_graph_params & params) {
+    if (!this->params.allow_reuse(params)) {
+        if (debug > 1) {
+            LLAMA_LOG_DEBUG("%s: cannot reuse graph due to incompatible graph parameters\n", __func__);
         }
+
+        return false;
     }
+
+    if (debug > 1) {
+        LLAMA_LOG_DEBUG("%s: checking compatibility of %d inputs:\n", __func__, (int) inputs.size());
+    }
+
+    bool res = true;
+
+    for (auto & input : inputs) {
+        const bool cur = input->can_reuse(params);
+
+        if (debug > 1) {
+            LLAMA_LOG_DEBUG("%s: can_reuse = %d\n", "placeholder", cur);
+        }
+
+        res = res && cur;
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: can reuse graph = %d\n", __func__, res);
+    }
+
+    return res;
+}
+
+llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) {
+    inputs.emplace_back(std::move(input));
+    return inputs.back().get();
+}
+
+void llm_graph_result::set_params(const llm_graph_params & params) {
+    this->params = params;
 }
 
 //
@@ -384,15 +526,17 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     n_ctx_orig       (cparams.n_ctx_orig_yarn),
     pooling_type     (cparams.pooling_type),
     rope_type        (hparams.rope_type),
-    ctx0             (params.ctx),
     sched            (params.sched),
     backend_cpu      (params.backend_cpu),
     cvec             (params.cvec),
     loras            (params.loras),
-    mstate           (params.mstate),
+    mctx             (params.mctx),
     cross            (params.cross),
     cb_func          (params.cb),
-    res              (std::make_unique<llm_graph_result>()) {
+    res              (params.res),
+    ctx0             (res->get_ctx()),
+    gf               (res->get_gf()) {
+        res->set_params(params);
     }
 
 void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
@@ -554,12 +698,20 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     switch (type_op) {
         case LLM_FFN_SILU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_swiglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_swiglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_silu(ctx0, cur);
                 cb(cur, "ffn_silu", il);
             } break;
         case LLM_FFN_GELU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_geglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_geglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_gelu", il);
                 if (act_scales != NULL) {
@@ -568,7 +720,11 @@ ggml_tensor * llm_graph_context::build_ffn(
                 }
             } break;
         case LLM_FFN_RELU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_reglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_reglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_relu", il);
             } break;
@@ -582,32 +738,19 @@ ggml_tensor * llm_graph_context::build_ffn(
             } break;
         case LLM_FFN_SWIGLU:
             {
-                // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
-                int64_t split_point = cur->ne[0] / 2;
-                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
-                ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
-                ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
-
-                x0 = ggml_silu(ctx0, x0);
-                cb(cur, "ffn_silu", il);
-
-                cur = ggml_mul(ctx0, x0, x1);
-                cb(cur, "ffn_mul", il);
+                cur = ggml_swiglu(ctx0, cur);
+                cb(cur, "ffn_swiglu", il);
             } break;
         case LLM_FFN_GEGLU:
             {
-                // Split into two equal parts
-                int64_t split_point = cur->ne[0] / 2;
-                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
-                ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
-                ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
-
-                x0 = ggml_gelu(ctx0, x0);
-                cb(x0, "ffn_gelu", il);
-
-                cur = ggml_mul(ctx0, x0, x1);
+                cur = ggml_geglu(ctx0, cur);
                 cb(cur, "ffn_geglu", il);
             } break;
+        case LLM_FFN_REGLU:
+            {
+                cur = ggml_reglu(ctx0, cur);
+                cb(cur, "ffn_reglu", il);
+            } break;
     }
 
     if (gate && type_gate == LLM_FFN_PAR) {
@@ -737,12 +880,18 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
     switch (type_op) {
         case LLM_FFN_SILU:
-            {
+            if (gate_exps) {
+                cur = ggml_swiglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_swiglu", il);
+            } else {
                 cur = ggml_silu(ctx0, cur);
                 cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
-            {
+            if (gate_exps) {
+                cur = ggml_geglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_geglu", il);
+            } else {
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_moe_gelu", il);
             } break;
@@ -750,11 +899,6 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
-    if (gate_exps) {
-        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
-        cb(cur, "ffn_moe_gate_par", il);
-    }
-
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -763,20 +907,28 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(cur, "ffn_moe_weighted", il);
     }
 
-    // aggregate experts
-    ggml_tensor * moe_out = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
-        ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
+    ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };
 
-        if (i == 0) {
-            moe_out = cur_expert;
-        } else {
-            moe_out = ggml_add(ctx0, moe_out, cur_expert);
-        }
+    assert(n_expert_used > 0);
+
+    // order the views before the adds
+    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
+        cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
+
+        ggml_build_forward_expand(gf, cur_experts[i]);
     }
 
-    if (n_expert_used == 1) {
+    // aggregate experts
+    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
+    //       to avoid potentially a large number of add nodes during warmup
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
+    ggml_tensor * moe_out = cur_experts[0];
+
+    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
+        moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
+    }
+
+    if (hparams.n_expert_used == 1) {
         // avoid returning a non-contiguous tensor
         moe_out = ggml_cont(ctx0, moe_out);
     }
@@ -950,11 +1102,11 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_state);
+    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);
 
-    const auto n_kv = kv_state->get_n_kv();
+    const auto n_kv = mctx_cur->get_n_kv();
 
     auto & cur = inp->pos_bucket;
 
@@ -981,35 +1133,7 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t
     return pos_bias;
 }
 
-llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
-    const auto * mem_state = static_cast<const llama_memory_hybrid_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(hparams, cparams, mem_state);
-
-    {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
-
-        const auto n_kv = inp->mem_state->get_state_attn()->get_n_kv();
-
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
-        ggml_set_input(inp->self_kq_mask);
-
-        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
-    }
-
-    {
-        const auto n_rs = mem_state->get_state_recr()->get_n_rs();
-
-        inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
-        ggml_set_input(inp->s_copy);
-    }
-
-    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
-}
-
 ggml_tensor * llm_graph_context::build_attn_mha(
-         ggml_cgraph * gf,
          ggml_tensor * q,
          ggml_tensor * k,
          ggml_tensor * v,
@@ -1019,13 +1143,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
+    // split the batch into streams if needed
+    const auto n_stream = k->ne[3];
+
+    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
+
     q = ggml_permute(ctx0, q, 0, 2, 1, 3);
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
     v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
-    const auto n_tokens = q->ne[1];
-    const auto n_head   = q->ne[2];
-    const auto n_kv     = k->ne[1];
+    const auto n_kv = k->ne[1];
 
     ggml_tensor * cur;
 
@@ -1067,7 +1194,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 #endif
         }
 
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1112,7 +1239,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
 
-        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        // recombine streams
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1129,8 +1257,7 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-    //cb(inp_kq_mask, "KQ_mask", -1);
+    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
     ggml_set_input(inp->kq_mask);
 
     inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;
@@ -1140,7 +1267,6 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_no_cache * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1160,11 +1286,15 @@ ggml_tensor * llm_graph_context::build_attn(
 
     const auto & kq_mask = inp->get_kq_mask();
 
+    // [TAG_NO_CACHE_PAD]
+    // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
+    assert(!ubatch.equal_seqs());
+
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1182,29 +1312,44 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unified_impl(
+           ggml_context * ctx0,
+     const llama_ubatch & ubatch,
+    const llama_hparams & hparams,
+    const llama_cparams & cparams,
+    const llama_kv_cache_unified_context * mctx_cur) {
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_state);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, mctx_cur);
 
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
 
-        const auto n_kv = kv_state->get_n_kv();
+        const auto n_kv     = mctx_cur->get_n_kv();
+        const auto n_tokens = ubatch.n_tokens;
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
     }
 
+    return inp;
+}
+
+llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+
+    auto inp = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+
     return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1220,21 +1365,24 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+    const auto * mctx_cur = inp->mctx;
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
 
     ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1254,7 +1402,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1267,28 +1414,41 @@ ggml_tensor * llm_graph_context::build_attn(
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
     ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
-    ggml_build_forward_expand(gf, v_cur);
 
-    const auto * kv_state_iswa = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+    if (k_cur) {
+        ggml_build_forward_expand(gf, k_cur);
+    }
+
+    if (v_cur) {
+        ggml_build_forward_expand(gf, v_cur);
+    }
+
+    const auto * mctx_iswa = inp->mctx;
 
     const bool is_swa = hparams.is_swa(il);
 
-    const auto * kv_state = is_swa ? kv_state_iswa->get_swa() : kv_state_iswa->get_base();
+    const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base();
 
-    // store to KV cache
-    {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+    // optionally store to KV cache
+    if (k_cur) {
+        const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+    }
+
+    if (v_cur) {
+        const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
     ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1311,7 +1471,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
     ggml_set_input(inp->cross_kq_mask);
 
     inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
@@ -1321,7 +1481,6 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_cross * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1343,7 +1502,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1361,66 +1520,23 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_mem_hybrid * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * wo,
-        ggml_tensor * wo_b,
-        ggml_tensor * q_cur,
-        ggml_tensor * k_cur,
-        ggml_tensor * v_cur,
-        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
-            float     kq_scale,
-            int       il) const {
-    // these nodes are added to the graph together so that they are not reordered
-    // by doing so, the number of splits in the graph is reduced
-    ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
-    ggml_build_forward_expand(gf, v_cur);
-
-    const auto * kv_state = static_cast<const llama_memory_hybrid_state *>(mstate)->get_state_attn();
-
-    // store to KV cache
-    {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
-    }
-
-    const auto & kq_mask = inp->get_kq_mask();
-
-    ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
-    cb(cur, "kqv_out", il);
-
-    if (wo) {
-        cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
-        }
-    }
-
-    if (wo_b) {
-        cur = ggml_add(ctx0, cur, wo_b);
-    }
-
-    return cur;
-}
-
+// TODO: maybe separate the inner implementation into a separate function
+//       like with the non-sliding window equivalent
+//       once sliding-window hybrid caches are a thing.
 llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_state);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
+
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
     {
-        const auto n_kv = kv_state->get_base()->get_n_kv();
+        const auto n_kv = mctx_cur->get_base()->get_n_kv();
 
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1429,10 +1545,12 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
     {
         GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
 
-        const auto n_kv = kv_state->get_swa()->get_n_kv();
+        const auto n_kv = mctx_cur->get_swa()->get_n_kv();
 
-        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1442,7 +1560,6 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 }
 
 ggml_tensor * llm_graph_context::build_rs(
-        ggml_cgraph * gf,
         ggml_tensor * s,
         ggml_tensor * state_copy,
             int32_t   state_size,
@@ -1451,7 +1568,7 @@ ggml_tensor * llm_graph_context::build_rs(
            uint32_t   kv_head,
            uint32_t   kv_size,
             int32_t   rs_zero,
-               bool   avoid_copies) const {
+        const llm_graph_get_rows_fn & get_state_rows) const {
 
     ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
 
@@ -1460,19 +1577,11 @@ ggml_tensor * llm_graph_context::build_rs(
     ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
     ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
 
-    ggml_tensor * output_states;
-
-    if (!avoid_copies) {
-        // copy states
-        // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-        // {state_size, kv_size} -> {state_size, n_seqs}
-        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
-        ggml_build_forward_expand(gf, output_states);
-    } else {
-        // FIXME: make the gathering operation happen before the copy below
-        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
-        output_states = states;
-    }
+    // copy states
+    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
+    // {state_size, kv_size} -> {state_size, n_seqs}
+    ggml_tensor * output_states = get_state_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+    ggml_build_forward_expand(gf, output_states);
 
     // copy extra states which won't be changed further (between n_seqs and n_kv)
     ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
@@ -1484,58 +1593,53 @@ ggml_tensor * llm_graph_context::build_rs(
     return output_states;
 }
 
-llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
+           ggml_context * ctx0,
+    const llama_memory_recurrent_context * mctx_cur) {
 
-    auto inp = std::make_unique<llm_graph_input_rs>(kv_state);
+    auto inp = std::make_unique<llm_graph_input_rs>(mctx_cur);
 
-    const auto n_rs = kv_state->get_n_rs();
+    const auto n_rs = mctx_cur->get_n_rs();
 
     inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
     ggml_set_input(inp->s_copy);
 
+    return inp;
+}
+
+llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+    auto inp = build_rs_inp_impl(ctx0, mctx_cur);
+
     return (llm_graph_input_rs *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_rs(
         llm_graph_input_rs * inp,
-        ggml_cgraph * gf,
         ggml_tensor * s,
             int32_t   state_size,
             int32_t   n_seqs,
-               bool   avoid_copies) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+        const llm_graph_get_rows_fn & get_state_rows) const {
+    const auto * kv_state = inp->mctx;
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies);
-}
-
-ggml_tensor * llm_graph_context::build_rs(
-        llm_graph_input_mem_hybrid * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * s,
-            int32_t   state_size,
-            int32_t   n_seqs,
-               bool   avoid_copies) const {
-    const auto * kv_state = static_cast<const llama_memory_hybrid_state *>(mstate)->get_state_recr();
-
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies);
+    return build_rs(s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
     llm_graph_input_rs * inp,
-           ggml_cgraph * gf,
     const llama_ubatch & ubatch,
-                 int   il) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+                   int   il) const {
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     const auto token_shift_count = hparams.token_shift_count;
 
     const int64_t n_seqs  = ubatch.n_seqs;
 
-    ggml_tensor * token_shift_all = kv_state->get_r_l(il);
+    ggml_tensor * token_shift_all = mctx_cur->get_r_l(il);
 
     ggml_tensor * token_shift = build_rs(
-            inp, gf, token_shift_all,
+            inp, token_shift_all,
             hparams.n_embd_r(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
@@ -1547,24 +1651,34 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;
 
     const int64_t n_seqs = ubatch.n_seqs;
 
-    const auto kv_head = kv_state->get_head();
+    const auto kv_head = mctx_cur->get_head();
 
     return ggml_cpy(
         ctx0,
         ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
-        ggml_view_1d(ctx0, kv_state->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(kv_state->get_r_l(il)))
+        ggml_view_1d(ctx0, mctx_cur->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(mctx_cur->get_r_l(il)))
     );
 }
 
+llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
+    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
+
+    auto inp_rs   = build_rs_inp_impl(ctx0, mctx_cur->get_recr());
+    auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
+
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
+
+    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
+}
+
 void llm_graph_context::build_pooling(
-        ggml_cgraph * gf,
         ggml_tensor * cls,
         ggml_tensor * cls_b,
         ggml_tensor * cls_out,

examples/talk-llama/llama-graph.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "llama-arch.h"
+#include "llama-batch.h"
 #include "llama-hparams.h"
 #include "llama-adapter.h"
 
@@ -14,15 +15,14 @@ struct ggml_cgraph;
 struct ggml_context;
 struct ggml_tensor;
 
-struct llama_ubatch;
 struct llama_cparams;
 
-struct llama_memory_state_i;
+struct llama_memory_context_i;
 
-class llama_kv_cache_unified_state;
-class llama_kv_cache_unified_iswa_state;
-class llama_memory_recurrent_state;
-class llama_memory_hybrid_state;
+class llama_kv_cache_unified_context;
+class llama_kv_cache_unified_iswa_context;
+class llama_memory_recurrent_context;
+class llama_memory_hybrid_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -38,6 +38,7 @@ enum llm_ffn_op_type {
     LLM_FFN_RELU_SQR,
     LLM_FFN_SWIGLU,
     LLM_FFN_GEGLU,
+    LLM_FFN_REGLU,
 };
 
 enum llm_ffn_gate_type {
@@ -68,6 +69,8 @@ struct llama_cross {
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
 };
 
+struct llm_graph_params;
+
 //
 // llm_graph_input
 //
@@ -77,11 +80,19 @@ public:
     virtual ~llm_graph_input_i() = default;
 
     virtual void set_input(const llama_ubatch * ubatch) = 0;
+
+    // return true if the resulting input tensors using the provided graph parameters would be
+    //   the same as the previous input tensors that we have currently stored in the object
+    virtual bool can_reuse(const llm_graph_params & params) {
+        // returning false here by default will prevent from reusing the graph if the check
+        //   for the input type has not been implemented yet
+        GGML_UNUSED(params);
+        return false;
+    }
 };
 
 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
-
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
     llm_graph_input_embd()          = default;
@@ -89,6 +100,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
 };
@@ -100,6 +113,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
     const uint32_t n_pos_per_embd = 1;
@@ -136,7 +151,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 public:
     llm_graph_input_pos_bucket_kv(
             const llama_hparams & hparams,
-            const llama_kv_cache_unified_state * kv_state) : hparams(hparams), kv_state(kv_state) {}
+            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -144,7 +159,8 @@ public:
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
     const llama_hparams & hparams;
-    const llama_kv_cache_unified_state * kv_state;
+
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -152,17 +168,19 @@ public:
     llm_graph_input_out_ids(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
+            uint32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
     virtual ~llm_graph_input_out_ids() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * out_ids; // I32 [n_outputs]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const int32_t n_outputs;
+    const uint32_t n_outputs;
 };
 
 class llm_graph_input_mean : public llm_graph_input_i {
@@ -191,14 +209,14 @@ public:
 
 class llm_graph_input_rs : public llm_graph_input_i {
 public:
-    llm_graph_input_rs(const llama_memory_recurrent_state * mem_state) : mem_state(mem_state) {}
+    llm_graph_input_rs(const llama_memory_recurrent_context * mctx) : mctx(mctx) {}
     virtual ~llm_graph_input_rs() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_memory_recurrent_state * mem_state;
+    const llama_memory_recurrent_context * mctx;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -226,8 +244,8 @@ public:
 
     ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
 
-    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch]
-    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch]
+    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
+    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -238,24 +256,32 @@ public:
     llm_graph_input_attn_kv_unified(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_state * kv_state) :
+            const llama_kv_cache_unified_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_state(kv_state) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified_state * kv_state;
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
@@ -263,27 +289,39 @@ public:
     llm_graph_input_attn_kv_unified_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_state * kv_state) :
+            const llama_kv_cache_unified_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_state(kv_state) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified_iswa() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
+    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
+    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }
+
     ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
+    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
+
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified_iswa_state * kv_state;
+    const llama_kv_cache_unified_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -295,8 +333,8 @@ public:
 
     ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
 
-    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch]
-    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
 
     const llama_cross * cross = nullptr;
 };
@@ -304,28 +342,23 @@ public:
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
     llm_graph_input_mem_hybrid(
-            const llama_hparams & hparams,
-            const llama_cparams & cparams,
-            const llama_memory_hybrid_state * mem_state) :
-        hparams(hparams),
-        cparams(cparams),
-        mem_state(mem_state) {
-    }
+            std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
+            std::unique_ptr<llm_graph_input_rs>              inp_rs,
+            const llama_memory_hybrid_context *              mctx) :
+        inp_attn(std::move(inp_attn)),
+        inp_rs(std::move(inp_rs)),
+        mctx(mctx) { }
     virtual ~llm_graph_input_mem_hybrid() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * s_copy; // I32 [kv_size]
+    std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>              inp_rs;
 
-    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+    llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs              * get_recr() const { return inp_rs.get(); }
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
-
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
-
-    const llama_memory_hybrid_state * mem_state;
+    const llama_memory_hybrid_context * mctx;
 };
 
 //
@@ -338,40 +371,108 @@ public:
 // along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.
 //   these are used by the llama_context to extact the relevant data, based on the compute parameters
 
-class llm_graph_result_i {
-public:
-    virtual ~llm_graph_result_i() = default;
+// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
+using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
 
-    virtual ggml_tensor * get_tokens()      = 0;
-    virtual ggml_tensor * get_logits()      = 0;
-    virtual ggml_tensor * get_embd()        = 0;
-    virtual ggml_tensor * get_embd_pooled() = 0;
+class llm_graph_result;
 
-    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
+struct llm_graph_params {
+    llm_arch arch = LLM_ARCH_UNKNOWN;
+
+    llama_hparams hparams;
+    llama_cparams cparams;
+
+    llama_ubatch ubatch; // note: intentionally make a copy
+
+    llm_graph_type gtype;
+
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
+
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
+
+    uint32_t n_outputs;
+
+    llm_graph_cb cb;
+
+    llm_graph_result * res;
+
+    // return true if the "other" params would result in a graph with the same topology as with the current params
+    //   having the same topology allows us to reuse the graph in some cases
+    bool allow_reuse(const llm_graph_params & other) const {
+        // first check the ubatch
+        bool can_reuse_ubatch =
+            ubatch.equal_seqs() == other.ubatch.equal_seqs() &&
+            ubatch.n_tokens     == other.ubatch.n_tokens &&
+            ubatch.n_seq_tokens == other.ubatch.n_seq_tokens &&
+            ubatch.n_seqs       == other.ubatch.n_seqs &&
+            ubatch.n_seqs_unq   == other.ubatch.n_seqs_unq &&
+            (
+                (!ubatch.token && !other.ubatch.token) ||
+                (!ubatch.embd  && !other.ubatch.embd)
+            );
+
+        if (can_reuse_ubatch && !ubatch.equal_seqs()) {
+            if (!ubatch.data) {
+                // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
+                //   therefore we cannot perform the sequence id check. normally should never happen
+                can_reuse_ubatch = false;
+            } else {
+                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                    can_reuse_ubatch &= ubatch.seq_id_unq[s] == other.ubatch.seq_id_unq[s];
+                }
+            }
+        }
+
+        if (!can_reuse_ubatch) {
+            return false;
+        }
+
+        return
+            cparams.embeddings  == other.cparams.embeddings  &&
+            cparams.causal_attn == other.cparams.causal_attn &&
+            arch      == other.arch  &&
+            gtype     == other.gtype &&
+            cvec      == other.cvec  &&
+            loras     == other.loras &&
+            cross     == other.cross &&
+            n_outputs == other.n_outputs;
+    }
 };
 
-using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
-
-
-class llm_graph_result : public llm_graph_result_i {
+class llm_graph_result {
 public:
+    llm_graph_result(int64_t max_nodes);
+
     virtual ~llm_graph_result() = default;
 
-    ggml_tensor * get_tokens()      override { return t_tokens; }
-    ggml_tensor * get_logits()      override { return t_logits; }
-    ggml_tensor * get_embd()        override { return t_embd; }
-    ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
+    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_logits()      const { return t_logits; }
+    ggml_tensor * get_embd()        const { return t_embd; }
+    ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
 
-    void set_inputs(const llama_ubatch * ubatch) override {
-        for (auto & input : inputs) {
-            input->set_input(ubatch);
-        }
-    }
+    ggml_cgraph  * get_gf()  const { return gf; }
+    ggml_context * get_ctx() const { return ctx_compute.get(); }
 
-    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
-        inputs.emplace_back(std::move(input));
-        return inputs.back().get();
-    }
+    int64_t get_max_nodes() const;
+
+    void reset();
+
+    void set_inputs(const llama_ubatch * ubatch);
+
+    // try to update the existing graph result using the new graph parameters in order to reuse it
+    // this can only be done if we determine that the resulting graph using the new graph parameters
+    //   would be identical to the existing graph. in that case, we simply have to update the memory
+    //   contexts of the input tensors of the graph and we can reuse it for another computation
+    // return true if the graph was updated and can be reused
+    bool can_reuse(const llm_graph_params & params);
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input);
+
+    void set_params(const llm_graph_params & params);
 
     // important graph nodes
     ggml_tensor * t_tokens      = nullptr;
@@ -380,36 +481,34 @@ public:
     ggml_tensor * t_embd_pooled = nullptr;
 
     std::vector<llm_graph_input_ptr> inputs;
+
+    ggml_context_ptr ctx_compute;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    ggml_cgraph * gf;
+
+    int64_t max_nodes;
+
+private:
+    // keep a copy of the previous graph parameters
+    // we will use this to determine whether the graph can be reused by comparing them with the new parameters
+    // note: these are updated after constructing the new graph
+    llm_graph_params params;
+
+    // env: LLAMA_GRAPH_RESULT_DEBUG
+    int debug = 0;
 };
 
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result>;
+
 //
 // llm_graph_context
 //
 
-// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
-using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
-
-struct llm_graph_params {
-    ggml_context * ctx;
-
-    const llm_arch arch;
-
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
-    const llama_ubatch  & ubatch;
-
-    ggml_backend_sched_t sched;
-    ggml_backend_t backend_cpu;
-
-    const llama_adapter_cvec   * cvec;
-    const llama_adapter_loras  * loras;
-    const llama_memory_state_i * mstate;
-    const llama_cross          * cross;
-
-    uint32_t n_outputs;
-
-    const llm_graph_cb & cb;
-};
+// used in build_rs to properly order writes and avoid unnecessary copies
+using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;
 
 struct llm_graph_context {
     const llm_arch arch;
@@ -447,22 +546,24 @@ struct llm_graph_context {
     const enum llama_pooling_type pooling_type;
     const enum llama_rope_type    rope_type;
 
-    ggml_context * ctx0 = nullptr;
-
     ggml_backend_sched_t sched;
 
     ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
 
-    const llama_adapter_cvec   * cvec;
-    const llama_adapter_loras  * loras;
-    const llama_memory_state_i * mstate;
-    const llama_cross          * cross;
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
     const llm_graph_cb & cb_func;
 
-    std::unique_ptr<llm_graph_result> res;
+    llm_graph_result * res;
+
+    ggml_context * ctx0 = nullptr;
+    ggml_cgraph  * gf   = nullptr;
 
     llm_graph_context(const llm_graph_params & params);
+    virtual ~llm_graph_context() = default;
 
     void cb(ggml_tensor * cur, const char * name, int il) const;
 
@@ -540,14 +641,11 @@ struct llm_graph_context {
     ggml_tensor * build_inp_pos_bucket_dec() const;
     ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
 
-    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
-
     //
     // attention
     //
 
     ggml_tensor * build_attn_mha(
-             ggml_cgraph * gf,
              ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
              ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
              ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
@@ -560,7 +658,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_no_cache * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -575,7 +672,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -588,14 +684,14 @@ struct llm_graph_context {
 
     llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
 
+    // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified_iswa * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
@@ -605,7 +701,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_cross * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -616,18 +711,6 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
-    ggml_tensor * build_attn(
-            llm_graph_input_mem_hybrid * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * wo,
-            ggml_tensor * wo_b,
-            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
-            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                  float   kq_scale,
-                    int   il) const;
     //
     // recurrent
     //
@@ -639,7 +722,6 @@ struct llm_graph_context {
     //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
     //         `llama_memory_recurrent`
     ggml_tensor * build_rs(
-            ggml_cgraph * gf,
             ggml_tensor * s,
             ggml_tensor * state_copy,
                 int32_t   state_size,
@@ -648,43 +730,37 @@ struct llm_graph_context {
                uint32_t   kv_head,
                uint32_t   kv_size,
                 int32_t   rs_zero,
-                   bool   avoid_copies = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     llm_graph_input_rs * build_rs_inp() const;
 
     ggml_tensor * build_rs(
             llm_graph_input_rs * inp,
-            ggml_cgraph * gf,
             ggml_tensor * s,
                 int32_t   state_size,
                 int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
-
-    ggml_tensor * build_rs(
-            llm_graph_input_mem_hybrid * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * s,
-                int32_t   state_size,
-                int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
         llm_graph_input_rs * inp,
-               ggml_cgraph * gf,
         const llama_ubatch & ubatch,
-                     int   il) const;
+                       int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(
              ggml_tensor * token_shift,
       const llama_ubatch & ubatch,
                      int   il) const;
+    //
+    // hybrid
+    //
+
+    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
 
     //
     // pooling
     //
 
     void build_pooling(
-            ggml_cgraph * gf,
             ggml_tensor * cls,
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,

examples/talk-llama/llama-hparams.cpp

@@ -65,15 +65,61 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
+bool llama_hparams::is_n_embd_k_gqa_variable() const {
+    const uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_k_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool llama_hparams::is_n_embd_v_gqa_variable() const {
+    const uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_v_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa_max() const {
+    uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_k_gqa(il));
+    }
+
+    return val;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa_max() const {
+    uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_v_gqa(il));
+    }
+
+    return val;
+}
+
 uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
     }
 
+    if (n_shortconv_l_cache != 0) {
+        // for LFM2 models
+        return n_embd * (n_shortconv_l_cache - 1);
+    }
+
     // TODO: maybe support other convolution strides than 1
     // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
-    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+    // Corresponds to Mamba's conv_states size
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * (ssm_d_inner + 2*ssm_n_group*ssm_d_state);
 }
 
 uint32_t llama_hparams::n_embd_s() const {

examples/talk-llama/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
@@ -55,6 +55,8 @@ struct llama_hparams {
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;
 
+    uint32_t n_shortconv_l_cache  = 0;
+
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
@@ -96,7 +98,7 @@ struct llama_hparams {
     float    rope_freq_scale_train;
     float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
-    float    rope_yarn_log_mul;
+    float    rope_yarn_log_mul = 0.0f;
 
     std::array<int, 4> rope_sections;
 
@@ -114,6 +116,7 @@ struct llama_hparams {
     uint32_t ssm_d_inner = 0;
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
+    uint32_t ssm_n_group = 0;
 
     // for hybrid state space models
     std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
@@ -143,6 +146,12 @@ struct llama_hparams {
     uint32_t n_attn_temp_floor_scale = 8192;
     float    f_attn_temp_scale       = 0.1;
 
+    // gemma3n altup
+    uint32_t n_altup      = 4; // altup_num_inputs
+    uint32_t i_altup_act  = 0; // altup_active_idx
+    uint32_t laurel_rank  = 64;
+    uint32_t n_embd_altup = 256;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -182,6 +191,14 @@ struct llama_hparams {
     // dimension of value embeddings across all k-v heads
     uint32_t n_embd_v_gqa(uint32_t il = 0) const;
 
+    // true if any layer has a different n_embd_k_gqa/n_embd_v_gqa
+    bool is_n_embd_k_gqa_variable() const;
+    bool is_n_embd_v_gqa_variable() const;
+
+    // return the maximum n_embd_k_gqa/n_embd_v_gqa across all layers
+    uint32_t n_embd_k_gqa_max() const;
+    uint32_t n_embd_v_gqa_max() const;
+
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
     uint32_t n_embd_r() const;

examples/talk-llama/llama-kv-cache-unified-iswa.cpp

@@ -18,16 +18,17 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
                      bool   v_trans,
                      bool   offload,
                      bool   swa_full,
+                     bool   unified,
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
-                 uint32_t   n_pad) : hparams(model.hparams) {
+                 uint32_t   n_pad) : hparams(model.hparams), unified(unified) {
     llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
     llama_kv_cache_unified::layer_filter_cb filter_swa  = [&](int32_t il) { return  model.hparams.is_swa(il); };
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*n_seq_max + n_ubatch, n_pad));
+    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {
@@ -41,14 +42,14 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
 
     kv_base = std::make_unique<llama_kv_cache_unified>(
             model, std::move(filter_base), type_k, type_v,
-            v_trans, offload, size_base, n_seq_max, n_pad,
+            v_trans, offload, unified, size_base, n_seq_max, n_pad,
             0, LLAMA_SWA_TYPE_NONE);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
     kv_swa = std::make_unique<llama_kv_cache_unified>(
             model, std::move(filter_swa), type_k, type_v,
-            v_trans, offload, size_swa, n_seq_max, n_pad,
+            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
             hparams.n_swa, hparams.swa_type);
 }
 
@@ -95,11 +96,16 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
     GGML_UNUSED(embd_all);
 
     // first try simple split
     do {
+        if (!unified) {
+            // requires equal splits, so we skip the simple split
+            break;
+        }
+
         balloc.split_reset();
 
         std::vector<llama_ubatch> ubatches;
@@ -113,20 +119,25 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_alloc
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
-        if (heads_base.empty()) {
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
-        if (heads_swa.empty()) {
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
 
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // if it fails, try equal split
@@ -135,7 +146,7 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_alloc
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch);
+            auto ubatch = balloc.split_equal(n_ubatch, !unified);
 
             if (ubatch.n_tokens == 0) {
                 break;
@@ -144,34 +155,39 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_alloc
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
-        if (heads_base.empty()) {
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
-        if (heads_swa.empty()) {
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
 
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() {
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(this);
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(this, lctx, optimize);
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this, lctx, optimize);
 }
 
 bool llama_kv_cache_unified_iswa::get_can_shift() const {
@@ -197,46 +213,46 @@ llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
 }
 
 //
-// llama_kv_cache_unified_iswa_state
+// llama_kv_cache_unified_iswa_context
 //
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(llama_memory_status status) : status(status) {}
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv) :
-    state_base(kv->get_base()->init_full()),
-    state_swa (kv->get_swa ()->init_full()),
-    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
+    ctx_base(kv->get_base()->init_full()),
+    ctx_swa (kv->get_swa ()->init_full()),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv,
         llama_context * lctx,
         bool optimize) :
-    state_base(kv->get_base()->init_update(lctx, optimize)),
-    state_swa (kv->get_swa ()->init_update(lctx, optimize)),
-    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
+    ctx_base(kv->get_base()->init_update(lctx, optimize)),
+    ctx_swa (kv->get_swa ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv,
-        std::vector<uint32_t> heads_base,
-        std::vector<uint32_t> heads_swa,
+        slot_info_vec_t sinfos_base,
+        slot_info_vec_t sinfos_swa,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    state_base(new llama_kv_cache_unified_state(kv->get_base(), std::move(heads_base), this->ubatches)),
-    state_swa (new llama_kv_cache_unified_state(kv->get_swa (), std::move(heads_swa),  this->ubatches)),
-    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
+    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_state:: ~llama_kv_cache_unified_iswa_state() = default;
+llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default;
 
-bool llama_kv_cache_unified_iswa_state::next() {
+bool llama_kv_cache_unified_iswa_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    state_base->next();
-    state_swa ->next();
+    ctx_base->next();
+    ctx_swa ->next();
 
     if (++i_next >= ubatches.size()) {
         return false;
@@ -245,35 +261,35 @@ bool llama_kv_cache_unified_iswa_state::next() {
     return true;
 }
 
-bool llama_kv_cache_unified_iswa_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_kv_cache_unified_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
 
     bool res = true;
 
-    res = res & state_base->apply();
-    res = res & state_swa ->apply();
+    res = res & ctx_base->apply();
+    res = res & ctx_swa ->apply();
 
     return res;
 }
 
-llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const {
+llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_iswa_state::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_base() const {
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_state *>(state_base.get());
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_base.get());
 }
 
-const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_swa()  const {
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_state *>(state_swa.get());
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_swa.get());
 }

examples/talk-llama/llama-kv-cache-unified-iswa.h

@@ -20,6 +20,7 @@ public:
                          bool   v_trans,
                          bool   offload,
                          bool   swa_full,
+                         bool   unified,
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
@@ -31,14 +32,14 @@ public:
     // llama_memory_i
     //
 
-    llama_memory_state_ptr init_batch(
+    llama_memory_context_ptr init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) override;
 
-    llama_memory_state_ptr init_full() override;
+    llama_memory_context_ptr init_full() override;
 
-    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
 
     bool get_can_shift() const override;
 
@@ -68,36 +69,40 @@ public:
 private:
     const llama_hparams & hparams;
 
+    const bool unified;
+
     std::unique_ptr<llama_kv_cache_unified> kv_base;
     std::unique_ptr<llama_kv_cache_unified> kv_swa;
 };
 
-class llama_kv_cache_unified_iswa_state : public llama_memory_state_i {
+class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
 public:
-    // used for errors
-    llama_kv_cache_unified_iswa_state(llama_memory_status status);
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
 
-    // used to create a full-cache state
-    llama_kv_cache_unified_iswa_state(
+    // used for errors
+    llama_kv_cache_unified_iswa_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv);
 
-    // used to create an update state
-    llama_kv_cache_unified_iswa_state(
+    // used to create an update context
+    llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv,
             llama_context * lctx,
             bool optimize);
 
-    // used to create a state from a batch
-    llama_kv_cache_unified_iswa_state(
+    // used to create a batch processing context from a batch
+    llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv,
-            std::vector<uint32_t> heads_base,
-            std::vector<uint32_t> heads_swa,
+            slot_info_vec_t sinfos_base,
+            slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_iswa_state();
+    virtual ~llama_kv_cache_unified_iswa_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
@@ -107,11 +112,11 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_iswa_state specific API
+    // llama_kv_cache_unified_iswa_context specific API
     //
 
-    const llama_kv_cache_unified_state * get_base() const;
-    const llama_kv_cache_unified_state * get_swa()  const;
+    const llama_kv_cache_unified_context * get_base() const;
+    const llama_kv_cache_unified_context * get_swa()  const;
 
 private:
     //llama_kv_cache_unified_iswa * kv;
@@ -121,8 +126,8 @@ private:
 
     std::vector<llama_ubatch> ubatches;
 
-    const llama_memory_state_ptr state_base;
-    const llama_memory_state_ptr state_swa;
+    const llama_memory_context_ptr ctx_base;
+    const llama_memory_context_ptr ctx_swa;
 
     const llama_memory_status status;
 };

examples/talk-llama/llama-kv-cache-unified.h

@@ -24,8 +24,6 @@ public:
     // this callback is used to filter out layers that should not be included in the cache
     using layer_filter_cb = std::function<bool(int32_t il)>;
 
-    using ubatch_heads = std::vector<uint32_t>;
-
     struct defrag_info {
         bool empty() const {
             return ids.empty();
@@ -37,6 +35,63 @@ public:
         std::vector<uint32_t> ids;
     };
 
+    struct stream_copy_info {
+        bool empty() const {
+            assert(ssrc.size() == sdst.size());
+            return ssrc.empty();
+        }
+
+        std::vector<uint32_t> ssrc;
+        std::vector<uint32_t> sdst;
+    };
+
+    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
+    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
+    struct slot_info {
+        // data for ggml_set_rows
+        using idx_vec_t = std::vector<uint32_t>;
+
+        // number of streams: ns = s1 - s0 + 1
+        llama_seq_id s0;
+        llama_seq_id s1;
+
+        std::vector<llama_seq_id> strm; // [ns]
+        std::vector<idx_vec_t>    idxs; // [ns]
+
+        uint32_t head() const {
+            GGML_ASSERT(idxs.size() == 1);
+            GGML_ASSERT(!idxs[0].empty());
+
+            return idxs[0][0];
+        }
+
+        void resize(size_t n) {
+            strm.resize(n);
+            idxs.resize(n);
+        }
+
+        size_t size() const {
+            GGML_ASSERT(idxs.size() == strm.size());
+            GGML_ASSERT(!idxs.empty());
+
+            return idxs[0].size();
+        }
+
+        size_t n_stream() const {
+            return strm.size();
+        }
+
+        bool empty() const {
+            return idxs.empty();
+        }
+
+        void clear() {
+            idxs.clear();
+        }
+    };
+
+    using slot_info_vec_t = std::vector<slot_info>;
+
     llama_kv_cache_unified(
             const llama_model &  model,
               layer_filter_cb && filter,
@@ -44,6 +99,7 @@ public:
                     ggml_type    type_v,
                          bool    v_trans,
                          bool    offload,
+                         bool    unified,
                      uint32_t    kv_size,
                      uint32_t    n_seq_max,
                      uint32_t    n_pad,
@@ -56,14 +112,14 @@ public:
     // llama_memory_i
     //
 
-    llama_memory_state_ptr init_batch(
+    llama_memory_context_ptr init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) override;
 
-    llama_memory_state_ptr init_full() override;
+    llama_memory_context_ptr init_full() override;
 
-    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
 
     bool get_can_shift() const override;
 
@@ -87,7 +143,8 @@ public:
     // llama_kv_cache_unified specific API
     //
 
-    uint32_t get_size() const;
+    uint32_t get_size()     const;
+    uint32_t get_n_stream() const;
 
     bool get_has_shift() const;
 
@@ -97,37 +154,48 @@ public:
 
     uint32_t get_n_kv() const;
 
+    // TODO: temporary
+    bool get_supports_set_rows() const;
+
     // get views of the current state of the cache
-    ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
-    ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
+    ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
+    ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
 
     //
     // preparation API
     //
 
-    // find places for the provided ubatches in the cache, returns the head locations
+    // find places for the provided ubatches in the cache, returns the slot infos
     // return empty vector on failure
-    ubatch_heads prepare(const std::vector<llama_ubatch> & ubatches);
+    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
-    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
+    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo, const stream_copy_info & sc_info);
 
-    // return the cell position where we can insert the ubatch
-    // return -1 on failure to find a contiguous slot of kv cells
-    int32_t find_slot(const llama_ubatch & ubatch) const;
+    // find a slot of kv cells that can hold the ubatch
+    // if cont == true, then the slot must be continuous
+    // return empty slot_info on failure
+    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
 
-    // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens)
-    void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch);
+    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
+    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
 
     //
-    // set_input API
+    // input API
     //
 
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+
+    void set_input_k_shift(ggml_tensor * dst) const;
+
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
-    void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
 private:
@@ -141,15 +209,15 @@ private:
 
         ggml_tensor * k;
         ggml_tensor * v;
+
+        std::vector<ggml_tensor *> k_stream;
+        std::vector<ggml_tensor *> v_stream;
     };
 
     bool v_trans = true;  // the value tensor is transposed
 
-    // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
-    // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
-    uint32_t head = 0;
-
     const uint32_t n_seq_max = 1;
+    const uint32_t n_stream  = 1;
 
     // required padding
     const uint32_t n_pad = 1;
@@ -157,14 +225,29 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    bool supports_set_rows = false;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
-    llama_kv_cells_unified cells;
+    // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
+    // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
+    std::vector<uint32_t> v_heads;
+
+    std::vector<llama_kv_cells_unified> v_cells;
+
+    // maps from a sequence id to a stream id
+    std::vector<uint32_t> seq_to_stream;
+
+    // pending stream copies that will be applied during the next update
+    stream_copy_info sc_info;
 
     std::vector<kv_layer> layers;
 
@@ -190,54 +273,60 @@ private:
                           float   freq_base,
                           float   freq_scale) const;
 
-    llm_graph_result_ptr build_graph_shift(
-            const llama_cparams & cparams,
-                   ggml_context * ctx,
-                    ggml_cgraph * gf) const;
+    ggml_cgraph * build_graph_shift(
+               llm_graph_result * res,
+                  llama_context * lctx) const;
 
-    llm_graph_result_ptr build_graph_defrag(
-            const llama_cparams & cparams,
-                   ggml_context * ctx,
-                    ggml_cgraph * gf,
+    ggml_cgraph * build_graph_defrag(
+               llm_graph_result * res,
+                  llama_context * lctx,
               const defrag_info & dinfo) const;
 
-    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
-    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
+    struct cell_ranges_t {
+        uint32_t strm;
 
-    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
-    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+        std::vector<std::pair<uint32_t, uint32_t>> data; // ranges, from inclusive, to exclusive
+    };
+
+    void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
 };
 
-class llama_kv_cache_unified_state : public llama_memory_state_i {
+class llama_kv_cache_unified_context : public llama_memory_context_i {
 public:
     // some shorthands
-    using ubatch_heads = llama_kv_cache_unified::ubatch_heads;
-    using defrag_info  = llama_kv_cache_unified::defrag_info;
+    using slot_info_vec_t  = llama_kv_cache_unified::slot_info_vec_t;
+    using defrag_info      = llama_kv_cache_unified::defrag_info;
+    using stream_copy_info = llama_kv_cache_unified::stream_copy_info;
 
     // used for errors
-    llama_kv_cache_unified_state(llama_memory_status status);
+    llama_kv_cache_unified_context(llama_memory_status status);
 
-    // used to create a full-cache state
-    llama_kv_cache_unified_state(
+    // used to create a full-cache context
+    llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv);
 
-    // used to create an update state
-    llama_kv_cache_unified_state(
+    // used to create an update context
+    llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv,
             llama_context * lctx,
             bool do_shift,
-            defrag_info dinfo);
+            defrag_info dinfo,
+            stream_copy_info sc_info);
 
-    // used to create a decode state from a batch
-    llama_kv_cache_unified_state(
+    // used to create a batch procesing context from a batch
+    llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv,
-            ubatch_heads heads,
+            slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_state();
+    virtual ~llama_kv_cache_unified_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
@@ -247,21 +336,29 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_state specific API
+    // llama_kv_cache_unified_context specific API
     //
 
     uint32_t get_n_kv() const;
 
+    // TODO: temporary
+    bool get_supports_set_rows() const;
+
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
-    void set_input_k_shift(ggml_tensor * dst) const;
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
 
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+    void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
@@ -272,21 +369,23 @@ private:
     llama_context * lctx;
 
     //
-    // update state
+    // update context
     //
 
     bool do_shift = false;
 
     defrag_info dinfo;
 
+    stream_copy_info sc_info;
+
     //
-    // batch processing state
+    // batch processing context
     //
 
-    // the index of the next ubatch to process
-    size_t i_next = 0;
+    // the index of the cur ubatch to process
+    size_t i_cur = 0;
 
-    ubatch_heads heads;
+    slot_info_vec_t sinfos;
 
     std::vector<llama_ubatch> ubatches;
 
@@ -297,7 +396,4 @@ private:
     // a heuristic, to avoid attending the full cache if it is not yet utilized
     // as the cache gets filled, the benefit from this heuristic disappears
     int32_t n_kv;
-
-    // the beginning of the current slot in which the ubatch will be inserted
-    int32_t head;
 };

examples/talk-llama/llama-kv-cells.h

@@ -7,6 +7,7 @@
 #include <cassert>
 #include <vector>
 #include <set>
+#include <map>
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -104,10 +105,30 @@ public:
         res.resize(n);
 
         for (uint32_t j = 0; j < n; ++j) {
-            res.pos[j] = pos[i + j];
-            res.seq[j] = seq[i + j];
+            const auto idx = i + j;
 
-            assert(shift[i + j] == 0);
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
+        }
+
+        return res;
+    }
+
+    // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells_unified res;
+
+        res.resize(idxs.size());
+
+        for (uint32_t j = 0; j < idxs.size(); ++j) {
+            const auto idx = idxs[j];
+
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
         }
 
         return res;
@@ -118,26 +139,58 @@ public:
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
-            if (pos[i + j] == -1 && other.pos[j] != -1) {
+            const auto idx = i + j;
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
                 used.insert(i + j);
             }
 
-            if (pos[i + j] != -1 && other.pos[j] == -1) {
+            if (pos[idx] != -1 && other.pos[j] == -1) {
                 used.erase(i + j);
             }
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_rm(i + j);
             }
 
-            pos[i + j] = other.pos[j];
-            seq[i + j] = other.seq[j];
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_add(i + j);
             }
 
-            assert(shift[i + j] == 0);
+            assert(shift[idx] == 0);
+        }
+    }
+
+    // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+        assert(idxs.size() == other.pos.size());
+
+        for (uint32_t j = 0; j < other.pos.size(); ++j) {
+            const auto idx = idxs[j];
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
+                used.insert(idx);
+            }
+
+            if (pos[idx] != -1 && other.pos[j] == -1) {
+                used.erase(idx);
+            }
+
+            if (pos[idx] != -1) {
+                seq_pos_rm(idx);
+            }
+
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
+
+            if (pos[idx] != -1) {
+                seq_pos_add(idx);
+            }
+
+            assert(shift[idx] == 0);
         }
     }
 
@@ -164,7 +217,7 @@ public:
         assert(seq_id >= 0);
 
         seq[i].reset(seq_id);
-        seq_pos[seq_id].erase(pos[i]);
+        seq_pos_dec(seq_id, pos[i]);
 
         if (seq[i].none()) {
             pos[i] = -1;
@@ -187,7 +240,7 @@ public:
             seq[i].reset();
 
             seq[i].set(seq_id);
-            seq_pos[seq_id].insert(pos[i]);
+            seq_pos_inc(seq_id, pos[i]);
 
             return false;
         }
@@ -232,7 +285,7 @@ public:
         assert(!seq[i].test(seq_id));
 
         seq[i].set(seq_id);
-        seq_pos[seq_id].insert(pos[i]);
+        seq_pos_inc(seq_id, pos[i]);
     }
 
     // return the sequence id of this cell
@@ -259,7 +312,9 @@ public:
             return -1;
         }
 
-        return *seq_pos[seq_id].begin();
+        assert(seq_pos[seq_id].begin()->second > 0);
+
+        return seq_pos[seq_id].begin()->first;
     }
 
     // the maximum position of sequence seq_id currently present in any of the cells
@@ -272,7 +327,9 @@ public:
             return -1;
         }
 
-        return *seq_pos[seq_id].rbegin();
+        assert(seq_pos[seq_id].rbegin()->second > 0);
+
+        return seq_pos[seq_id].rbegin()->first;
     }
 
     // note: call only if the cell is not empty
@@ -389,17 +446,36 @@ private:
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
     std::vector<seq_set_t> seq;
 
-    // the set seq_pos[s] tells us which positions are currently present for sequence s
+    // the set seq_pos[s][p] tells us how many times the position p is currently present for sequence s
+    // if the position p is not present, seq_pos[s][p] is not set
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_SEQ];
+    //
+    // note that we cannot a use an std::set because in some cases a position can occur more than once for the same seq:
+    //  - during performing a cache reuse via (rm + add)
+    //  - some vision models have input embeddings with repeating positions
+    //
+    std::map<llama_pos, int> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
+    void seq_pos_dec(llama_seq_id s, llama_pos p) {
+        auto it = seq_pos[s].find(p);
+        assert(it != seq_pos[s].end());
+
+        if (--it->second == 0) {
+            seq_pos[s].erase(it);
+        }
+    }
+
+    void seq_pos_inc(llama_seq_id s, llama_pos p) {
+        seq_pos[s][p]++;
+    }
+
     // remove cell i
     void seq_pos_rm(uint32_t i) {
         for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
-                seq_pos[s].erase(pos[i]);
+                seq_pos_dec(s, pos[i]);
             }
         }
     }
@@ -408,7 +484,7 @@ private:
     void seq_pos_add(uint32_t i) {
         for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
-                seq_pos[s].insert(pos[i]);
+                seq_pos_inc(s, pos[i]);
             }
         }
     }

examples/talk-llama/llama-memory-hybrid.cpp

@@ -38,6 +38,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
+        1,
         kv_size,
         n_seq_max,
         n_pad,
@@ -56,7 +57,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         n_seq_max
     )) {}
 
-llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
     do {
         balloc.split_reset();
 
@@ -70,7 +71,7 @@ llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ball
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch);
+                ubatch = balloc.split_equal(n_ubatch, false);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -80,33 +81,38 @@ llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ball
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         // prepare the recurrent batches first
         if (!mem_recr->prepare(ubatches)) {
-            // TODO: will the recurrent cache be in an undefined state at this point?
+            // TODO: will the recurrent cache be in an undefined context at this point?
             LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
-            return std::make_unique<llama_memory_hybrid_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+            return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
         }
 
         // prepare the attention cache
         auto heads_attn = mem_attn->prepare(ubatches);
         if (heads_attn.empty()) {
             LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__);
-            return std::make_unique<llama_memory_hybrid_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+            return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
         }
 
-        return std::make_unique<llama_memory_hybrid_state>(
+        return std::make_unique<llama_memory_hybrid_context>(
                 this, std::move(heads_attn), std::move(ubatches));
     } while(false);
 
-    return std::make_unique<llama_memory_hybrid_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_memory_hybrid::init_full() {
-    return std::make_unique<llama_memory_hybrid_state>(this);
+llama_memory_context_ptr llama_memory_hybrid::init_full() {
+    return std::make_unique<llama_memory_hybrid_context>(this);
 }
 
-llama_memory_state_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) {
-    return std::make_unique<llama_memory_hybrid_state>(this, lctx, optimize);
+llama_memory_context_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_memory_hybrid_context>(this, lctx, optimize);
 }
 
 bool llama_memory_hybrid::get_can_shift() const {
@@ -176,39 +182,39 @@ llama_memory_recurrent * llama_memory_hybrid::get_mem_recr() const {
     return mem_recr.get();
 }
 
-llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_status status) : status(status) {}
+llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_status status) : status(status) {}
 
-llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_hybrid * mem) :
-    state_attn(mem->get_mem_attn()->init_full()),
-    state_recr(mem->get_mem_recr()->init_full()),
-    status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) {
+llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_hybrid * mem) :
+    ctx_attn(mem->get_mem_attn()->init_full()),
+    ctx_recr(mem->get_mem_recr()->init_full()),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
 
-llama_memory_hybrid_state::llama_memory_hybrid_state(
+llama_memory_hybrid_context::llama_memory_hybrid_context(
         llama_memory_hybrid * mem,
               llama_context * lctx,
                        bool   optimize) :
-    state_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
-    state_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
-    status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) {
+    ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
+    ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
 
-llama_memory_hybrid_state::llama_memory_hybrid_state(
+llama_memory_hybrid_context::llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)),
-    state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(),                        this->ubatches)),
-    status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) {
+    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
 
-bool llama_memory_hybrid_state::next() {
+bool llama_memory_hybrid_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    state_attn->next();
-    state_recr->next();
+    ctx_attn->next();
+    ctx_recr->next();
 
     if (++i_next >= ubatches.size()) {
         return false;
@@ -217,30 +223,30 @@ bool llama_memory_hybrid_state::next() {
     return true;
 }
 
-bool llama_memory_hybrid_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_memory_hybrid_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
 
     bool res = true;
 
-    res = res & state_attn->apply();
-    res = res & state_recr->apply();
+    res = res & ctx_attn->apply();
+    res = res & ctx_recr->apply();
 
     return res;
 }
 
-llama_memory_status llama_memory_hybrid_state::get_status() const {
+llama_memory_status llama_memory_hybrid_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_memory_hybrid_state::get_ubatch() const {
+const llama_ubatch & llama_memory_hybrid_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_state * llama_memory_hybrid_state::get_state_attn() const {
-    return static_cast<const llama_kv_cache_unified_state *>(state_attn.get());
+const llama_kv_cache_unified_context * llama_memory_hybrid_context::get_attn() const {
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_attn.get());
 }
 
-const llama_memory_recurrent_state * llama_memory_hybrid_state::get_state_recr() const {
-    return static_cast<const llama_memory_recurrent_state *>(state_recr.get());
+const llama_memory_recurrent_context * llama_memory_hybrid_context::get_recr() const {
+    return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
 }

examples/talk-llama/llama-memory-hybrid.h

@@ -49,14 +49,14 @@ public:
     // llama_memory_i
     //
 
-    llama_memory_state_ptr init_batch(
+    llama_memory_context_ptr init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) override;
 
-    llama_memory_state_ptr init_full() override;
+    llama_memory_context_ptr init_full() override;
 
-    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
 
     bool get_can_shift() const override;
 
@@ -90,27 +90,29 @@ private:
     const std::unique_ptr<llama_memory_recurrent> mem_recr;
 };
 
-class llama_memory_hybrid_state : public llama_memory_state_i {
+class llama_memory_hybrid_context : public llama_memory_context_i {
 public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
     // init failure
-    explicit llama_memory_hybrid_state(llama_memory_status status);
+    explicit llama_memory_hybrid_context(llama_memory_status status);
 
     // init full
-    explicit llama_memory_hybrid_state(llama_memory_hybrid * mem);
+    explicit llama_memory_hybrid_context(llama_memory_hybrid * mem);
 
     // init update
-    explicit llama_memory_hybrid_state(
+    explicit llama_memory_hybrid_context(
         llama_memory_hybrid * mem,
               llama_context * lctx,
                        bool   optimize);
 
     // init success
-    llama_memory_hybrid_state(
+    llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches);
 
-    ~llama_memory_hybrid_state() = default;
+    ~llama_memory_hybrid_context() = default;
 
     bool next()  override;
     bool apply() override;
@@ -119,11 +121,11 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_memory_hybrid_state
+    // llama_memory_hybrid_context
     //
 
-    const llama_kv_cache_unified_state * get_state_attn() const;
-    const llama_memory_recurrent_state * get_state_recr() const;
+    const llama_kv_cache_unified_context * get_attn() const;
+    const llama_memory_recurrent_context * get_recr() const;
 
 private:
     // the index of the next ubatch to process
@@ -131,8 +133,8 @@ private:
 
     std::vector<llama_ubatch> ubatches;
 
-    const llama_memory_state_ptr state_attn;
-    const llama_memory_state_ptr state_recr;
+    const llama_memory_context_ptr ctx_attn;
+    const llama_memory_context_ptr ctx_recr;
 
     const llama_memory_status status;
 };

examples/talk-llama/llama-memory-recurrent.cpp

@@ -25,9 +25,6 @@ llama_memory_recurrent::llama_memory_recurrent(
                  uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
-    LLAMA_LOG_INFO("%s: mem_size = %u, n_seq_max = %u, type_r = '%s', type_s = '%s', n_layer = %d\n",
-            __func__, mem_size, n_seq_max, ggml_type_name(type_r), ggml_type_name(type_s), n_layer);
-
     head = 0;
     size = mem_size;
     used = 0;
@@ -84,7 +81,7 @@ llama_memory_recurrent::llama_memory_recurrent(
 
         ggml_context * ctx = ctx_for_buft(buft);
         if (!ctx) {
-            throw std::runtime_error("failed to create ggml context for kv cache");
+            throw std::runtime_error("failed to create ggml context for rs cache");
         }
 
         ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size);
@@ -102,10 +99,10 @@ llama_memory_recurrent::llama_memory_recurrent(
 
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
         if (!buf) {
-            throw std::runtime_error("failed to allocate buffer for kv cache");
+            throw std::runtime_error("failed to allocate buffer for rs cache");
         }
         ggml_backend_buffer_clear(buf, 0);
-        LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
+        LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
         bufs.emplace_back(buf);
     }
 
@@ -113,8 +110,8 @@ llama_memory_recurrent::llama_memory_recurrent(
         const size_t memory_size_r = size_r_bytes();
         const size_t memory_size_s = size_s_bytes();
 
-        LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
-                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f),
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u seqs), R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f), mem_size, n_layer, n_seq_max,
                 ggml_type_name(type_r), (float)memory_size_r / (1024.0f * 1024.0f),
                 ggml_type_name(type_s), (float)memory_size_s / (1024.0f * 1024.0f));
     }
@@ -362,42 +359,52 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_memory_state_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
-    std::vector<llama_ubatch> ubatches;
+llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    do {
+        balloc.split_reset();
 
-    while (true) {
-        llama_ubatch ubatch;
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            llama_ubatch ubatch;
 
-        if (embd_all) {
-            // if all tokens are output, split by sequence
-            ubatch = balloc.split_seq(n_ubatch);
-        } else {
-            ubatch = balloc.split_equal(n_ubatch);
+            if (embd_all) {
+                // if all tokens are output, split by sequence
+                ubatch = balloc.split_seq(n_ubatch);
+            } else {
+                ubatch = balloc.split_equal(n_ubatch, false);
+            }
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        if (ubatch.n_tokens == 0) {
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
             break;
         }
 
-        ubatches.push_back(std::move(ubatch)); // NOLINT
-    }
+        if (!prepare(ubatches)) {
+            break;
+        }
 
-    if (!prepare(ubatches)) {
-        return std::make_unique<llama_memory_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
+        return std::make_unique<llama_memory_recurrent_context>(this, std::move(ubatches));
+    } while (false);
 
-    return std::make_unique<llama_memory_recurrent_state>(this, std::move(ubatches));
+    return std::make_unique<llama_memory_recurrent_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_memory_recurrent::init_full() {
-    return std::make_unique<llama_memory_recurrent_state>(this);
+llama_memory_context_ptr llama_memory_recurrent::init_full() {
+    return std::make_unique<llama_memory_recurrent_context>(this);
 }
 
-llama_memory_state_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) {
+llama_memory_context_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) {
     GGML_UNUSED(lctx);
     GGML_UNUSED(optimize);
 
-    return std::make_unique<llama_memory_recurrent_state>(LLAMA_MEMORY_STATUS_NO_UPDATE);
+    return std::make_unique<llama_memory_recurrent_context>(LLAMA_MEMORY_STATUS_NO_UPDATE);
 }
 
 bool llama_memory_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
@@ -439,7 +446,7 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) {
     // A slot should be always be contiguous.
 
     // can only process batches with an equal number of new tokens in each sequence
-    GGML_ASSERT(ubatch.equal_seqs);
+    GGML_ASSERT(ubatch.equal_seqs());
 
     int32_t min = size - 1;
     int32_t max = 0;
@@ -761,6 +768,8 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
     // Iterate and write all the keys first, each row is a cell
     // Get whole range at a time
     for (uint32_t il = 0; il < n_layer; ++il) {
+        // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+        if (r_l[il] == nullptr) continue;
 
         // Write key type
         const int32_t r_type_i = (int32_t)r_l[il]->type;
@@ -780,6 +789,8 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 
     if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+            if (s_l[il] == nullptr) continue;
 
             // Write value type
             const int32_t s_type_i = (int32_t)s_l[il]->type;
@@ -800,6 +811,9 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
         // When v is transposed, we also need the element size and get the element ranges from each row
         const uint32_t mem_size = size;
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+            if (s_l[il] == nullptr) continue;
+
             const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Write value type
@@ -944,6 +958,8 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
     // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block
     for (uint32_t il = 0; il < n_layer; ++il) {
+        // skip null layers
+        if (r_l[il] == nullptr) continue;
 
         // Read type of key
         int32_t r_type_i_ref;
@@ -971,11 +987,14 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
     if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers
+            if (s_l[il] == nullptr) continue;
 
             // Read type of value
             int32_t s_type_i_ref;
             io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
             const int32_t s_type_i = (int32_t)s_l[il]->type;
+
             if (s_type_i != s_type_i_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_type_i_ref, il);
                 return false;
@@ -998,6 +1017,9 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
     } else {
         // For each layer, read the values for each cell (transposed)
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers
+            if (s_l[il] == nullptr) continue;
+
             const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Read type of value
@@ -1040,22 +1062,22 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 }
 
 //
-// llama_memory_recurrent_state
+// llama_memory_recurrent_context
 //
 
-llama_memory_recurrent_state::llama_memory_recurrent_state(llama_memory_status status) : status(status) {}
+llama_memory_recurrent_context::llama_memory_recurrent_context(llama_memory_status status) : status(status) {}
 
-llama_memory_recurrent_state::llama_memory_recurrent_state(
+llama_memory_recurrent_context::llama_memory_recurrent_context(
         llama_memory_recurrent * mem) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), is_full(true) {
 }
 
-llama_memory_recurrent_state::llama_memory_recurrent_state(
+llama_memory_recurrent_context::llama_memory_recurrent_context(
         llama_memory_recurrent * mem,
         std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), ubatches(std::move(ubatches)) {}
 
-llama_memory_recurrent_state::~llama_memory_recurrent_state() = default;
+llama_memory_recurrent_context::~llama_memory_recurrent_context() = default;
 
-bool llama_memory_recurrent_state::next() {
+bool llama_memory_recurrent_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     if (++i_next >= ubatches.size()) {
@@ -1065,48 +1087,56 @@ bool llama_memory_recurrent_state::next() {
     return true;
 }
 
-bool llama_memory_recurrent_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_memory_recurrent_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    // no ubatches -> this is an update
+    if (ubatches.empty()) {
+        // recurrent cache never performs updates
+        assert(status == LLAMA_MEMORY_STATUS_NO_UPDATE);
+
+        return true;
+    }
 
     mem->find_slot(ubatches[i_next]);
 
     return true;
 }
 
-llama_memory_status llama_memory_recurrent_state::get_status() const {
+llama_memory_status llama_memory_recurrent_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_memory_recurrent_state::get_ubatch() const {
+const llama_ubatch & llama_memory_recurrent_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-uint32_t llama_memory_recurrent_state::get_n_rs() const {
+uint32_t llama_memory_recurrent_context::get_n_rs() const {
     return is_full ? mem->size : mem->n;
 }
 
-uint32_t llama_memory_recurrent_state::get_head() const {
+uint32_t llama_memory_recurrent_context::get_head() const {
     return is_full ? 0 : mem->head;
 }
 
-int32_t llama_memory_recurrent_state::get_rs_z() const {
+int32_t llama_memory_recurrent_context::get_rs_z() const {
     return is_full ? 0 : mem->rs_z;
 }
 
-uint32_t llama_memory_recurrent_state::get_size() const {
+uint32_t llama_memory_recurrent_context::get_size() const {
     return mem->size;
 }
 
-ggml_tensor * llama_memory_recurrent_state::get_r_l(int32_t il) const {
+ggml_tensor * llama_memory_recurrent_context::get_r_l(int32_t il) const {
     return mem->r_l[il];
 }
 
-ggml_tensor * llama_memory_recurrent_state::get_s_l(int32_t il) const {
+ggml_tensor * llama_memory_recurrent_context::get_s_l(int32_t il) const {
     return mem->s_l[il];
 }
 
-int32_t llama_memory_recurrent_state::s_copy(int i) const {
+int32_t llama_memory_recurrent_context::s_copy(int i) const {
     return  mem->cells[i + mem->head].src0;
 }

examples/talk-llama/llama-memory-recurrent.h

@@ -11,8 +11,8 @@
 // llama_memory_recurrent
 //
 
-// TODO: extract the cache state used for graph computation into llama_memory_recurrent_state_i
-//       see the implementation of llama_kv_cache_unified_state_i for an example how to do it
+// TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
+//       see the implementation of llama_kv_cache_unified_context_i for an example how to do it
 class llama_memory_recurrent : public llama_memory_i {
 public:
 
@@ -34,14 +34,14 @@ public:
     // llama_memory_i
     //
 
-    llama_memory_state_ptr init_batch(
+    llama_memory_context_ptr init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) override;
 
-    llama_memory_state_ptr init_full() override;
+    llama_memory_context_ptr init_full() override;
 
-    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
 
     void clear(bool data) override;
 
@@ -125,24 +125,24 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-class llama_memory_recurrent_state : public llama_memory_state_i {
+class llama_memory_recurrent_context : public llama_memory_context_i {
 public:
     // used for errors
-    llama_memory_recurrent_state(llama_memory_status status);
+    llama_memory_recurrent_context(llama_memory_status status);
 
-    // used to create a full-cache state
-    llama_memory_recurrent_state(
+    // used to create a full-cache or update context
+    llama_memory_recurrent_context(
             llama_memory_recurrent * mem);
 
-    // used to create a state from a batch
-    llama_memory_recurrent_state(
+    // used to create a batch processing context from a batch
+    llama_memory_recurrent_context(
             llama_memory_recurrent * mem,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_memory_recurrent_state();
+    virtual ~llama_memory_recurrent_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
@@ -152,7 +152,7 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_memory_recurrent_state specific API
+    // llama_memory_recurrent_context specific API
     //
 
     uint32_t get_n_rs() const;

examples/talk-llama/llama-memory.cpp

@@ -40,3 +40,20 @@ llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_me
     // if either status has an update, then the combined status has an update
     return has_update ? LLAMA_MEMORY_STATUS_SUCCESS : LLAMA_MEMORY_STATUS_NO_UPDATE;
 }
+
+bool llama_memory_status_is_fail(llama_memory_status status) {
+    switch (status) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                return false;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return true;
+            }
+    }
+
+    return false;
+}

examples/talk-llama/llama-memory.h

@@ -3,7 +3,6 @@
 #include "llama.h"
 
 #include <memory>
-#include <vector>
 
 struct llama_ubatch;
 
@@ -28,23 +27,24 @@ enum llama_memory_status {
     LLAMA_MEMORY_STATUS_FAILED_COMPUTE,
 };
 
-// helper function for combining the status of two memory states
+// helper function for combining the status of two memory contexts
 // useful for implementing hybrid memory types (e.g. iSWA)
 llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1);
 
-// the interface for managing the memory state during batch processing
+// helper function for checking if a memory status indicates a failure
+bool llama_memory_status_is_fail(llama_memory_status status);
+
+// the interface for managing the memory context during batch processing
 // this interface is implemented per memory type. see:
-//   - llama_kv_cache_unified_state
-//   - llama_kv_cache_unified_iswa_state
+//   - llama_kv_cache_unified_context
+//   - llama_kv_cache_unified_iswa_context
 //   ...
 //
-// the only method that can mutate the memory and the memory state is llama_memory_i::apply()
-//
-// TODO: rename to llama_memory_context_i ?
-struct llama_memory_state_i {
-    virtual ~llama_memory_state_i() = default;
+// the only method that should mutate the memory and the memory context is llama_memory_i::apply()
+struct llama_memory_context_i {
+    virtual ~llama_memory_context_i() = default;
 
-    // consume the current ubatch from the state and proceed to the next one
+    // consume the current ubatch from the context and proceed to the next one
     // return false if we are done
     virtual bool next() = 0;
 
@@ -55,11 +55,11 @@ struct llama_memory_state_i {
     // get the current ubatch
     virtual const llama_ubatch & get_ubatch() const = 0;
 
-    // get the status of the memory state - used for error handling and checking if any updates would be applied
+    // get the status of the memory context - used for error handling and checking if any updates would be applied
     virtual llama_memory_status get_status() const = 0;
 };
 
-using llama_memory_state_ptr = std::unique_ptr<llama_memory_state_i>;
+using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>;
 
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
@@ -67,19 +67,19 @@ struct llama_memory_i {
     virtual ~llama_memory_i() = default;
 
     // split the input batch into a set of ubatches and verify that they can fit into the cache
-    // return a state object containing the ubatches and KV cache state required to process them
-    // check the llama_memory_state_i::get_status() for the result
-    virtual llama_memory_state_ptr init_batch(
+    // return a context object containing the ubatches and memory state required to process them
+    // check the llama_memory_context_i::get_status() for the result
+    virtual llama_memory_context_ptr init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) = 0;
 
     // simulate full cache, used for allocating worst-case compute buffers
-    virtual llama_memory_state_ptr init_full() = 0;
+    virtual llama_memory_context_ptr init_full() = 0;
 
     // prepare for any pending memory updates, such as shifts, defrags, etc.
     // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
-    virtual llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) = 0;
+    virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0;
 
     // getters
     virtual bool get_can_shift() const = 0;

examples/talk-llama/llama-model.h

@@ -32,16 +32,21 @@ enum llm_type {
     LLM_TYPE_190M,
     LLM_TYPE_220M,
     LLM_TYPE_250M,
+    LLM_TYPE_256M,
     LLM_TYPE_270M,
     LLM_TYPE_335M,
+    LLM_TYPE_350M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
+    LLM_TYPE_0_3B,
     LLM_TYPE_0_5B,
     LLM_TYPE_0_6B,
     LLM_TYPE_1B,
+    LLM_TYPE_1_2B,
     LLM_TYPE_1_3B,
     LLM_TYPE_1_4B,
     LLM_TYPE_1_5B,
@@ -93,8 +98,13 @@ enum llm_type {
     LLM_TYPE_57B_A14B,
     LLM_TYPE_17B_16E, // llama4 Scout
     LLM_TYPE_17B_128E, // llama4 Maverick
+    LLM_TYPE_A13B,
+    LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
     LLM_TYPE_235B_A22B,
+    LLM_TYPE_300B_A47B, // Ernie MoE big
+    LLM_TYPE_E2B,
+    LLM_TYPE_E4B,
 };
 
 std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
@@ -150,6 +160,12 @@ struct llama_layer_convnext {
     struct ggml_tensor * gamma = nullptr;
 };
 
+struct llama_layer_shortconv {
+    struct ggml_tensor * in_proj  = nullptr;
+    struct ggml_tensor * conv     = nullptr;
+    struct ggml_tensor * out_proj = nullptr;
+};
+
 struct llama_layer {
     // normalization
     struct ggml_tensor * attn_norm       = nullptr;
@@ -169,6 +185,10 @@ struct llama_layer {
     struct ggml_tensor * ffn_sub_norm    = nullptr;
     struct ggml_tensor * attn_norm_cross = nullptr;
     struct ggml_tensor * attn_norm_enc   = nullptr;
+    struct ggml_tensor * ssm_norm        = nullptr;
+    struct ggml_tensor * ssm_dt_norm     = nullptr;
+    struct ggml_tensor * ssm_b_norm      = nullptr;
+    struct ggml_tensor * ssm_c_norm      = nullptr;
 
     // attention
     struct ggml_tensor * wq        = nullptr;
@@ -316,9 +336,24 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_scale   = nullptr;
     struct ggml_tensor * ffn_down_scale = nullptr;
 
+    // altup & laurel
+    struct ggml_tensor * per_layer_inp_gate   = nullptr;
+    struct ggml_tensor * per_layer_proj       = nullptr;
+    struct ggml_tensor * per_layer_post_norm  = nullptr;
+    struct ggml_tensor * altup_correct_coef   = nullptr;
+    struct ggml_tensor * altup_correct_scale  = nullptr;
+    struct ggml_tensor * altup_predict_coef   = nullptr;
+    struct ggml_tensor * altup_router         = nullptr;
+    struct ggml_tensor * altup_router_norm    = nullptr;
+    struct ggml_tensor * laurel_l             = nullptr;
+    struct ggml_tensor * laurel_r             = nullptr;
+    struct ggml_tensor * laurel_post_norm     = nullptr;
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
+
+    struct llama_layer_shortconv shortconv;
 };
 
 struct llama_model {
@@ -354,6 +389,13 @@ struct llama_model {
     struct ggml_tensor * conv1d   = nullptr;
     struct ggml_tensor * conv1d_b = nullptr;
 
+    // gemma3n altup
+    struct ggml_tensor * tok_embd_per_layer   = nullptr;
+    struct ggml_tensor * altup_proj           = nullptr;
+    struct ggml_tensor * altup_unembd_proj    = nullptr;
+    struct ggml_tensor * per_layer_model_proj = nullptr;
+    struct ggml_tensor * per_layer_proj_norm  = nullptr;
+
     std::vector<llama_layer> layers;
 
     llama_model_params params;
@@ -412,10 +454,7 @@ struct llama_model {
     llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
-    llm_graph_result_ptr build_graph(
-            const llm_graph_params & params,
-                       ggml_cgraph * gf,
-                    llm_graph_type   type) const;
+    ggml_cgraph * build_graph(const llm_graph_params & params) const;
 
 private:
     struct impl;

examples/talk-llama/llama-quant.cpp

@@ -1,5 +1,4 @@
 #include "llama-quant.h"
-
 #include "llama-impl.h"
 #include "llama-model.h"
 #include "llama-model-loader.h"
@@ -27,6 +26,56 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
+static std::string remap_layer(const std::string & orig_name, const std::vector<int> & prune, std::map<int, std::string> & mapped, int & next_id) {
+    if (prune.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const int blk = std::stoi(match[1]);
+        std::string new_name = orig_name;
+
+        if (mapped.count(blk)) {
+            // Already mapped, do nothing
+        } else if (std::find(prune.begin(), prune.end(), blk) != prune.end()) {
+            mapped[blk] = "";
+        } else if (blk < prune.front()) {
+            mapped[blk] = std::to_string(blk);
+            next_id = blk + 1;
+        } else {
+            mapped[blk] = std::to_string(next_id);
+            ++next_id;
+        }
+
+        return mapped[blk].empty() ? mapped[blk] : new_name.replace(match.position(1), match.length(1), mapped[blk]);
+    }
+
+    return orig_name;
+}
+
+static std::string remap_imatrix (const std::string & orig_name, const std::map<int, std::string> & mapped) {
+    if (mapped.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const std::string blk(match[1]);
+        std::string new_name = orig_name;
+
+        for (const auto & p : mapped) {
+            if (p.second == blk) {
+                LLAMA_LOG_DEBUG("(blk.%d imatrix) ", p.first);
+                return new_name.replace(match.position(1), match.length(1), std::to_string(p.first));
+            }
+        }
+        GGML_ABORT("\n%s: imatrix mapping error for %s\n", __func__, orig_name.c_str());
+    }
+
+    return orig_name;
+}
+
 struct quantize_state_impl {
     const llama_model                 & model;
     const llama_model_quantize_params * params;
@@ -174,7 +223,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                 new_type = GGML_TYPE_Q6_K;
             }
         }
-    } else if (name == "token_embd.weight") {
+    } else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
         if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
             new_type = qs.params->token_embedding_type;
         } else {
@@ -568,6 +617,11 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     gguf_context_ptr ctx_out { gguf_init_empty() };
 
+    std::vector<int> prune_list = {};
+    if (params->prune_layers) {
+        prune_list = *static_cast<const std::vector<int> *>(params->prune_layers);
+    }
+
     // copy the KV pairs from the input file
     gguf_set_kv     (ctx_out.get(), ml.meta.get());
     gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
@@ -597,12 +651,32 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         }
     }
 
+    std::map<int, std::string> mapped;
+    int blk_id = 0;
+    int pruned_attention_w = 0;
+
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
     tensors.reserve(ml.weights_map.size());
     for (const auto & it : ml.weights_map) {
+        const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
+        if (remapped_name.empty()) {
+            if (it.first.find("attn_v.weight") != std::string::npos ||
+                it.first.find("attn_qkv.weight") != std::string::npos ||
+                it.first.find("attn_kv_b.weight") != std::string::npos) {
+                    pruned_attention_w++;
+            }
+            LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
+            continue;
+        } else if (remapped_name != it.first) {
+            ggml_set_name(it.second.tensor, remapped_name.c_str());
+            LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
+        }
         tensors.push_back(&it.second);
     }
+    if (!prune_list.empty()) {
+        gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_BLOCK_COUNT).c_str(), blk_id);
+    }
 
     // keep_split requires that the weights are sorted by split index
     if (params->keep_split) {
@@ -640,7 +714,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;
@@ -681,7 +755,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         for (size_t i = 0; i < ctx_outs.size(); ++i) {
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
-            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors);
+            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), (int32_t)tensors.size());
         }
     }
 
@@ -756,6 +830,13 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // NOTE: can't use LLM_TN here because the layer number is not known
         quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
 
+        // these are very small (e.g. 4x4)
+        quantize &= name.find("altup")  == std::string::npos;
+        quantize &= name.find("laurel") == std::string::npos;
+
+        // these are not too big so keep them as it is
+        quantize &= name.find("per_layer_model_proj") == std::string::npos;
+
         // do not quantize positional embeddings and token types (BERT)
         quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD,    "weight");
         quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");
@@ -763,6 +844,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize Mamba's small yet 2D weights
         // NOTE: can't use LLM_TN here because the layer number is not known
         quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
+        quantize &= name.find("shortconv.conv.weight") == std::string::npos;
 
         // do not quantize RWKV's small yet 2D weights
         quantize &= name.find("time_mix_first.weight") == std::string::npos;
@@ -802,8 +884,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                         if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
                             if  (qtype != new_type) {
                                 LLAMA_LOG_DEBUG("(overriding %s) ", ggml_type_name(new_type));
-                                new_type = qtype;
-                                break; // if two or more types are specified for the tensor, first match wins
+                                new_type = qtype; // if two or more types are specified for the same tensor, the last match wins
                             }
                         }
                     }
@@ -832,7 +913,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             const float * imatrix = nullptr;
             if (imatrix_data) {
-                auto it = imatrix_data->find(tensor->name);
+                auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
                 if (it == imatrix_data->end()) {
                     LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
                 } else {
@@ -947,6 +1028,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
         /*.tensor_type                 =*/ nullptr,
+        /*.prune_layers                =*/ nullptr
     };
 
     return result;

examples/talk-llama/llama-vocab.cpp

@@ -11,6 +11,7 @@
 #include <cassert>
 #include <cctype>
 #include <cfloat>
+#include <cmath>
 #include <cstdarg>
 #include <cstring>
 #include <forward_list>
@@ -351,6 +352,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 break;
             case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
             case LLAMA_VOCAB_PRE_TYPE_QWEN2:
+            case LLAMA_VOCAB_PRE_TYPE_HUNYUAN:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
@@ -403,6 +405,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_KIMI_K2:
+                regex_exprs = {
+                    // K2 trigger pattern - this will activate the custom K2 handler in unicode.cpp
+                    // The custom handler implements all K2 patterns with proper Han character exclusion
+                    "\\p{Han}+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_SUPERBPE:
                 regex_exprs = {
                     "\\p{N}+",
@@ -1195,6 +1204,284 @@ private:
     const llm_tokenizer_rwkv & tokenizer;
 };
 
+struct llm_tokenizer_plamo2 : llm_tokenizer {
+    llm_tokenizer_plamo2(const llama_vocab & vocab) {
+        build(vocab);
+    }
+
+    void build(const llama_vocab & vocab) {
+        // Reset internal structures
+        tokens_.clear();
+        bytes_.assign(256, 0);
+        to_suffix_id_.clear();
+        table_.clear();
+
+        // Build token list and byte mapping
+        std::unordered_map<std::string, float> suffix_to_score;
+        std::unordered_map<std::string, llama_token> token_to_id;
+
+        for (size_t token_id = 0; token_id < vocab.n_tokens(); ++token_id) {
+            const auto & entry = vocab.get_token_data(token_id);
+            tokens_.push_back(entry.text);
+            token_to_id[entry.text] = static_cast<llama_token>(token_id);
+
+            // Handle byte tokens
+            if (vocab.is_byte(token_id)) {
+                if (entry.text.length() == 6 && entry.text.substr(0, 3) == "<0x" && entry.text.back() == '>') {
+                    std::string hex_str = entry.text.substr(3, 2);
+                    int byte_val = std::stoi(hex_str, nullptr, 16);
+                    bytes_[byte_val] = static_cast<llama_token>(token_id);
+                }
+                continue;
+            }
+
+            // Add token and all its suffixes to suffix_to_score
+            suffix_to_score[entry.text] = entry.score;
+
+            // Extract suffixes character by character (UTF-8 aware)
+            std::vector<uint32_t> cpts = unicode_cpts_from_utf8(entry.text);
+            for (size_t i = 1; i < cpts.size(); ++i) {
+                std::string suffix;
+                for (size_t j = i; j < cpts.size(); ++j) {
+                    suffix += unicode_cpt_to_utf8(cpts[j]);
+                }
+                if (suffix_to_score.find(suffix) == suffix_to_score.end()) {
+                    suffix_to_score[suffix] = std::numeric_limits<float>::quiet_NaN();
+                }
+            }
+        }
+
+        // Check that all byte tokens are set
+        for (int i = 0; i < 256; ++i) {
+            if (bytes_[i] == 0) {
+                throw std::runtime_error("Byte token for <0x" + std::to_string(i) + "> is not set");
+            }
+        }
+
+        // Build suffix list in lexicographical order of reversed strings
+        std::vector<std::string> suffixes;
+        for (const auto & pair : suffix_to_score) {
+            suffixes.push_back(pair.first);
+        }
+        suffixes.push_back("");  // Empty suffix
+
+        std::sort(suffixes.begin(), suffixes.end(), [](const std::string & a, const std::string & b) {
+            std::string rev_a(a.rbegin(), a.rend());
+            std::string rev_b(b.rbegin(), b.rend());
+            return rev_a < rev_b;
+        });
+
+        // Build suffix_to_id and to_suffix_id_
+        std::unordered_map<std::string, int32_t> suffix_to_id;
+        int32_t num_pieces = 0;
+
+        for (const auto & suffix : suffixes) {
+            suffix_to_id[suffix] = num_pieces;
+            if (!suffix.empty()) {
+                std::vector<uint32_t> cpts = unicode_cpts_from_utf8(suffix);
+
+                std::string remaining;
+                for (size_t i = 1; i < cpts.size(); ++i) {
+                    remaining += unicode_cpt_to_utf8(cpts[i]);
+                }
+
+                int64_t piece_code = (static_cast<int64_t>(cpts[0]) << 32) | suffix_to_id[remaining];
+                to_suffix_id_[piece_code] = num_pieces;
+
+                // Count number of pieces for this suffix
+                int32_t pieces_for_suffix = 1; // sentinel row
+                for (int32_t piece_length = static_cast<int32_t>(cpts.size()); piece_length > 0; --piece_length) {
+                    std::string piece;
+                    for (int32_t i = 0; i < piece_length; ++i) {
+                        piece += unicode_cpt_to_utf8(cpts[i]);
+                    }
+                    if (suffix_to_score.find(piece) != suffix_to_score.end()) {
+                        pieces_for_suffix++;
+                    }
+                }
+                num_pieces += pieces_for_suffix;
+            } else {
+                num_pieces++;  // Empty suffix contributes one piece (sentinel row)
+            }
+        }
+
+        // Build flattened table
+        table_.resize(num_pieces, std::vector<int32_t>(4, 0));
+        int32_t table_idx = 0;
+
+        for (const auto & suffix : suffixes) {
+            // Add all prefixes of the suffix to the table (in decreasing order of length)
+            std::vector<uint32_t> cpts = unicode_cpts_from_utf8(suffix);
+            for (int32_t piece_length = static_cast<int32_t>(cpts.size()); piece_length > 0; --piece_length) {
+                std::string piece;
+                for (int32_t i = 0; i < piece_length; ++i) {
+                    piece += unicode_cpt_to_utf8(cpts[i]);
+                }
+
+                auto score_it = suffix_to_score.find(piece);
+                if (score_it == suffix_to_score.end()) {
+                    continue;
+                }
+
+                table_[table_idx][TABLE_PIECE_LENGTH] = piece_length;
+                auto token_it = token_to_id.find(piece);
+                table_[table_idx][TABLE_TOKEN_ID] = (token_it != token_to_id.end()) ? token_it->second : -1;
+
+                float score = score_it->second;
+                table_[table_idx][TABLE_SCORE] = std::isfinite(score) ?
+                    static_cast<int32_t>(std::round(score * 1e4)) : INVALID_SCORE;
+                table_[table_idx][TABLE_PIECE_ID] = suffix_to_id[piece];
+
+                table_idx++;
+            }
+
+            // Add sentinel row
+            table_[table_idx][TABLE_PIECE_LENGTH] = 1;
+            table_[table_idx][TABLE_TOKEN_ID] = -1;
+            table_[table_idx][TABLE_SCORE] = UNKNOWN_SCORE;
+            table_idx++;
+        }
+    }
+
+    std::vector<llama_token> encode(const std::string & text) const {
+        std::vector<uint32_t> unicode_data = unicode_cpts_from_utf8(text);
+        // Skip the first code point if it is a BOM (Byte Order Mark)
+        if (!unicode_data.empty() && unicode_data[0] == 0xFEFF) {
+            unicode_data.erase(unicode_data.begin());
+        }
+
+        if (unicode_data.empty()) {
+            return {};
+        }
+
+        const size_t data_len = unicode_data.size();
+
+        // Initialize scores array (dynamic programming)
+        std::vector<int64_t> scores(data_len + 1, static_cast<int64_t>(1) << 60);
+        scores[data_len] = 0;
+
+        // Path array to track best tokenization
+        std::vector<std::vector<int32_t>> path(data_len + 1, std::vector<int32_t>(3, 0));
+
+        int32_t suffix_id = 0;
+
+        // Process from end to beginning
+        for (int i = static_cast<int>(data_len) - 1; i >= 0; --i) {
+            uint32_t c = unicode_data[i];
+
+            // Find next suffix ID
+            for (size_t p = suffix_id; p < table_.size(); ++p) {
+                int64_t piece_code = (static_cast<int64_t>(c) << 32) | table_[p][TABLE_PIECE_ID];
+                auto it = to_suffix_id_.find(piece_code);
+                suffix_id = (it != to_suffix_id_.end()) ? it->second : 0;
+
+                if (suffix_id > 0 || table_[p][TABLE_SCORE] == UNKNOWN_SCORE) {
+                    break;
+                }
+            }
+
+            // Update best path
+            for (size_t p = suffix_id; p < table_.size(); ++p) {
+                int32_t score = table_[p][TABLE_SCORE];
+                if (score > INVALID_SCORE) {
+                    int32_t piece_length = table_[p][TABLE_PIECE_LENGTH];
+                    int64_t s = scores[i + piece_length] - score;
+
+                    if (s < scores[i]) {
+                        scores[i] = s;
+                        path[i][PATH_TOKEN_LENGTH] = piece_length;
+                        path[i][PATH_TOKEN_ID] = table_[p][TABLE_TOKEN_ID];
+                        path[i][PATH_NUM_TOKENS] = path[i + piece_length][PATH_NUM_TOKENS] + 1;
+
+                        if (score == UNKNOWN_SCORE) {
+                            // Add UTF-8 byte count
+                            path[i][PATH_NUM_TOKENS] += (c >= 0x80) + (c >= 0x800) + (c >= 0x10000);
+                        }
+                    }
+                }
+
+                if (score == UNKNOWN_SCORE) {
+                    break;
+                }
+            }
+        }
+
+        // Decode the best path
+        std::vector<llama_token> token_ids;
+        token_ids.reserve(path[0][PATH_NUM_TOKENS]);
+
+        int pos = 0;
+        while (pos < static_cast<int>(data_len)) {
+            if (path[pos][PATH_TOKEN_ID] >= 0) {
+                token_ids.push_back(path[pos][PATH_TOKEN_ID]);
+            } else {
+                // Fall back to byte tokens
+                uint32_t c = unicode_data[pos];
+                int s = 1 + (c >= 0x80) + (c >= 0x800) + (c >= 0x10000);
+
+                for (int i = 0; i < s; ++i) {
+                    uint8_t b;
+                    if (s == 1) {
+                        b = c;
+                    } else {
+                        if (i == 0) {
+                            b = (0xF00 >> s) & 0xFF;
+                        } else {
+                            b = 0x80;
+                        }
+                    }
+                    token_ids.push_back(bytes_[b | ((c >> ((s - i - 1) * 6)) & 0x3F)]);
+                }
+            }
+
+            assert(path[pos][PATH_TOKEN_LENGTH] > 0);
+            pos += path[pos][PATH_TOKEN_LENGTH];
+        }
+
+        return token_ids;
+    }
+private:
+    // Constants for table structure
+    static constexpr int32_t TABLE_PIECE_LENGTH = 0;
+    static constexpr int32_t TABLE_TOKEN_ID     = 1;
+    static constexpr int32_t TABLE_SCORE        = 2;
+    static constexpr int32_t TABLE_PIECE_ID     = 3;
+
+    // Constants for path array
+    static constexpr int32_t PATH_TOKEN_LENGTH  = 0;
+    static constexpr int32_t PATH_TOKEN_ID      = 1;
+    static constexpr int32_t PATH_NUM_TOKENS    = 2;
+
+    // Score constants
+    static constexpr int32_t INVALID_SCORE = -20000000;
+    static constexpr int32_t UNKNOWN_SCORE = -10000000;
+
+    // List of tokens in the vocabulary
+    std::vector<std::string> tokens_;
+
+    // Mapping from byte code point to token ID (for byte fallback)
+    std::vector<llama_token> bytes_;
+
+    // Mapping from piece code to suffix ID
+    std::unordered_map<int64_t, int32_t> to_suffix_id_;
+
+    // Flattened table representing the Trie structure
+    // Each row contains: [piece_length, token_id, score, piece_id]
+    std::vector<std::vector<int32_t>> table_;
+};
+
+struct llm_tokenizer_plamo2_session {
+    llm_tokenizer_plamo2_session(const llm_tokenizer_plamo2 & tokenizer) : tokenizer(tokenizer) {}
+
+    void tokenize(const std::string & text, std::vector<llama_token> & output) {
+        std::vector<llama_token> tokens = tokenizer.encode(text);
+        output.insert(output.end(), tokens.begin(), tokens.end());
+    }
+
+private:
+    const llm_tokenizer_plamo2 & tokenizer;
+};
+
 //
 // impl
 //
@@ -1498,6 +1785,16 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             special_unk_id = LLAMA_TOKEN_NULL;
             special_sep_id = LLAMA_TOKEN_NULL;
             special_pad_id = LLAMA_TOKEN_NULL;
+        } else if (tokenizer_model == "plamo2") {
+            type = LLAMA_VOCAB_TYPE_PLAMO2;
+
+            // PLaMo-2 default special tokens (these will be overridden by model config)
+            special_bos_id = 1;  // <|plamo:bos|>
+            special_eos_id = 2;  // <|plamo:eos|>
+            special_unk_id = 0;  // <|plamo:unk|>
+            special_sep_id = LLAMA_TOKEN_NULL;
+            special_pad_id = 3;  // <|plamo:pad|>
+            special_mask_id = LLAMA_TOKEN_NULL;
         } else {
             throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str()));
         }
@@ -1522,7 +1819,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe"||
                     tokenizer_pre == "falcon3"  ||
-                    tokenizer_pre == "pixtral") {
+                    tokenizer_pre == "falcon-h1" ||
+                    tokenizer_pre == "pixtral"  ||
+                    tokenizer_pre == "midm-2.0" ||
+                    tokenizer_pre == "lfm2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;
@@ -1554,7 +1854,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "jina-de" ||
                     tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v2-es" ||
-                    tokenizer_pre == "jina-v2-de") {
+                    tokenizer_pre == "jina-v2-de" ||
+                    tokenizer_pre == "a.x-4.0") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
@@ -1624,6 +1925,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                 tokenizer_pre == "exaone") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_EXAONE;
+            } else if (
+                tokenizer_pre == "exaone4") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                 tokenizer_pre == "chameleon") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
@@ -1656,6 +1960,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "seed-coder") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "hunyuan") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "kimi-k2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -1839,6 +2151,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<EOT>"
                         || t.first == "_<EOT>"
                         || t.first == "<｜end▁of▁sentence｜>" // DeepSeek
+                        || t.first == "<end_of_utterance>" // smoldocling
                    ) {
                     special_eot_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1998,6 +2311,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"
                     || t.first == "<|end_of_text|>"
+                    || t.first == "<end_of_utterance>" // smoldocling
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2134,13 +2448,14 @@ enum llama_vocab_type llama_vocab::impl::get_type() const {
 
 std::string llama_vocab::impl::type_name() const{
     switch (type) {
-        case LLAMA_VOCAB_TYPE_NONE: return "no vocab";
-        case LLAMA_VOCAB_TYPE_SPM:  return "SPM";
-        case LLAMA_VOCAB_TYPE_BPE:  return "BPE";
-        case LLAMA_VOCAB_TYPE_WPM:  return "WPM";
-        case LLAMA_VOCAB_TYPE_UGM:  return "UGM";
-        case LLAMA_VOCAB_TYPE_RWKV: return "RWKV";
-        default:                    return "unknown";
+        case LLAMA_VOCAB_TYPE_NONE:   return "no vocab";
+        case LLAMA_VOCAB_TYPE_SPM:    return "SPM";
+        case LLAMA_VOCAB_TYPE_BPE:    return "BPE";
+        case LLAMA_VOCAB_TYPE_WPM:    return "WPM";
+        case LLAMA_VOCAB_TYPE_UGM:    return "UGM";
+        case LLAMA_VOCAB_TYPE_RWKV:   return "RWKV";
+        case LLAMA_VOCAB_TYPE_PLAMO2: return "PLaMo2";
+        default:                      return "unknown";
     }
 }
 
@@ -2223,6 +2538,9 @@ void llama_vocab::impl::init_tokenizer(enum llama_vocab_type type) {
         case LLAMA_VOCAB_TYPE_RWKV:
             tokenizer = std::make_unique<llm_tokenizer_rwkv>(vocab);
             break;
+        case LLAMA_VOCAB_TYPE_PLAMO2:
+            tokenizer = std::make_unique<llm_tokenizer_plamo2>(vocab);
+            break;
         default:
             GGML_ABORT("unsupported vocab type");
     }
@@ -2555,6 +2873,23 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
                         std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
 
+#ifdef PRETOKENIZERDEBUG
+                        LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", text.length(), fragment.offset, fragment.length, text.c_str());
+#endif
+
+                        session.tokenize(text, output);
+                    } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
+                        output.push_back(fragment.token);
+                    }
+                }
+            } break;
+        case LLAMA_VOCAB_TYPE_PLAMO2:
+            {
+                llm_tokenizer_plamo2_session session(*static_cast<const llm_tokenizer_plamo2 *>(tokenizer.get()));
+                for (const auto & fragment : fragment_buffer) {
+                    if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
+                        std::string text = fragment.raw_text.substr(fragment.offset, fragment.length);
+
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", text.length(), fragment.offset, fragment.length, text.c_str());
 #endif
@@ -2653,6 +2988,24 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
                 memcpy(buf, result.data(), result.size());
                 return (int)result.size();
             }
+            case LLAMA_VOCAB_TYPE_PLAMO2: {
+                // PLaMo-2 uses similar token handling as BPE/SPM
+                if (vocab.is_byte(token)) {
+                    // Handle byte tokens like <0xXX>
+                    if (token_text.length() == 6 && token_text.substr(0, 3) == "<0x" && token_text.back() == '>') {
+                        int hex_val = std::stoi(token_text.substr(3, 2), nullptr, 16);
+                        if (length < 1) {
+                            return -1;
+                        }
+                        buf[0] = static_cast<char>(hex_val);
+                        return 1;
+                    }
+                }
+
+                // Normal token - just copy the text
+                std::string result = token_text;
+                return _try_copy(result.data(), result.size());
+            }
             default:
                 GGML_ABORT("fatal error");
         }
@@ -2897,6 +3250,12 @@ llama_token llama_vocab::byte_to_token(uint8_t ch) const {
         case LLAMA_VOCAB_TYPE_BPE: {
             return pimpl->token_to_id.at(unicode_byte_to_utf8(ch));
         }
+        case LLAMA_VOCAB_TYPE_PLAMO2: {
+            // PLaMo-2 uses byte tokens in format <0xXX>
+            char hex_str[8];
+            snprintf(hex_str, sizeof(hex_str), "<0x%02X>", ch);
+            return pimpl->token_to_id.at(hex_str);
+        }
         default:
             GGML_ABORT("fatal error");
     }
@@ -2998,6 +3357,10 @@ llama_token llama_vocab::token_fim_sep() const {
     return pimpl->special_fim_sep_id;
 }
 
+llama_token llama_vocab::token_mask() const {
+    return pimpl->special_mask_id;
+}
+
 bool llama_vocab::get_add_space_prefix() const {
     return pimpl->add_space_prefix;
 }
@@ -3238,6 +3601,10 @@ llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab) {
     return vocab->token_fim_sep();
 }
 
+llama_token llama_vocab_mask(const struct llama_vocab* vocab) {
+    return vocab->token_mask();
+}
+
 // deprecated
 const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token) {
     return llama_vocab_get_text(vocab, token);
@@ -3374,4 +3741,3 @@ int32_t llama_detokenize(
                         bool   unparse_special) {
     return vocab->detokenize(tokens, n_tokens, text, text_len_max, remove_special, unparse_special);
 }
-

examples/talk-llama/llama-vocab.h

@@ -6,6 +6,48 @@
 #include <vector>
 #include <memory>
 
+// pre-tokenization types
+enum llama_vocab_pre_type {
+    LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
+    LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
+    LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
+    LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
+    LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
+    LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
+    LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
+    LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
+    LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
+    LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
+    LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
+    LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
+    LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM3       = 16,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
+    LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
+    LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
+    LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
+    LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
+    LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+    LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
+    LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
+    LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+    LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
+    LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
+    LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
+    LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
+    LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
+    LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+    LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+    LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
+    LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
+    LLAMA_VOCAB_PRE_TYPE_KIMI_K2        = 37,
+};
+
 struct LLM_KV;
 struct llama_model_loader;
 
@@ -59,6 +101,7 @@ struct llama_vocab {
     llama_token token_sep() const;
     llama_token token_nl () const;
     llama_token token_pad() const;
+    llama_token token_mask() const;
 
     llama_token token_prefix() const;
     llama_token token_middle() const;

examples/talk-llama/llama.h

@@ -71,52 +71,13 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-        LLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on Unigram
-        LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization
-    };
-
-    // pre-tokenization types
-    enum llama_vocab_pre_type {
-        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
-        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
-        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
-        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
-        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
-        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
-        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
-        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
-        LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
-        LLAMA_VOCAB_PRE_TYPE_CHATGLM3       = 16,
-        LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
-        LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
-        LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
-        LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
-        LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
-        LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
-        LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
-        LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
-        LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
-        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
-        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
-        LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
-        LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
-        LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
-        LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
-        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
-        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
+        LLAMA_VOCAB_TYPE_NONE   = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM    = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
+        LLAMA_VOCAB_TYPE_BPE    = 2, // GPT-2 tokenizer based on byte-level BPE
+        LLAMA_VOCAB_TYPE_WPM    = 3, // BERT tokenizer based on WordPiece
+        LLAMA_VOCAB_TYPE_UGM    = 4, // T5 tokenizer based on Unigram
+        LLAMA_VOCAB_TYPE_RWKV   = 5, // RWKV tokenizer based on greedy tokenization
+        LLAMA_VOCAB_TYPE_PLAMO2 = 6, // PLaMo-2 tokenizer based on Aho-Corasick with dynamic programming
     };
 
     enum llama_rope_type {
@@ -374,6 +335,9 @@ extern "C" {
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
                           // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
                           //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
+        bool kv_unified;  // use a unified buffer across the input sequences when computing the attention
+                          // try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
+                          // ref: https://github.com/ggml-org/llama.cpp/pull/14363
     };
 
     // model quantization parameters
@@ -390,6 +354,7 @@ extern "C" {
         void * imatrix;                       // pointer to importance matrix data
         void * kv_overrides;                  // pointer to vector containing overrides
         void * tensor_types;                  // pointer to vector containing tensor types
+        void * prune_layers;                  // pointer to vector containing layer indices to prune
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -763,7 +728,7 @@ extern "C" {
     //   - lazily on next llama_decode()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    DEPRECATED(void llama_kv_self_seq_div(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -943,12 +908,14 @@ extern "C" {
     // Requires the context to have a memory.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the memory state is restored to the state before this call
+    // Upon fatal-error or abort, the ubatches that managed to be been processed will remain in the memory state of the context
+    //   To handle this correctly, query the memory state using llama_memory_seq_pos_min() and llama_memory_seq_pos_max()
+    // Upon other return values, the memory state is restored to the state before this call
     //    0 - success
     //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    //    2 - aborted
+    //    2 - aborted     (processed ubatches will remain in the context's memory)
     //   -1 - invalid input batch
-    // < -1 - error
+    // < -1 - fatal error (processed ubatches will remain in the context's memory)
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -989,6 +956,7 @@ extern "C" {
     // in the order they have appeared in the batch.
     // Rows: number of tokens for which llama_batch.logits[i] != 0
     // Cols: n_vocab
+    // TODO: deprecate in favor of llama_get_logits_ith() (ref: https://github.com/ggml-org/llama.cpp/pull/14853#issuecomment-3113143522)
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
     // Logits for the ith token. For positive indices, Equivalent to:
@@ -1003,6 +971,7 @@ extern "C" {
     // in the order they have appeared in the batch.
     // shape: [n_outputs*n_embd]
     // Otherwise, returns NULL.
+    // TODO: deprecate in favor of llama_get_embeddings_ith() (ref: https://github.com/ggml-org/llama.cpp/pull/14853#issuecomment-3113143522)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
     // Get the embeddings for the ith token. For positive indices, Equivalent to:
@@ -1041,6 +1010,7 @@ extern "C" {
     LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator
     LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line
     LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding
+    LLAMA_API llama_token llama_vocab_mask(const struct llama_vocab * vocab); // mask
 
     LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
     LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
@@ -1426,6 +1396,7 @@ extern "C" {
 
         int32_t n_p_eval;
         int32_t n_eval;
+        int32_t n_reused; // number of times a ggml compute graph had been reused
     };
 
     struct llama_perf_sampler_data {

examples/talk-llama/unicode.cpp

@@ -557,6 +557,178 @@ static std::vector<size_t> unicode_regex_split_stl(const std::string & text, con
     return bpe_offsets;
 }
 
+// K2 system regex patterns (from tokenization_kimi.py):
+// [\p{Han}]+|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+
+static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets;
+    bpe_offsets.reserve(offsets.size());
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
+        };
+
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
+
+            // Pattern 1: [\p{Han}]+ (Chinese characters)
+            if (unicode_cpt_is_han(cpt)) {
+                while (unicode_cpt_is_han(_get_cpt(pos))) {
+                    pos++;
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 2 & 3: Letter words excluding Han characters with optional contractions
+            // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?:'s|'t|'re|'ve|'m|'ll|'d)?
+            // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?:'s|'t|'re|'ve|'m|'ll|'d)?
+            // Check if current char is a letter OR if current char could be a leading char and next char is a letter
+            bool is_letter_pattern = (flags.is_letter && !unicode_cpt_is_han(cpt)) ||
+                                     (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number) &&
+                                      _get_flags(pos + 1).is_letter && !unicode_cpt_is_han(_get_cpt(pos + 1)));
+
+            if (is_letter_pattern) {
+                // Handle optional leading non-letter/non-number character
+                bool has_leading_char = false;
+                if (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number)) {
+                    has_leading_char = true;
+                    pos++;
+                }
+
+                // Match letter sequence (excluding Han characters)
+                bool has_letters = false;
+                while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
+                    has_letters = true;
+                    pos++;
+                }
+
+                // Only proceed if we found letters (after potentially skipping leading char)
+                if (has_letters || (!has_leading_char && _get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos)))) {
+                    if (!has_letters) pos++; // consume the first letter if we didn't already
+
+                    // Continue consuming letters
+                    while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
+                        pos++;
+                    }
+
+                    // Check for optional contractions (?:'s|'t|'re|'ve|'m|'ll|'d)
+                    if (_get_cpt(pos) == '\'' && pos + 1 < offset_end) {
+                        uint32_t cpt_next = unicode_tolower(_get_cpt(pos + 1));
+                        if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                            pos += 2;
+                        } else if (pos + 2 < offset_end) {
+                            uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos + 2));
+                            if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                                (cpt_next == 'v' && cpt_next_next == 'e') ||
+                                (cpt_next == 'l' && cpt_next_next == 'l')) {
+                                pos += 3;
+                            }
+                        }
+                    }
+
+                    _add_token(pos);
+                    continue;
+                } else if (has_leading_char) {
+                    // We consumed a leading char but found no letters, backtrack
+                    pos--;
+                }
+            }
+
+            // Pattern 4: \p{N}{1,3} (numbers 1-3 digits)
+            if (flags.is_number) {
+                size_t ini = pos;
+                while (_get_flags(pos).is_number) {
+                    if (++pos - ini >= 3) {
+                        _add_token(pos);
+                        ini = pos;
+                    }
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 5:  ?[^\s\p{L}\p{N}]+[\r\n]* (optional space + non-word chars + optional newlines)
+            auto flags2 = (cpt == ' ' ? _get_flags(pos + 1) : flags);
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
+                pos += (cpt == ' ');
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
+                    flags2 = _get_flags(++pos);
+                }
+                // Match optional [\r\n]*
+                uint32_t cpt2 = _get_cpt(pos);
+                while (cpt2 == '\r' || cpt2 == '\n') {
+                    cpt2 = _get_cpt(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Count whitespace characters
+            size_t num_whitespaces = 0;
+            size_t last_end_r_or_n = 0;
+            while (_get_flags(pos + num_whitespaces).is_whitespace) {
+                uint32_t cpt2 = _get_cpt(pos + num_whitespaces);
+                if (cpt2 == '\r' || cpt2 == '\n') {
+                    last_end_r_or_n = pos + num_whitespaces + 1;
+                }
+                num_whitespaces++;
+            }
+
+            // Pattern 6: \s*[\r\n]+ (whitespace with newlines)
+            if (last_end_r_or_n > 0) {
+                pos = last_end_r_or_n;
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 7: \s+(?!\S) (trailing whitespace)
+            if (num_whitespaces > 1 && _get_cpt(pos + num_whitespaces) != OUT_OF_RANGE) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 8: \s+ (general whitespace)
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // No matches - consume single character
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
 static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
     std::vector<size_t> bpe_offsets;
 
@@ -567,6 +739,9 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
             regex_expr == "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+") {
 
         bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
+    } else if (regex_expr == "\\p{Han}+") {
+        // K2's first pattern - handle all K2 patterns together
+        bpe_offsets = unicode_regex_split_custom_kimi_k2(text, offsets);
     }
 
     return bpe_offsets;
@@ -672,6 +847,38 @@ uint32_t unicode_tolower(uint32_t cpt) {
     return cpt;  // Return the original code point if no lowercase mapping is found
 }
 
+bool unicode_cpt_is_han(uint32_t cpt) {
+    // Han character ranges (Chinese/CJK characters)
+    // CJK Unified Ideographs (most common)
+    if (cpt >= 0x4E00 && cpt <= 0x9FFF) return true;
+
+    // CJK Extension A
+    if (cpt >= 0x3400 && cpt <= 0x4DBF) return true;
+
+    // CJK Extension B
+    if (cpt >= 0x20000 && cpt <= 0x2A6DF) return true;
+
+    // CJK Extension C
+    if (cpt >= 0x2A700 && cpt <= 0x2B73F) return true;
+
+    // CJK Extension D
+    if (cpt >= 0x2B740 && cpt <= 0x2B81F) return true;
+
+    // CJK Extension E
+    if (cpt >= 0x2B820 && cpt <= 0x2CEAF) return true;
+
+    // CJK Extension F
+    if (cpt >= 0x2CEB0 && cpt <= 0x2EBEF) return true;
+
+    // CJK Compatibility Ideographs
+    if (cpt >= 0xF900 && cpt <= 0xFAFF) return true;
+
+    // CJK Compatibility Ideographs Supplement
+    if (cpt >= 0x2F800 && cpt <= 0x2FA1F) return true;
+
+    return false;
+}
+
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {

examples/talk-llama/unicode.h

@@ -63,4 +63,6 @@ uint8_t     unicode_utf8_to_byte(const std::string & utf8);
 
 uint32_t unicode_tolower(uint32_t cpt);
 
+bool unicode_cpt_is_han(uint32_t cpt);
+
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

examples/whisper.android.java/app/src/main/jni/whisper/CMakeLists.txt

@@ -5,6 +5,17 @@ project(whisper.cpp)
 set(CMAKE_CXX_STANDARD 17)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)
 
+# Get whisper.cpp version
+file(READ "${WHISPER_LIB_DIR}/CMakeLists.txt" MAIN_CMAKE_CONTENT)
+string(REGEX MATCH "project\\(\"whisper\\.cpp\" VERSION ([0-9]+\\.[0-9]+\\.[0-9]+)\\)" VERSION_MATCH "${MAIN_CMAKE_CONTENT}")
+if(CMAKE_MATCH_1)
+    set(WHISPER_VERSION ${CMAKE_MATCH_1} PARENT_SCOPE)
+else()
+    set(WHISPER_VERSION "unknown" PARENT_SCOPE)
+endif()
+
+message(STATUS " Whisper version: ${WHISPER_VERSION}")
+
 set(SOURCE_FILES
     ${WHISPER_LIB_DIR}/src/whisper.cpp
     ${CMAKE_SOURCE_DIR}/jni.c
@@ -26,6 +37,7 @@ function(build_library target_name)
 
     target_link_libraries(${target_name} ${LOG_LIB} android ggml)
     target_compile_definitions(${target_name} PUBLIC GGML_USE_CPU)
+    target_compile_definitions(${target_name} PRIVATE WHISPER_VERSION="${WHISPER_VERSION}")
 
     if (${target_name} STREQUAL "whisper_v8fp16_va")
         target_compile_options(${target_name} PRIVATE -march=armv8.2-a+fp16)

examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -5,6 +5,17 @@ project(whisper.cpp)
 set(CMAKE_CXX_STANDARD 17)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../..)
 
+# Get whisper.cpp version
+file(READ "${WHISPER_LIB_DIR}/CMakeLists.txt" MAIN_CMAKE_CONTENT)
+string(REGEX MATCH "project\\(\"whisper\\.cpp\" VERSION ([0-9]+\\.[0-9]+\\.[0-9]+)\\)" VERSION_MATCH "${MAIN_CMAKE_CONTENT}")
+if(CMAKE_MATCH_1)
+    set(WHISPER_VERSION ${CMAKE_MATCH_1} PARENT_SCOPE)
+else()
+    set(WHISPER_VERSION "unknown" PARENT_SCOPE)
+endif()
+
+message(STATUS " Whisper version: ${WHISPER_VERSION}")
+
 # Path to external GGML, otherwise uses the copy in whisper.cpp.
 option(GGML_HOME "whisper: Path to external GGML source" OFF)
 
@@ -26,6 +37,7 @@ function(build_library target_name)
     )
 
     target_compile_definitions(${target_name} PUBLIC GGML_USE_CPU)
+    target_compile_definitions(${target_name} PRIVATE WHISPER_VERSION="${WHISPER_VERSION}")
 
     if (${target_name} STREQUAL "whisper_v8fp16_va")
         target_compile_options(${target_name} PRIVATE -march=armv8.2-a+fp16)

examples/whisper.wasm/README.md

@@ -22,7 +22,7 @@ audio is limited to 120 seconds.
 
 ## Live demo
 
-Link: https://ggerganov.github.io/whisper.cpp/
+Link: https://ggml.ai/whisper.cpp/
 
 ![image](https://user-images.githubusercontent.com/1991296/197348344-1a7fead8-3dae-4922-8b06-df223a206603.png)
 
@@ -52,6 +52,16 @@ cp bin/libmain.js        /path/to/html/
 cp bin/libmain.worker.js /path/to/html/
 ```
 
+> 📝 **Note:** By default this example is built with `WHISPER_WASM_SINGLE_FILE=ON`
+> which means that that a separate .wasm file will not be generated. Instead, the
+> WASM module is embedded in the main JS file as a base64 encoded string. To
+> generate a separate .wasm file, you need to disable this option by passing
+> `-DWHISPER_WASM_SINGLE_FILE=OFF`:
+> ```console
+> emcmake cmake .. -DWHISPER_WASM_SINGLE_FILE=OFF
+> ```
+> This will generate a `libmain.wasm` file in the build/bin directory.
+
 > 📝 **Note:** As of Emscripten 3.1.58 (April 2024), separate worker.js files are no
 > longer generated and the worker is embedded in the main JS file. So the worker
 > file will not be geneated for versions later than `3.1.58`.

ggml/CMakeLists.txt

@@ -131,6 +131,7 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_NNPA             "ggml: enable nnpa"             OFF)  # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -173,6 +174,8 @@ option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental,
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
 option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
+option(GGML_MUSA_GRAPHS                     "ggml: use MUSA graph, experimental, unstable"    OFF)
+option(GGML_MUSA_MUDNN_COPY                 "ggml: enable muDNN for accelerated copy"         OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -180,7 +183,8 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
-option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
+option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
+option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -265,12 +269,12 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cpp.h
     include/ggml-cuda.h
-    include/ggml-kompute.h
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
     include/ggml-vulkan.h
+    include/ggml-webgpu.h
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
@@ -359,6 +363,13 @@ write_basic_package_version_file(
     VERSION ${GGML_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
+target_compile_definitions(ggml-base PRIVATE
+    GGML_VERSION="${GGML_INSTALL_VERSION}"
+    GGML_COMMIT="${GGML_BUILD_COMMIT}"
+)
+message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
+message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")
+
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)

ggml/cmake/ggml-config.cmake.in

@@ -1,152 +1,189 @@
+@PACKAGE_INIT@
 
 @GGML_VARIABLES_EXPANDED@
 
-@PACKAGE_INIT@
-
-set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
-set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
-#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
-
-find_package(Threads REQUIRED)
-
-find_library(GGML_LIBRARY ggml
-    REQUIRED
-    HINTS ${GGML_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH)
-
-add_library(ggml::ggml UNKNOWN IMPORTED)
-set_target_properties(ggml::ggml
-    PROPERTIES
-        IMPORTED_LOCATION "${GGML_LIBRARY}")
-
-find_library(GGML_BASE_LIBRARY ggml-base
-    REQUIRED
-    HINTS ${GGML_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH)
-
-add_library(ggml::ggml-base UNKNOWN IMPORTED)
-set_target_properties(ggml::ggml-base
-    PROPERTIES
-        IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
-
+# Find all dependencies before creating any target.
+include(CMakeFindDependencyMacro)
+find_dependency(Threads)
 if (NOT GGML_SHARED_LIB)
+    set(GGML_CPU_INTERFACE_LINK_LIBRARIES "")
+    set(GGML_CPU_INTERFACE_LINK_OPTIONS   "")
+
     if (APPLE AND GGML_ACCELERATE)
-        find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
+        if(NOT ACCELERATE_FRAMEWORK)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${ACCELERATE_FRAMEWORK})
     endif()
 
-    if (GGML_OPENMP)
-        find_package(OpenMP REQUIRED)
+    if (GGML_OPENMP_ENABLED)
+        find_dependency(OpenMP)
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
     endif()
 
     if (GGML_CPU_HBM)
-        find_library(memkind memkind REQUIRED)
+        find_library(memkind memkind)
+        if(NOT memkind)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES memkind)
     endif()
 
     if (GGML_BLAS)
-        find_package(BLAS REQUIRED)
+        find_dependency(BLAS)
         list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${BLAS_LIBRARIES})
         list(APPEND GGML_CPU_INTERFACE_LINK_OPTIONS   ${BLAS_LINKER_FLAGS})
     endif()
 
     if (GGML_CUDA)
-        find_package(CUDAToolkit REQUIRED)
+        set(GGML_CUDA_INTERFACE_LINK_LIBRARIES "")
+        find_dependency(CUDAToolkit)
+        if (GGML_STATIC)
+            list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cudart_static>)
+            if (WIN32)
+                list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas> $<LINK_ONLY:CUDA::cublasLt>)
+            else()
+                list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cublas_static> $<LINK_ONLY:CUDA::cublasLt_static>)
+            endif()
+        endif()
+        if (NOT GGML_CUDA_NO_VMM)
+            list(APPEND GGML_CUDA_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:CUDA::cuda_driver>)
+        endif()
     endif()
 
     if (GGML_METAL)
-        find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-        find_library(METAL_FRAMEWORK    Metal REQUIRED)
-        find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+        find_library(FOUNDATION_LIBRARY Foundation)
+        find_library(METAL_FRAMEWORK    Metal)
+        find_library(METALKIT_FRAMEWORK MetalKit)
+        if(NOT FOUNDATION_LIBRARY OR NOT METAL_FRAMEWORK OR NOT METALKIT_FRAMEWORK)
+            set(${CMAKE_FIND_PACKAGE_NAME}_FOUND 0)
+            return()
+        endif()
+        set(GGML_METAL_INTERFACE_LINK_LIBRARIES
+            ${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    endif()
 
-        list(APPEND GGML_METAL_INTERFACE_LINK_LIBRARIES
-                    ${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    if (GGML_OPENCL)
+        find_dependency(OpenCL)
+        set(GGML_OPENCL_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:OpenCL::OpenCL>)
     endif()
 
     if (GGML_VULKAN)
-        find_package(Vulkan REQUIRED)
-        list(APPEND GGML_VULKAN_INTERFACE_LINK_LIBRARIES Vulkan::Vulkan)
+        find_dependency(Vulkan)
+        set(GGML_VULKAN_INTERFACE_LINK_LIBRARIES $<LINK_ONLY:Vulkan::Vulkan>)
     endif()
 
     if (GGML_HIP)
-        find_package(hip     REQUIRED)
-        find_package(hipblas REQUIRED)
-        find_package(rocblas REQUIRED)
-        list(APPEND GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
+        find_dependency(hip)
+        find_dependency(hipblas)
+        find_dependency(rocblas)
+        set(GGML_HIP_INTERFACE_LINK_LIBRARIES hip::host roc::rocblas roc::hipblas)
     endif()
 
     if (GGML_SYCL)
+        set(GGML_SYCL_INTERFACE_LINK_LIBRARIES "")
         find_package(DNNL)
         if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
             list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES DNNL::dnnl)
         endif()
         if (WIN32)
-            find_package(IntelSYCL REQUIRED)
-            find_package(MKL       REQUIRED)
+            find_dependency(IntelSYCL)
+            find_dependency(MKL)
             list(APPEND GGML_SYCL_INTERFACE_LINK_LIBRARIES IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
         endif()
     endif()
 endif()
 
-set(_ggml_all_targets "")
-foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
-    string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
-    string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
+set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
+set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
+#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
 
-    find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
+if(NOT TARGET ggml::ggml)
+    find_package(Threads REQUIRED)
+
+    find_library(GGML_LIBRARY ggml
         REQUIRED
         HINTS ${GGML_LIB_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
-    message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
-
-    add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
-    set_target_properties(ggml::${_ggml_backend}
+    add_library(ggml::ggml UNKNOWN IMPORTED)
+    set_target_properties(ggml::ggml
         PROPERTIES
-            INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
-            IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
-            IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
-            INTERFACE_COMPILE_FEATURES c_std_90
-            POSITION_INDEPENDENT_CODE ON)
+            IMPORTED_LOCATION "${GGML_LIBRARY}")
 
-    string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
-    if(is_cpu_variant)
-        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
-        set_target_properties(ggml::${_ggml_backend}
-           PROPERTIES
-               INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+    find_library(GGML_BASE_LIBRARY ggml-base
+        REQUIRED
+        HINTS ${GGML_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH)
 
-        if(GGML_CPU_INTERFACE_LINK_OPTIONS)
-            set_target_properties(ggml::${_ggml_backend}
-                PROPERTIES
-                    INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
-        endif()
+    add_library(ggml::ggml-base UNKNOWN IMPORTED)
+    set_target_properties(ggml::ggml-base
+        PROPERTIES
+            IMPORTED_LOCATION "${GGML_BASE_LIBRARY}")
 
-    else()
-        list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+    set(_ggml_all_targets "")
+    foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
+        string(REPLACE "-" "_" _ggml_backend_pfx "${_ggml_backend}")
+        string(TOUPPER "${_ggml_backend_pfx}" _ggml_backend_pfx)
+
+        find_library(${_ggml_backend_pfx}_LIBRARY ${_ggml_backend}
+            REQUIRED
+            HINTS ${GGML_LIB_DIR}
+            NO_CMAKE_FIND_ROOT_PATH)
+
+        message(STATUS "Found ${${_ggml_backend_pfx}_LIBRARY}")
+
+        add_library(ggml::${_ggml_backend} UNKNOWN IMPORTED)
         set_target_properties(ggml::${_ggml_backend}
             PROPERTIES
-                INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+                INTERFACE_INCLUDE_DIRECTORIES "${GGML_INCLUDE_DIR}"
+                IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+                IMPORTED_LOCATION "${${_ggml_backend_pfx}_LIBRARY}"
+                INTERFACE_COMPILE_FEATURES c_std_90
+                POSITION_INDEPENDENT_CODE ON)
 
-        if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+        string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
+        if(is_cpu_variant)
+            list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
+            set_target_properties(ggml::${_ggml_backend}
+            PROPERTIES
+                INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
+
+            if(GGML_CPU_INTERFACE_LINK_OPTIONS)
+                set_target_properties(ggml::${_ggml_backend}
+                    PROPERTIES
+                        INTERFACE_LINK_OPTIONS "${GGML_CPU_INTERFACE_LINK_OPTIONS}")
+            endif()
+
+        else()
+            list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
             set_target_properties(ggml::${_ggml_backend}
                 PROPERTIES
-                    INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+                    INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
+
+            if(${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS)
+                set_target_properties(ggml::${_ggml_backend}
+                    PROPERTIES
+                        INTERFACE_LINK_OPTIONS "${${_ggml_backend_pfx}_INTERFACE_LINK_OPTIONS}")
+            endif()
         endif()
-    endif()
 
-    list(APPEND _ggml_all_targets ggml::${_ggml_backend})
-endforeach()
+        list(APPEND _ggml_all_targets ggml::${_ggml_backend})
+    endforeach()
 
-list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
-set_target_properties(ggml::ggml
-    PROPERTIES
-        INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
+    list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
+    set_target_properties(ggml::ggml
+        PROPERTIES
+            INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
 
-add_library(ggml::all INTERFACE IMPORTED)
-set_target_properties(ggml::all
-    PROPERTIES
-        INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
+    add_library(ggml::all INTERFACE IMPORTED)
+    set_target_properties(ggml::all
+        PROPERTIES
+            INTERFACE_LINK_LIBRARIES "${_ggml_all_targets}")
+
+endif()
 
 check_required_components(ggml)

ggml/include/ggml-backend.h

@@ -339,7 +339,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
 
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml/include/ggml-cpu.h

@@ -101,6 +101,7 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
+    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -133,6 +134,7 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

ggml/include/ggml-kompute.h

@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif

ggml/include/ggml-webgpu.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_WEBGPU_NAME "WebGPU"
+
+// Needed for examples in ggml
+GGML_BACKEND_API ggml_backend_t ggml_backend_webgpu_init(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_webgpu_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml.h

@@ -314,6 +314,13 @@
 extern "C" {
 #endif
 
+    // Function type used in fatal error callbacks
+    typedef void (*ggml_abort_callback_t)(const char * error_message);
+
+    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+    // Returns the old callback for chaining
+    GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
+
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
@@ -470,6 +477,7 @@ extern "C" {
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
         GGML_OP_GET_ROWS_BACK,
+        GGML_OP_SET_ROWS,
         GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
@@ -481,12 +489,13 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
         GGML_OP_POOL_2D_BACK,
-        GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_UPSCALE,
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
         GGML_OP_ROLL,
@@ -519,6 +528,8 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
 
+        GGML_OP_GLU,
+
         GGML_OP_COUNT,
     };
 
@@ -542,6 +553,16 @@ extern "C" {
         GGML_UNARY_OP_COUNT,
     };
 
+    enum ggml_glu_op {
+        GGML_GLU_OP_REGLU,
+        GGML_GLU_OP_GEGLU,
+        GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_GEGLU_ERF,
+        GGML_GLU_OP_GEGLU_QUICK,
+
+        GGML_GLU_OP_COUNT,
+    };
+
     enum ggml_object_type {
         GGML_OBJECT_TYPE_TENSOR,
         GGML_OBJECT_TYPE_GRAPH,
@@ -627,6 +648,9 @@ extern "C" {
 
     // misc
 
+    GGML_API const char * ggml_version(void);
+    GGML_API const char * ggml_commit(void);
+
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
     GGML_API int64_t ggml_time_ms(void);
     GGML_API int64_t ggml_time_us(void);
@@ -657,6 +681,7 @@ extern "C" {
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
     GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
@@ -687,6 +712,9 @@ extern "C" {
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
+    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
+    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -758,6 +786,7 @@ extern "C" {
     GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
     GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1086,6 +1115,89 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // gated linear unit ops
+    // A: n columns, r rows,
+    // result is n / 2 columns, r rows,
+    // expects gate in second half of row, unless swapped is true
+    GGML_API struct ggml_tensor * ggml_glu(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_glu_op     op,
+             bool                 swapped);
+
+    GGML_API struct ggml_tensor * ggml_reglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_reglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // A: n columns, r rows,
+    // B: n columns, r rows,
+    GGML_API struct ggml_tensor * ggml_glu_split(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             struct ggml_tensor * b,
+             enum ggml_glu_op     op);
+
+    GGML_API struct ggml_tensor * ggml_reglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1185,6 +1297,19 @@ extern "C" {
             struct ggml_tensor  * a,
             float                 s);
 
+    // x = s * a + b
+    GGML_API struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    GGML_API struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
             struct ggml_context * ctx,
@@ -1375,6 +1500,23 @@ extern "C" {
             struct ggml_tensor  * b,  // row indices
             struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
+    // a TD  [n_embd, ne1,    ne2,    ne3]
+    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
+    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
+    //
+    // undefined behavior if destination rows overlap
+    //
+    // broadcast:
+    //   ne2 % ne11 == 0
+    //   ne3 % ne12 == 0
+    //
+    // return view(a)
+    GGML_API struct ggml_tensor * ggml_set_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // destination
+            struct ggml_tensor  * b,  // source
+            struct ggml_tensor  * c); // row indices
+
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
         struct ggml_tensor      * a);
@@ -1412,8 +1554,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
@@ -1723,6 +1871,17 @@ extern "C" {
             struct ggml_tensor  * b,
             int                   stride);
 
+    GGML_API struct ggml_tensor * ggml_conv_2d_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+            struct ggml_tensor  * b,   // input data [W, H, C, N]
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -1765,6 +1924,12 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+
+        GGML_SCALE_MODE_COUNT
+    };
+
+    enum ggml_scale_flag {
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
     };
 
     // interpolate
@@ -1777,14 +1942,26 @@ extern "C" {
 
     // interpolate
     // interpolate scale to specified dimensions
-    GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
             int                   ne3,
-            enum ggml_scale_mode  mode);
+            enum ggml_scale_mode  mode),
+        "use ggml_interpolate instead");
+
+    // Up- or downsamples the input to the specified size.
+    // 2D scale modes (eg. bilinear) are applied to the first two dimensions.
+    GGML_API struct ggml_tensor * ggml_interpolate(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            uint32_t              mode); // ggml_scale_mode [ | ggml_scale_flag...]
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1847,11 +2024,17 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    1]
-    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
-    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1890,7 +2073,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:

ggml/src/CMakeLists.txt

@@ -365,12 +365,12 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
-ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
+ggml_add_backend(WebGPU)
 ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)

ggml/src/ggml-alloc.c

@@ -22,21 +22,6 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
     return t->view_src != NULL;
 }
 
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
 // ops that return true for this function must not use restrict pointers for their backend implementations
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {

ggml/src/ggml-backend-reg.cpp

@@ -45,6 +45,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef GGML_USE_WEBGPU
+#include "ggml-webgpu.h"
+#endif
+
 #ifdef GGML_USE_OPENCL
 #include "ggml-opencl.h"
 #endif
@@ -61,10 +65,6 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -177,6 +177,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_VULKAN
         register_backend(ggml_backend_vk_reg());
 #endif
+#ifdef GGML_USE_WEBGPU
+        register_backend(ggml_backend_webgpu_reg());
+#endif
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
@@ -189,9 +192,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -575,7 +575,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);

ggml/src/ggml-backend.cpp

@@ -352,21 +352,6 @@ ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
 
 // backend copy
 
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
 void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
@@ -662,6 +647,7 @@ struct ggml_backend_sched {
     // pipeline parallelism support
     int n_copies;
     int cur_copy;
+    int next_copy;
     ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
     struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
     int n_graph_inputs;
@@ -817,8 +803,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1447,8 +1434,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         }
     }
 
-    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
-
     return GGML_STATUS_SUCCESS;
 }
 
@@ -1549,10 +1534,10 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
-    ggml_backend_sched_split_graph(sched, measure_graph);
-
     ggml_backend_sched_synchronize(sched);
 
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
@@ -1564,6 +1549,10 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
+    GGML_ASSERT(!sched->is_alloc);
+
+    sched->cur_copy = sched->next_copy;
+    sched->next_copy = (sched->next_copy + 1) % sched->n_copies;
 
     ggml_backend_sched_split_graph(sched, graph);
 
@@ -1604,7 +1593,7 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
         // if the graph is not already allocated, always use copy 0 after a synchronization
         // this ensures that during generation the same copy is used every time,
         // which avoids changes in the graph that could cause CUDA or other graphs to be disabled
-        sched->cur_copy = 0;
+        sched->next_copy = 0;
     }
 }
 
@@ -1826,7 +1815,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
     ggml_free(copy.ctx_unallocated);
 }
 
-bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node) {
     struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
     if (copy.buffer == NULL) {
         return false;
@@ -1837,28 +1826,45 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     assert(g1->n_nodes == g2->n_nodes);
 
-    for (int i = 0; i < g1->n_nodes; i++) {
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
+    if (test_node != nullptr) {
+        // Compute the whole graph and only test the output for a specific tensor
+        ggml_backend_graph_compute(backend1, g1);
+        ggml_backend_graph_compute(backend2, g2);
 
-        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-        ggml_backend_graph_compute(backend1, &g1v);
-        ggml_backend_graph_compute(backend2, &g2v);
-
-        if (ggml_is_view_op(t1->op)) {
-            continue;
+        int test_node_idx = -1;
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            if (t1 == test_node) {
+                test_node_idx = i;
+                break;
+            }
         }
+        GGML_ASSERT(test_node_idx != -1);
 
-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
+        callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
+    } else {
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            struct ggml_tensor * t2 = g2->nodes[i];
+
+            assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+            struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+            struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+            ggml_backend_graph_compute(backend1, &g1v);
+            ggml_backend_graph_compute(backend2, &g2v);
+
+            if (ggml_is_view_op(t1->op)) {
+                continue;
+            }
+
+            // compare results, calculate rms etc
+            if (!callback(i, t1, t2, user_data)) {
+                break;
+            }
         }
     }
-
     ggml_backend_graph_copy_free(copy);
 
     return true;

ggml/src/ggml-cann/acl_tensor.cpp

@@ -77,6 +77,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     for (int i = 0; i < final_dims; i++) {
         acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
     }
+    size_t elem_offset = offset / ggml_element_size(tensor);
+    acl_storage_len += elem_offset;
 
     // Reverse ne and stride.
     std::reverse(acl_ne, acl_ne + final_dims);
@@ -84,7 +86,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
 
     aclTensor* acl_tensor = aclCreateTensor(
         acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
-        offset / ggml_element_size(tensor), format, &acl_storage_len, 1,
+        elem_offset, format, &acl_storage_len, 1,
         tensor->data);
 
     return acl_tensor;

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,8 +65,9 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v2.h>
+#include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_zero.h>
 #include <float.h>
 
 #include <cmath>
@@ -98,7 +99,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclT
     }
 }
 
-void ggml_cann_unary_op(
+void ggml_cann_op_unary(
     std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
     ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
@@ -110,6 +111,42 @@ void ggml_cann_unary_op(
     ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
+void ggml_cann_op_unary_gated(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor *acl_src0 = nullptr, *acl_src1 = nullptr;
+    if(src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+
+        acl_src0 = ggml_cann_create_tensor(src0);
+        acl_src1 = ggml_cann_create_tensor(src1);
+    } else {
+        int64_t ne[] = {src0->ne[0] / 2, src0->ne[1], src0->ne[2], src0->ne[3]};
+        size_t nb[] = {src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]};
+        acl_src0 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, 0);
+        acl_src1 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, ne[0] * ggml_element_size(src0));
+        if (swapped) {
+            std::swap(acl_src0, acl_src1);
+        }
+    }
+
+    unary_op(ctx, acl_src0, acl_dst);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst, acl_src1);
+
+    ggml_cann_release_resources(ctx, acl_src0, acl_dst);
+    if(src1)
+        ggml_cann_release_resources(ctx, acl_src1);
+}
+
 /**
  * @brief Repeats elements of a tensor along each dimension according to the
  * specified repeat array.
@@ -804,10 +841,11 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
         nb[i] = nb[i - 1] * ne[i - 1];
     }
 
-    ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
     aclTensor* zero =
         ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
     return zero;
+    GGML_UNUSED(n_bytes);
 }
 
 /**
@@ -1783,8 +1821,27 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_weight_tensor;
+
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
+    if (weightToNZ && is_matmul_weight(weight)) {
+        int64_t acl_stride[2] = {1, transpose_ne[1]};
+
+        // Reverse ne.
+        std::reverse(transpose_ne, transpose_ne + n_dims);
+
+        std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
+
+        acl_weight_tensor = aclCreateTensor(
+            transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
+            0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
+    } else {
+        acl_weight_tensor =
+            ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
+    }
     aclTensor* acl_dst =
         ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
 
@@ -2654,6 +2711,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
     }
 
+#ifdef ASCEND_310P
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    if (src0->type == GGML_TYPE_F16) {
+        src0_row.type = GGML_TYPE_F32;
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = ori_src0_nb[0];
+    src0_row.nb[1] = ori_src0_nb[1];
+    src0_row.nb[2] = ori_src0_nb[1];
+    src0_row.nb[3] = ori_src0_nb[1];
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            src0_row.data = src0_tmp_ptr;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+#endif
+
     std::vector<aclTensor*> src0_tensor_vec;
     std::vector<aclTensor*> src1_tensor_vec;
     std::vector<aclTensor*> dst_tensor_vec;
@@ -2701,9 +2819,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     }
 
     size_t GROUP_SIZE = 128;
-    // GroupedMatmulV2 required tensor_list.size < 128
+    // GroupedMatmulV3 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV2
+        // split and call GroupedMatmulV3
         size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
         std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
         std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2713,7 +2831,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
         aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
             nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
 
         ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <unordered_set>
 #include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
@@ -1020,6 +1021,37 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  */
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
+ *
+ * @details Checks whether the given tensor serves as weight parameters in matrix multiplication operations,
+ *          typically within neural network layers. The function maintains a static set of canonical weight
+ *          naming suffixes from Transformer-based architectures. Uses substring matching to identify weight
+ *          tensors even with hierarchical naming patterns.
+ *
+ * @param tensor Pointer to the target ggml_tensor object (const-qualified).
+ */
+static bool is_matmul_weight(const ggml_tensor* tensor) {
+    std::string name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{
+        "output.weight",
+        "attn_q.weight",
+        "attn_k.weight",
+        "attn_v.weight",
+        "attn_output.weight",
+        "ffn_gate.weight",
+        "ffn_up.weight",
+        "ffn_down.weight"
+    };
+
+    for (const auto& suffix : weight_suffixes) {
+        if (name.find(suffix) != std::string::npos) {
+            return true;
+        }
+    }
+    return false;
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.
@@ -1066,7 +1098,7 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param dst The destination tensor. Its src[0] is treated as the input tensor.
  */
 template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
-    void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
     aclTensor* acl_src = ggml_cann_create_tensor(src);
@@ -1077,49 +1109,125 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
 }
 
 /**
- * @brief   Applies a unary operation to a ggml tensor using the CANN backend.
+ * @brief Applies a unary operation to a ggml tensor using the CANN backend.
  *
- * @details This function performs a unary operation on the input tensor using
- * a user-provided lambda or callable object `unary_op`, which accepts the CANN
- * context and two ACL tensors (source and destination). Internally, this function
- * creates ACL representations of the ggml tensors and invokes the unary operation.
- * The result is stored in the destination tensor `dst`. This utility abstracts the
- * common boilerplate of tensor conversion and cleanup when implementing unary ops.
+ * @details This function applies a unary operation to the input tensor using
+ * a user-provided lambda or callable `unary_op`. The lambda receives the
+ * CANN backend context and two ACL tensors: the source and the destination.
  *
- * @param unary_op A callable that performs the unary operation using CANN APIs.
- * @param ctx The CANN context used for operations.
- * @param dst The destination tensor where the result will be stored.
- *            The source tensor is retrieved from `dst->src[0]`.
+ * Internally, this function handles the conversion from GGML tensors to ACL tensors,
+ * calls the provided unary op, and manages resource cleanup. The input is assumed
+ * to be `dst->src[0]`, and the result is written to `dst`.
+ *
+ * This utility simplifies writing unary op wrappers by abstracting tensor preparation.
+ *
+ * @param unary_op A callable that performs the unary operation using CANN ACL APIs.
+ * @param ctx The CANN context for operation execution.
+ * @param dst The destination ggml_tensor where the result will be stored.
+ *            The input tensor is assumed to be `dst->src[0]`.
+ *
+ * @see GGML_CANN_CALL_OP_UNARY
  */
-void ggml_cann_unary_op(
+void ggml_cann_op_unary(
     std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
     ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
 /**
- * @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
+ * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
  *
- * This macro defines an inline lambda wrapping a specific ACL operation name,
- * and passes it to the templated ggml_cann_unary_op function. It simplifies
- * calling unary ops by hiding the lambda boilerplate.
+ * @details This function performs a gated activation such as GEGLU or ReGLU.
+ * It supports two input modes:
+ *
+ * 1. **Dual input mode**: `dst->src[0]` and `dst->src[1]` are both valid tensors.
+ *    These are used directly as the value and gate tensors.
+ *
+ * 2. **Packed input mode**: Only `dst->src[0]` is valid, and it is assumed to
+ *    contain a concatenation of value and gate along the first dimension. This tensor
+ *    will be split into two equal halves to form the value and gate inputs.
+ *
+ * The function applies a user-provided unary operation (e.g., GELU) to the value tensor,
+ * then multiplies the result in-place with the gate tensor:
  *
- * Internally, the lambda will call:
  * @code
- * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * dst = unary_op(value) * gate;
  * @endcode
  *
+ * The `swapped` parameter (from `dst->op_params[1]`) allows flipping the
+ * order of value/gate in the packed input case.
+ *
+ * @param unary_op A callable that performs the unary operation using CANN ACL APIs.
+ *                 It receives (ctx, acl_value_tensor, acl_output_tensor).
+ * @param ctx      The CANN context used for execution.
+ * @param dst      The destination ggml_tensor. Source tensors are in `dst->src[0]` and optionally `src[1]`.
+ *
+ * @see GGML_CANN_CALL_OP_UNARY_GATED
+ */
+void ggml_cann_op_unary_gated(
+    std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
+    ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
+ *
+ * This macro wraps the specified ACLNN unary operator name into a lambda expression,
+ * and passes it to `ggml_cann_op_unary`, which handles the common logic for executing
+ * unary ops in the CANN backend.
+ *
+ * Internally, this macro expands to a lambda like:
+ * @code
+ * [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
+ *     GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * };
+ * @endcode
+ *
+ * This lambda is then passed to `ggml_cann_op_unary`, which applies the operation.
+ *
  * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
  *
- * @see ggml_cann_unary_op
+ * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                              \
     do {                                                              \
         auto lambda = [](ggml_backend_cann_context& ctx,              \
             aclTensor* acl_src,                                       \
             aclTensor* acl_dst) {                                     \
             GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
         };                                                            \
-        ggml_cann_unary_op(lambda, ctx, dst);                         \
+        ggml_cann_op_unary(lambda, ctx, dst);                         \
     }                                                                 \
     while (0)
+
+/**
+ * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
+ *
+ * This macro wraps the specified ACLNN unary operator name into a lambda expression,
+ * and passes it to `ggml_cann_op_unary_gated`, which handles the common logic for
+ * executing gated unary ops in the CANN backend.
+ *
+ * Internally, this macro expands to a lambda like:
+ * @code
+ * [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
+ *     GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * };
+ * @endcode
+ *
+ * This lambda is then passed to `ggml_cann_op_unary_gated`, which applies the operation.
+ *
+ * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
+ *
+ * @see ggml_cann_op_unary_gated
+ * @see GGML_CANN_CALL_ACLNN_OP
+ */
+#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                        \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_op_unary_gated(lambda, ctx, dst);                   \
+    }                                                                 \
+    while (0)
+
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -359,7 +359,7 @@ struct ggml_backend_cann_context {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
-        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
+        async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -24,6 +24,7 @@
 
 #include <acl/acl.h>
 #include <stdarg.h>
+#include <aclnnop/aclnn_trans_matmul_weight.h>
 
 #include <cmath>
 #include <cstdio>
@@ -1115,6 +1116,63 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
+static int CreateAclTensorWeight(const void *hostData, const std::vector<int64_t> &shape, void **deviceAddr,
+                      aclDataType dataType, aclTensor **tensor)
+{
+    uint64_t size = 1;
+    for (auto i : shape) {
+        size *= i;
+    }
+
+    const aclIntArray *mat2Size = aclCreateIntArray(shape.data(), shape.size());
+    ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(mat2Size, dataType, &size));
+
+    size *= sizeof(int16_t);
+
+    ACL_CHECK(aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+    aclrtMemcpy(*deviceAddr, size, hostData, size, ACL_MEMCPY_HOST_TO_DEVICE);
+
+    std::vector<int64_t> strides(shape.size(), 1);
+    for (int64_t i = shape.size() - 2; i >= 0; i--) {
+        strides[i] = shape[i + 1] * strides[i + 1];
+    }
+
+    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
+                              shape.data(), shape.size(), *deviceAddr);
+    return 0;
+}
+
+static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
+    aclrtStream stream;
+    ACL_CHECK(aclrtCreateStream(&stream));
+
+    std::vector<int64_t> weightTransposedShape = {tensor->ne[1], tensor->ne[0]};
+    void *weightTransposedDeviceAddr = nullptr;
+    aclTensor *weightTransposed = nullptr;
+    CreateAclTensorWeight(data, weightTransposedShape, &weightTransposedDeviceAddr,
+                          ggml_cann_type_mapping(tensor->type), &weightTransposed);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor *executor;
+    void *workspaceAddr = nullptr;
+
+    // TransMatmulWeight
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed, &workspaceSize, &executor));
+    std::unique_ptr<void, aclError (*)(void *)> workspaceAddrPtrTrans(nullptr, aclrtFree);
+    if (workspaceSize > 0) {
+        ACL_CHECK(aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
+        workspaceAddrPtrTrans.reset(workspaceAddr);
+    }
+    ACL_CHECK(aclnnTransMatmulWeight(workspaceAddr, workspaceSize, executor, stream));
+
+    size_t size = ggml_nelements(tensor) * ggml_element_size(tensor);
+
+    aclrtMemcpy((char *)tensor->data + offset, size,
+                weightTransposedDeviceAddr, size, ACL_MEMCPY_HOST_TO_DEVICE);
+    ACL_CHECK(aclDestroyTensor(weightTransposed));
+    aclrtFree(weightTransposedDeviceAddr);
+}
+
 // TODO: need handle tensor which has paddings.
 /**
  * @brief Set tensor data in a CANN buffer.
@@ -1139,9 +1197,16 @@ static void ggml_backend_cann_buffer_set_tensor(
     // For acl, synchronous functions use this default stream.
     // Why aclrtSynchronizeDevice?
 
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
+        if (weightToNZ && is_matmul_weight((const ggml_tensor*)tensor)) {
+            weight_format_to_nz(tensor, data, offset);
+        }
     } else {
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
@@ -1616,16 +1681,18 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(dst)) {
                 case GGML_UNARY_OP_ABS:
-                    GGML_CANN_CALL_UNARY_OP(Abs);
+                    GGML_CANN_CALL_OP_UNARY(Abs);
                     break;
                 case GGML_UNARY_OP_NEG:
-                    GGML_CANN_CALL_UNARY_OP(Neg);
+                    GGML_CANN_CALL_OP_UNARY(Neg);
                     break;
                 case GGML_UNARY_OP_GELU:
-                    GGML_CANN_CALL_UNARY_OP(Gelu);
+                case GGML_UNARY_OP_GELU_ERF:
+                    // aclnnGelu internally uses the erf-based approximation.
+                    GGML_CANN_CALL_OP_UNARY(Gelu);
                     break;
                 case GGML_UNARY_OP_SILU:
-                    GGML_CANN_CALL_UNARY_OP(Silu);
+                    GGML_CANN_CALL_OP_UNARY(Silu);
                     break;
                 case GGML_UNARY_OP_GELU_QUICK: {
                     auto lambda = [](ggml_backend_cann_context& ctx,
@@ -1633,31 +1700,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                         aclTensor* acl_dst) {
                         GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
                     };
-                    ggml_cann_unary_op(lambda, ctx, dst);
+                    ggml_cann_op_unary(lambda, ctx, dst);
                 } break;
                 case GGML_UNARY_OP_TANH:
-                    GGML_CANN_CALL_UNARY_OP(Tanh);
+                    GGML_CANN_CALL_OP_UNARY(Tanh);
                     break;
                 case GGML_UNARY_OP_RELU:
-                    GGML_CANN_CALL_UNARY_OP(Relu);
+                    GGML_CANN_CALL_OP_UNARY(Relu);
                     break;
                 case GGML_UNARY_OP_SIGMOID:
-                    GGML_CANN_CALL_UNARY_OP(Sigmoid);
+                    GGML_CANN_CALL_OP_UNARY(Sigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSIGMOID:
-                    GGML_CANN_CALL_UNARY_OP(Hardsigmoid);
+                    GGML_CANN_CALL_OP_UNARY(Hardsigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSWISH:
-                    GGML_CANN_CALL_UNARY_OP(Hardswish);
+                    GGML_CANN_CALL_OP_UNARY(Hardswish);
                     break;
                 case GGML_UNARY_OP_EXP:
-                    GGML_CANN_CALL_UNARY_OP(Exp);
+                    GGML_CANN_CALL_OP_UNARY(Exp);
                     break;
                 case GGML_UNARY_OP_ELU:
                     ggml_cann_elu(ctx, dst);
                     break;
                 case GGML_UNARY_OP_SGN:
-                    GGML_CANN_CALL_UNARY_OP(Sign);
+                    GGML_CANN_CALL_OP_UNARY(Sign);
                     break;
                 case GGML_UNARY_OP_STEP:
                     ggml_cann_step(ctx, dst);
@@ -1666,6 +1733,31 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    GGML_CANN_CALL_OP_UNARY_GATED(Relu);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                    // aclnnGelu internally uses the erf-based approximation.
+                    GGML_CANN_CALL_OP_UNARY_GATED(Gelu);
+                    break;
+                case GGML_GLU_OP_SWIGLU:
+                    GGML_CANN_CALL_OP_UNARY_GATED(Silu);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK: {
+                    auto lambda = [](ggml_backend_cann_context& ctx,
+                        aclTensor* acl_src,
+                        aclTensor* acl_dst) {
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
+                    };
+                    ggml_cann_op_unary_gated(lambda, ctx, dst);
+                } break;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_NORM:
             ggml_cann_norm(ctx, dst);
             break;
@@ -1708,7 +1800,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_binary_op<aclnn_mul>(ctx, dst);
             break;
         case GGML_OP_SQRT:
-            GGML_CANN_CALL_UNARY_OP(Sqrt);
+            GGML_CANN_CALL_OP_UNARY(Sqrt);
             break;
         case GGML_OP_CLAMP:
             ggml_cann_clamp(ctx, dst);
@@ -1753,16 +1845,16 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_argmax(ctx, dst);
             break;
         case GGML_OP_COS:
-            ggml_cann_unary_op<aclnn_cos>(ctx, dst);
+            ggml_cann_op_unary<aclnn_cos>(ctx, dst);
             break;
         case GGML_OP_SIN:
-            ggml_cann_unary_op<aclnn_sin>(ctx, dst);
+            ggml_cann_op_unary<aclnn_sin>(ctx, dst);
             break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cann_conv_transpose_1d(ctx, dst);
             break;
         case GGML_OP_LOG:
-            GGML_CANN_CALL_UNARY_OP(Log);
+            GGML_CANN_CALL_OP_UNARY(Log);
             break;
         case GGML_OP_MEAN:
             ggml_cann_mean(ctx, dst);
@@ -2036,10 +2128,23 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_SGN:
                 case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_GELU_ERF:
                     return true;
                 default:
                     return false;
             }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return true;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
                 case GGML_TYPE_F16:
@@ -2086,6 +2191,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                return false;
+            } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@@ -2182,12 +2294,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
@@ -2205,6 +2315,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_SCALE:
+            float bias;
+            memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
+            return bias == 0.0f; // TODO: support bias != 0.0f
+        case GGML_OP_SOFT_MAX:
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2227,6 +2345,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -1,30 +0,0 @@
-file(GLOB SRC_FILES
-    get_row_f32.cpp
-    get_row_f16.cpp
-    get_row_q4_0.cpp
-    get_row_q8_0.cpp
-    quantize_f32_q8_0.cpp
-    quantize_f16_q8_0.cpp
-    quantize_float_to_q4_0.cpp
-    dup.cpp
-)
-
-set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
-set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
-
-if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
-    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
-elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
-    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
-else()
-    message(FATAL_ERROR "ascendc_kernel_cmake does not exist, please check whether the compiler package is installed.")
-endif()
-include(${ASCENDC_CMAKE_DIR}/ascendc.cmake)
-
-ascendc_library(ascendc_kernels STATIC
-    ${SRC_FILES}
-)
-
-message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
-ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
-# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/ascendc_kernels.h

@@ -1,19 +0,0 @@
-#ifndef ASCENDC_KERNELS_H
-#define ASCENDC_KERNELS_H
-
-#include "aclrtlaunch_ascendc_get_row_f32.h"
-#include "aclrtlaunch_ascendc_get_row_f16.h"
-#include "aclrtlaunch_ascendc_get_row_q8_0.h"
-#include "aclrtlaunch_ascendc_get_row_q4_0.h"
-
-#include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
-#include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
-#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
-#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
-
-#include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16.h"
-#include "aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32.h"
-
-#endif  // ASCENDC_KERNELS_H

ggml/src/ggml-cann/kernels/dup.cpp

@@ -1,234 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
-
-template <typename SRC_T, typename DST_T>
-class DupByRows {
-   public:
-    __aicore__ inline DupByRows() {}
-    __aicore__ inline void init(GM_ADDR src, GM_ADDR dst, int64_t *input_ne_ub,
-                                size_t *input_nb_ub) {
-        /* Dup by rows when src is contigous on first dimension and dst is
-        contiguous, each kernel process one row.
-        */
-
-        // Input has four dims.
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        // param
-        num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
-        num_elem = input_ne_ub[0];
-
-        // index for (ne[1], ne[2], ne[3]): (idx_ne1, idx_ne2, idx_ne3)
-        idx_ne3 = op_block_idx / (input_ne_ub[1] * input_ne_ub[2]);
-        idx_ne2 = (op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2]))
-                  / (input_ne_ub[1]);
-        idx_ne1 = op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2])
-                - idx_ne2 * input_ne_ub[1];
-
-        // src may not contiguous in dim [1,2,3], so stride decited by ne&nb
-        src_stride = input_nb_ub[3] * idx_ne3 + input_nb_ub[2] * idx_ne2
-                     + input_nb_ub[1] * idx_ne1;
-
-        // dst is contiguous
-        dst_stride = op_block_idx * (input_ne_ub[0] * sizeof(DST_T));
-
-        src_gm.SetGlobalBuffer(reinterpret_cast<__gm__ SRC_T *>(src +
-                                                                src_stride));
-        dst_gm.SetGlobalBuffer(reinterpret_cast<__gm__ DST_T *>(dst +
-                                                                dst_stride));
-
-        pipe.InitBuffer(src_queue, BUFFER_NUM, (sizeof(SRC_T) * num_elem +
-                                                32 - 1) / 32 * 32);
-        pipe.InitBuffer(dst_queue, BUFFER_NUM, (sizeof(DST_T) * num_elem +
-                                                32 - 1) / 32 * 32);
-    }
-
-    __aicore__ inline void copy_in() {
-        LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-        const size_t elem_per_block = 32 / sizeof(SRC_T);
-        size_t tail = num_elem % elem_per_block;
-        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
-        DataCopy(src_local, src_gm, cpy_elements_len);
-        src_queue.EnQue(src_local);
-    }
-
-    __aicore__ inline void copy_out() {
-        LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-#ifdef ASCEND_310P
-        const size_t elem_per_block = 32 / sizeof(DST_T);
-        size_t tail = num_elem % elem_per_block;
-        size_t len = num_elem & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(dst_gm, dst_local, len);
-        }
-        if(tail != 0) {
-            for (size_t i = tail; i < elem_per_block; i++) {
-                dst_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
-            SetAtomicNone();
-        }
-#else
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(DST_T);
-        DataCopyPad(dst_gm, dst_local, dataCopyParams);
-#endif
-        dst_queue.FreeTensor(dst_local);
-    }
-
-    __aicore__ inline void dup() {
-        // main process, copy one row data from src to dst.
-        copy_in();
-
-        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
-        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
-
-        int32_t BLOCK_NUM = 32 / sizeof(DST_T);
-        DataCopy(dst_local, src_local, (num_elem + BLOCK_NUM - 1)
-                                        / BLOCK_NUM * BLOCK_NUM);
-        dst_queue.EnQue<DST_T>(dst_local);
-
-        src_queue.FreeTensor(src_local);
-        copy_out();
-    }
-
-    __aicore__ inline void dup_with_cast() {
-        // main process, copy one row data from src to dst.
-        // cast dtype from src to dst.
-        copy_in();
-
-        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
-        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
-
-        Cast(dst_local, src_local, RoundMode::CAST_NONE, num_elem);
-        dst_queue.EnQue<DST_T>(dst_local);
-
-        src_queue.FreeTensor(src_local);
-        copy_out();
-    }
-
-   private:
-
-    TPipe pipe;
-    GlobalTensor<SRC_T> src_gm;
-    GlobalTensor<DST_T> dst_gm;
-
-    int64_t num_rows;
-    int64_t num_elem;
-    int64_t idx_ne3;
-    int64_t idx_ne2;
-    int64_t idx_ne1;
-    int64_t src_stride;
-    int64_t dst_stride;
-
-    TQue<QuePosition::VECIN, BUFFER_NUM> src_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> dst_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<half, half> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<float, float> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<float, half> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup_with_cast();
-}
-
-extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32(
-                                                        GM_ADDR src_gm,
-                                                        GM_ADDR dst_gm,
-                                                        GM_ADDR input_ne_gm,
-                                                        GM_ADDR input_nb_gm,
-                                                        GM_ADDR output_ne_gm,
-                                                        GM_ADDR output_nb_gm) {
-
-    // copy params from gm to ub.
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    DupByRows<half, float> op;
-    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
-    op.dup_with_cast();
-}

ggml/src/ggml-cann/kernels/get_row_f16.cpp

@@ -1,197 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-class GET_ROW_F16 {
-   public:
-    __aicore__ inline GET_ROW_F16() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
-                                int64_t *output_ne_ub, size_t *output_nb_ub) {
-        // TODO, use template for F16/f32
-        int64_t op_block_num = GetBlockNum();
-        op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ half *)input);
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        uint64_t input_local_buffer_size = ((input_ne[0] * sizeof(half) + 31)
-                                             & ~31);
-        uint64_t output_local_buffer_size = ((input_ne[0] * sizeof(float) + 31)
-                                              & ~31);
-
-        local_buffer_elems = input_local_buffer_size / sizeof(half);
-
-        // TODO, consider long row that can't put in UB.
-        // All data should asign to 32. It's ok because all data is align to 32.
-        pipe.InitBuffer(input_queue, BUFFER_NUM, input_local_buffer_size);
-        pipe.InitBuffer(output_queue, BUFFER_NUM, output_local_buffer_size);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
-        size_t origin_len = len;
-        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        const size_t elem_per_block = 32 / sizeof(half);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if(tail != 0) {
-            len += elem_per_block;
-        }
-        DataCopy(input_local, input_gm[offset], len);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(output_gm[offset], output_local, len);
-        }
-
-        if(tail != 0) {
-#ifdef ASCEND_310P
-            for (size_t i = tail; i < elem_per_block; i++) {
-                output_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
-            SetAtomicNone();
-#else
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPad(output_gm[offset + len], output_local[len],
-                        dataCopyParams);
-#endif
-        }
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_row(int64_t idx) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3];
-
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3];
-
-        copy_in(input_offset, input_ne[0]);
-        LocalTensor<half> input_local = input_queue.DeQue<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        Cast(output_local, input_local, RoundMode::CAST_NONE,
-             local_buffer_elems);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset, input_ne[0]);
-
-        input_queue.FreeTensor(input_local);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            calculate_row(i);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    size_t local_buffer_elems;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<half> input_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    int64_t op_block_idx;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_f16(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
-    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_F16 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
-            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/get_row_f32.cpp

@@ -1,190 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-class GET_ROW_F32 {
-   public:
-    __aicore__ inline GET_ROW_F32() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
-                                int64_t *output_ne_ub, size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ float *)input);
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        uint64_t local_buffer_size = ((input_ne[0] * sizeof(float) + 31) & ~31);
-        local_buffer_elems = local_buffer_size / sizeof(float);
-
-        // TODO, consider long row that can't put in UB.
-        // All data should asign to 32. It's ok because all data is align to 32.
-        pipe.InitBuffer(input_queue, BUFFER_NUM, local_buffer_size);
-        pipe.InitBuffer(output_queue, BUFFER_NUM, local_buffer_size);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
-        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if(tail != 0) {
-            len += elem_per_block;
-        }
-        DataCopy(input_local, input_gm[offset], len);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        const size_t elem_per_block = 32 / sizeof(float);
-        size_t tail = len % elem_per_block;
-        len = len & ~(elem_per_block - 1);
-        if (len > 0) {
-            DataCopy(output_gm[offset], output_local, len);
-        }
-
-        if(tail != 0) {
-#ifdef ASCEND_310P
-            for (size_t i = tail; i < elem_per_block; i++) {
-                output_local[len + i].SetValue(0, 0);
-            }
-            SetAtomicAdd<float>();
-            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
-            SetAtomicNone();
-#else
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPad(output_gm[offset + len], output_local[len],
-                        dataCopyParams);
-#endif
-        }
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_row(int64_t idx) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3];
-
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3];
-
-        copy_in(input_offset, input_ne[0]);
-        LocalTensor<float> input_local = input_queue.DeQue<float>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        DataCopy(output_local, input_local, local_buffer_elems);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset, input_ne[0]);
-
-        input_queue.FreeTensor(input_local);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            calculate_row(i);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    size_t local_buffer_elems;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<float> input_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    int64_t op_block_idx;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_f32(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
-    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_F32 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
-            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/get_row_q4_0.cpp

@@ -1,204 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support 4bit get row
-    extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
-        GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-        GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-        GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support 4bit get row.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-
-#define QK4_0 32
-
-class GET_ROW_Q4_0 {
-   public:
-    __aicore__ inline GET_ROW_Q4_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
-                                size_t *indices_nb_ub, int64_t *output_ne_ub,
-                                size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-            scale_ne[i] = input_ne_ub[i];
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // one scale for a group.
-        scale_ne[0] /= QK4_0;
-
-        input_stride[0] = 1;
-        scale_stride[0] = 1;
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        group_size_in_row = input_ne[0] / QK4_0;
-        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
-                               input_ne[3] / 2;
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ int4b_t *)input);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK4_0 * sizeof(int4b_t));
-        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK4_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK4_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<int4b_t> input_local = input_queue.AllocTensor<int4b_t>();
-        // 32 * sizeof(int4b_t) = 16, which is not aligned to 32, why no error?
-        DataCopy(input_local, input_gm[offset], QK4_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        DataCopy(output_gm[offset], output_local, QK4_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3] +
-                                     group * QK4_0;
-        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
-                                     indices_ne1_idx * scale_stride[2] +
-                                     indices_ne2_idx * scale_stride[3] + group;
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3] +
-                                      group * QK4_0;
-
-        copy_in(input_offset);
-        LocalTensor<int4b_t> input_local = input_queue.DeQue<int4b_t>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        // TODO: cast more data to speed up.
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
-        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
-
-        // Only mul need compile by group.
-        half scale = scale_gm.GetValue(scale_offset);
-
-        Muls(output_local, output_local, (float)scale, QK4_0);
-
-        input_queue.FreeTensor(input_local);
-        cast_queue.FreeTensor(cast_local);
-        output_queue.EnQue(output_local);
-
-        copy_out(output_offset);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                calculate_group(i, j);
-            }
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t scale_ne[4];
-    size_t scale_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t ir;
-    int64_t dr;
-
-    int64_t group_size_in_row;
-
-    TPipe pipe;
-    GlobalTensor<int4b_t> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_Q4_0 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
-            indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/get_row_q8_0.cpp

@@ -1,191 +0,0 @@
-#include "kernel_operator.h"
-
-// optimize me. Use template to avoid copy code.
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-
-#define QK8_0 32
-
-class GET_ROW_Q8_0 {
-   public:
-    __aicore__ inline GET_ROW_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
-                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
-                                size_t *indices_nb_ub, int64_t *output_ne_ub,
-                                size_t *output_nb_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            indices_ne[i] = indices_ne_ub[i];
-            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
-            scale_ne[i] = input_ne_ub[i];
-            output_ne[i] = output_ne_ub[i];
-            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
-        }
-
-        // one scale for a group.
-        scale_ne[0] /= QK8_0;
-
-        input_stride[0] = 1;
-        scale_stride[0] = 1;
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        group_size_in_row = input_ne[0] / QK8_0;
-        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
-                               input_ne[3] * sizeof(int8_t);
-
-        // Indices has two dims. n_elements = all rows should get.
-        // dr, all rows should this thread get.
-        uint64_t n_elements =
-            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
-        dr = n_elements / op_block_num;
-
-        uint64_t tails = n_elements % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        input_gm.SetGlobalBuffer((__gm__ int8_t *)input);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
-        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
-        output_gm.SetGlobalBuffer((__gm__ float *)output);
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK8_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<int8_t> input_local = input_queue.AllocTensor<int8_t>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<float> output_local = output_queue.DeQue<float>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
-        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
-        const int64_t indices_ne1_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
-            indices_ne[0];
-        const int64_t indices_ne0_idx =
-            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
-             indices_ne1_idx * indices_ne[0]);
-
-        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
-                                       indices_ne1_idx * indices_stride[1] +
-                                       indices_ne2_idx * indices_stride[2];
-        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
-
-        const int64_t input_offset = selected_row_idx * input_stride[1] +
-                                     indices_ne1_idx * input_stride[2] +
-                                     indices_ne2_idx * input_stride[3] +
-                                     group * QK8_0;
-        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
-                                     indices_ne1_idx * scale_stride[2] +
-                                     indices_ne2_idx * scale_stride[3] + group;
-        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
-                                      indices_ne1_idx * output_stride[2] +
-                                      indices_ne2_idx * output_stride[3] +
-                                      group * QK8_0;
-
-        copy_in(input_offset);
-        LocalTensor<int8_t> input_local = input_queue.DeQue<int8_t>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
-
-        // TODO: cast more data to speed up.
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
-        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK8_0);
-
-        // Only mul need compile by group.
-        half scale = scale_gm.GetValue(scale_offset);
-        Muls(output_local, output_local, (float)scale, QK8_0);
-
-        input_queue.FreeTensor(input_local);
-        cast_queue.FreeTensor(cast_local);
-        output_queue.EnQue(output_local);
-
-        copy_out(output_offset);
-    }
-
-    __aicore__ inline void calculate() {
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                calculate_group(i, j);
-            }
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t scale_ne[4];
-    size_t scale_stride[4];
-
-    int64_t indices_ne[4];
-    size_t indices_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t ir;
-    int64_t dr;
-
-    int64_t group_size_in_row;
-
-    TPipe pipe;
-    GlobalTensor<int8_t> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int32_t> indices_gm;
-    GlobalTensor<float> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_get_row_q8_0(
-    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
-    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
-    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
-    int64_t input_ne_ub[4];
-    int64_t indices_ne_ub[4];
-    size_t indices_nb_ub[4];
-    int64_t output_ne_ub[4];
-    size_t output_nb_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
-    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-    copy_to_ub(output_nb_gm, output_nb_ub, 32);
-
-    GET_ROW_Q8_0 op;
-    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
-            indices_nb_ub, output_ne_ub, output_nb_ub);
-    op.calculate();
-}

ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp

@@ -1,218 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P
-    extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f16->8bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define QK8_0 32
-
-class QUANTIZE_F16_Q8_0 {
-   public:
-    __aicore__ inline QUANTIZE_F16_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / QK8_0;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t);
-
-        input_gm.SetGlobalBuffer((__gm__ half *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + ir *
-                                                 group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(half));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(work_queue, 1, 32);
-        pipe.InitBuffer(max_queue, 1, 32);
-        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
-        pipe.InitBuffer(scale_queue, 1, 32);
-        pipe.InitBuffer(cast_queue ,1 ,QK8_0 * sizeof(float));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + QK8_0 * group;
-
-        const int64_t output_offset = i1 * output_stride[1] +
-                                      i2 * output_stride[2] +
-                                      i3 * output_stride[3] + QK8_0 * group;
-
-        copy_in(input_offset);
-        LocalTensor<half> input_local = input_queue.DeQue<half>();
-        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
-
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
-        Abs(abs_local, cast_local, QK8_0);
-        ReduceMax(max_local, abs_local, work_local, QK8_0);
-
-        pipe_barrier(PIPE_ALL);
-        float d = max_local.GetValue(0);
-        d = d / ((1 << 7) - 1);
-        if (d != 0) {
-            Muls(cast_local, cast_local, 1.0f / d, QK8_0);
-        }
-
-        Cast(cast_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(input_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(output_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        abs_queue.FreeTensor(abs_local);
-        max_queue.FreeTensor(max_local);
-        cast_queue.FreeTensor(cast_local);
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                if (scale_local_offset == 16) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += 16;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<half> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, 1> work_queue;
-    TQue<QuePosition::VECOUT, 1> max_queue;
-    TQue<QuePosition::VECIN, 1> abs_queue;
-    TQue<QuePosition::VECOUT, 1> scale_queue;
-    TQue<QuePosition::VECOUT, 1> cast_queue;
-
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_F16_Q8_0 op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp

@@ -1,216 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support f32->8bit quantization
-    extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f32->8bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define QK8_0 32
-
-class QUANTIZE_F32_Q8_0 {
-   public:
-    __aicore__ inline QUANTIZE_F32_Q8_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / QK8_0;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t);
-
-        input_gm.SetGlobalBuffer((__gm__ float *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size +
-                                                 ir * group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(float));
-        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
-        pipe.InitBuffer(work_queue, 1, 32);
-        pipe.InitBuffer(max_queue, 1, 32);
-        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
-        pipe.InitBuffer(cast_queue, 1, QK8_0 * sizeof(half));
-        pipe.InitBuffer(scale_queue, 1, 32);
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        DataCopy(input_local, input_gm[offset], QK8_0);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
-        DataCopy(output_gm[offset], output_local, QK8_0);
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + QK8_0 * group;
-
-        const int64_t output_offset = i1 * output_stride[1] +
-                                      i2 * output_stride[2] +
-                                      i3 * output_stride[3] + QK8_0 * group;
-
-        copy_in(input_offset);
-        LocalTensor<float> input_local = input_queue.DeQue<float>();
-        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
-
-        Abs(abs_local, input_local, QK8_0);
-        ReduceMax(max_local, abs_local, work_local, QK8_0);
-        pipe_barrier(PIPE_ALL);
-        float d = max_local.GetValue(0);
-        d = d / ((1 << 7) - 1);
-        if (d != 0) {
-            Muls(input_local, input_local, 1.0f / d, QK8_0);
-        }
-
-        Cast(input_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(cast_local, input_local, RoundMode::CAST_ROUND, QK8_0);
-        Cast(output_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        abs_queue.FreeTensor(abs_local);
-        max_queue.FreeTensor(max_local);
-        cast_queue.FreeTensor(cast_local);
-
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                if (scale_local_offset == 16) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += 16;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<float> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, 1> work_queue;
-    TQue<QuePosition::VECOUT, 1> max_queue;
-    TQue<QuePosition::VECIN, 1> abs_queue;
-    TQue<QuePosition::VECIN, 1> cast_queue;
-    TQue<QuePosition::VECOUT, 1> scale_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_F32_Q8_0 op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -1,295 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-#ifdef ASCEND_310P // 310P not support float->4bit quantization
-    extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f32->4bit quantization.\n");
-    }
-
-    extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
-        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
-        printf("Ascend310P not support f16->4bit quantization.\n");
-    }
-#else
-
-#define BUFFER_NUM 2
-#define Group_Size 32
-
-template <typename SRC_T>
-class QUANTIZE_FLOAT_TO_Q4_0 {
-   public:
-    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
-        //                         permute=[0,0,0,0]):
-        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        // input stride of data elements
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        // output stride of data elements
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / Group_Size;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t) / 2;
-
-        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
-                                                 group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
-        pipe.InitBuffer(output_queue, BUFFER_NUM,
-                            Group_Size * sizeof(int8_t) / 2);
-        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
-        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
-        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
-        DataCopy(input_local, input_gm[offset], Group_Size);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
-        // and using DataCopyPad to avoid 32 bits align.
-        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
-        LocalTensor<int8_t> output_int8_local =
-                                    output_local.ReinterpretCast<int8_t>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
-        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
-
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<float> input_local) {
-        DataCopy(cast_local, input_local, Group_Size);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<half> input_local) {
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + Group_Size * group;
-
-        // output_offset is stride for output_gm which datatype is int8_t and
-        // divided by 2 is needed for int4b_t.
-        const int64_t output_offset = (i1 * output_stride[1] +
-                                       i2 * output_stride[2] +
-                                       i3 * output_stride[3] +
-                                       Group_Size * group) / 2;
-        copy_in(input_offset);
-
-        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
-        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
-        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
-        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
-        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
-
-        input_to_cast(cast_local, input_local);
-
-        ReduceMax(max_local, cast_local, work_local, Group_Size);
-        ReduceMin(min_local, cast_local, work_local, Group_Size);
-        const float max_value = max_local.GetValue(0);
-        const float min_value = min_local.GetValue(0);
-        float d = max_value;
-        if (min_value < 0 && (-1 * min_value) > max_value) {
-            d = min_value;
-        }
-
-        d = d / (-8);
-        if (d != 0) {
-            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
-        }
-
-        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
-        float scalar = 8.5f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
-        scalar = 15.0f;
-        Mins(cast_local, cast_local, scalar, Group_Size);
-        scalar = -8.0f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-
-        // float->half->int4b
-        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
-        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
-
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        max_queue.FreeTensor(max_local);
-        min_queue.FreeTensor(min_local);
-        int8_queue.FreeTensor(int8_local);
-        half_queue.FreeTensor(half_local);
-        cast_queue.FreeTensor(cast_local);
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                // Copy Group_Size/2 length data each time.
-                if (scale_local_offset == Group_Size / 2) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local,
-                                      Group_Size / 2);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += Group_Size / 2;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-        scale_queue.FreeTensor(scale_local);
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<SRC_T> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<half> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<float> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-#endif // #ifdef ASCEND_310P

ggml/src/ggml-cpu/CMakeLists.txt

@@ -5,7 +5,7 @@ function(ggml_add_cpu_backend_features cpu_name arch)
     # build, using set_source_files_properties() to set the arch flags is not possible
     set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
     add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
-    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
     set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
@@ -70,10 +70,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (GGML_OPENMP)
         find_package(OpenMP)
         if (OpenMP_FOUND)
+            set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
             target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
 
             target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
         else()
+            set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
             message(WARNING "OpenMP not found")
         endif()
     endif()
@@ -448,6 +450,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
+            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -455,6 +458,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             list(APPEND ARCH_FLAGS -march=z16)
         elseif (${S390X_M} MATCHES "9175|9176")
             # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
             message(STATUS "z17 target")
             list(APPEND ARCH_FLAGS -march=z17)
         else()
@@ -464,7 +468,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (GGML_VXE)
+            message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
+            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+        endif()
+
+        if (GGML_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
         endif()
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
@@ -486,9 +497,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.11.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "3fe9e5ab964c375c53839296eb71eaa2")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -581,4 +592,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (EMSCRIPTEN)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
+
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
+    endif()
 endfunction()

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,6 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
@@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
 
         // Quantize these floats
         const float iscale = 127.f / amax;
-        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
+        y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
         const float id = ( amax != 0.0f ) ? iscale : 0.f;
         const __m512 vscale = _mm512_set1_ps(id);
 
@@ -1090,7 +1091,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1113,8 +1114,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
         const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
-            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1137,7 +1138,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b
@@ -1498,8 +1499,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
-                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
             }
 
             // load b
@@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     va[k] = _mm512_add_epi8(va[k], off);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -62,7 +63,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const float32x4_t v  = vmulq_n_f32(srcv[j], id);
@@ -104,7 +105,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         int32x4_t accv = vdupq_n_s32(0);
 
@@ -120,7 +121,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             accv = vaddq_s32(accv, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vaddvq_s32(accv));
     }
 #else
     GGML_UNUSED(nb);
@@ -194,10 +195,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -274,10 +275,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     // dot product
                     sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                     svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
-                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                     svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
-                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -313,9 +314,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     // dot product
                     sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -354,9 +355,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     // dot product
                     sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
@@ -404,8 +405,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
         const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -423,7 +424,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -464,10 +465,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i];
 
             float32_t summs_t[4] = {
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s)
             };
             summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
 
@@ -490,10 +491,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             // mmla into int32x4_t
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -539,7 +540,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
         const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -562,8 +563,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
         const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@@ -582,7 +583,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -666,10 +667,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -694,7 +695,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -739,8 +740,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
-        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
 
         // extract the 5th bit via lookup table ((b) << 4)
         memcpy(&qh0, x0->qh, sizeof(qh0));
@@ -784,10 +785,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@@ -812,7 +813,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -864,10 +865,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -934,10 +935,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                     svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
-                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                     svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
-                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
@@ -960,9 +961,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
 
                     sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -1002,8 +1003,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
 
                     // scale creation
-                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
-                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+                    const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d);
 
                     // duplicate deq1 in first half of vector and deq2 in second half of vector
                     const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
@@ -1043,11 +1044,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
 
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -1059,7 +1060,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1217,7 +1218,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
         const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
 #if defined(__ARM_FEATURE_DOTPROD)
         sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -1362,7 +1363,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
         const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
 #if defined(__ARM_FEATURE_DOTPROD)
         sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1393,7 +1394,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -1425,9 +1426,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     switch (vector_length) {
         case 128:
             for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                 svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                 svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
 
                 const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1570,9 +1571,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         case 256:
         case 512:
             for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                 svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                 svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
 
                 const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1671,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sum = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1742,8 +1743,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1805,7 +1806,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3_sv = x[i].qs;
         const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask;
@@ -1981,7 +1982,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].hmask;
@@ -2112,7 +2113,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2258,18 +2259,18 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
                                                vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
                 const float32x4_t dmins = {
-                    GGML_FP16_TO_FP32(x0->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x0->dmin) * y1->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d,
                 };
                 vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
 
                 const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                 };
                 vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
             }
@@ -2289,8 +2290,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2377,8 +2378,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2478,9 +2479,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2520,8 +2521,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2630,9 +2631,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2827,10 +2828,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
 
                 const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                 };
 
                 visum = vsubq_s32(visum, vibias);
@@ -2858,7 +2859,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     svuint8_t q6h_1, q6h_2, q6h_3, q6h_4;
 
     for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q6 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3011,7 +3012,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q6 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3128,7 +3129,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -3199,7 +3200,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         float sumf1 = 0, sumf2 = 0;
@@ -3234,7 +3235,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -3284,7 +3285,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         const uint8x8_t scales8 = vld1_u8(x[i].scales);
@@ -3329,7 +3330,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3398,7 +3399,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3458,7 +3459,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -3521,7 +3522,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3557,7 +3558,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3630,7 +3631,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3691,7 +3692,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3786,7 +3787,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
         }
 
-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
     }
 
     *s = sumf;
@@ -3817,7 +3818,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -3905,7 +3906,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
         }
 
-        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
     }
 
     *s = sumf;
@@ -3952,7 +3953,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qh += 2;
         }
 
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -4003,13 +4004,13 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
 
         sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+            GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
     }
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4071,7 +4072,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         }
 
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
     }
 
     *s = sumf;
@@ -4079,7 +4080,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include <cmath>
@@ -51,7 +52,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < 8; j++) {
@@ -102,7 +103,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -145,7 +146,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < 4; j++) {
@@ -221,7 +222,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -311,7 +312,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -399,7 +400,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -514,7 +515,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -608,7 +609,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
                             const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -1117,7 +1118,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                                 }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }
@@ -1570,7 +1571,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -2039,7 +2040,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -2147,7 +2148,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                                 }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -474,7 +475,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
 
@@ -543,12 +544,12 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
         const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = __lasx_xvreplfr2vr_s( id );
 
@@ -576,7 +577,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
         const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
 
         // Convert int32 to int16
         ni0 = lsx_packs_w( ni0, ni1 );
@@ -667,7 +668,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
@@ -699,7 +700,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -717,7 +718,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 
@@ -766,7 +767,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -797,10 +798,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
         const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
@@ -834,7 +835,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -865,7 +866,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -902,7 +903,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -934,16 +935,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
         qx = __lasx_xvor_v(qx, bxhi);
 
-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -973,7 +974,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -1003,7 +1004,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
         __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
         __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
@@ -1023,7 +1024,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1047,8 +1048,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1116,8 +1117,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1170,7 +1171,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         // Set up scales
@@ -1294,7 +1295,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1330,8 +1331,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1438,9 +1439,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1593,9 +1594,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1624,7 +1625,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1713,7 +1714,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1780,7 +1781,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = __lasx_xvldi(0);
@@ -1820,7 +1821,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -1895,7 +1896,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -1980,7 +1981,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,7 +2050,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -2108,7 +2109,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -2168,7 +2169,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2213,7 +2214,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2279,7 +2280,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -2340,7 +2341,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2451,7 +2452,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
         accum1 += d * sumi1;
     }
@@ -2484,7 +2485,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -2530,9 +2531,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = lasx_madd_h(p16_1, mone);
         const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                 __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                 __lasx_xvffint_s_w(p_2), accum2);
     }
 
@@ -2540,7 +2541,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2595,7 +2596,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi1 = __lasx_xvadd_w(p_1, sumi1);
             sumi2 = __lasx_xvadd_w(p_2, sumi2);
         }
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
     }
 
@@ -2604,7 +2605,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/powerpc/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -67,7 +68,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float id = d ? 1.0f/d : 0.0f;
         const vector float vid = vec_splats(id);
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const vector float v  = vec_round(vec_mul(srcv[j], vid));
@@ -112,7 +113,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float id = d ? 1.0f/d : 0.0f;
         const vector float vid = vec_splats(id);
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vector int accv = vec_splats(0);
 
@@ -127,7 +128,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         accv = vec_add(accv, vec_sld(accv, accv, 4));
         accv = vec_add(accv, vec_sld(accv, accv, 8));
-        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0));
     }
 
 #else
@@ -170,8 +171,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -214,7 +215,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -249,12 +250,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -291,7 +292,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -326,8 +327,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
@@ -379,7 +380,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -415,12 +416,12 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
         vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
@@ -470,7 +471,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -502,8 +503,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char q8x0 = vec_xl( 0, x[ib].qs);
@@ -542,7 +543,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -574,11 +575,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -708,8 +709,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -770,7 +771,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -962,7 +963,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1005,11 +1006,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -1177,9 +1178,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1222,11 +1223,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         UNUSED(kmask1);
@@ -1394,9 +1395,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1432,7 +1433,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1591,7 +1592,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1659,7 +1660,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1742,7 +1743,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -1790,7 +1791,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1871,7 +1872,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -1939,7 +1940,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2033,7 +2034,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -2096,7 +2097,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2176,7 +2177,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2236,7 +2237,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2329,7 +2330,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2394,7 +2395,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2505,7 +2506,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -2546,8 +2547,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         __builtin_prefetch(y[ib].qs, 0, 1);
 
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -2582,7 +2583,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2620,7 +2621,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     for (int ibl = 0; ibl < nb; ++ibl) {
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d));
         vector float vyd = vec_splats(y[ibl].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2697,7 +2698,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/riscv/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -45,7 +46,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
 
@@ -85,7 +86,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d  = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
 
@@ -102,7 +103,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // set y[i].s
         int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = GGML_FP32_TO_FP16(sum*d);
+        y[i].s = GGML_CPU_FP32_TO_FP16(sum*d);
     }
 
 #else
@@ -160,7 +161,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
 #endif
@@ -177,7 +178,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -225,7 +226,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
 #endif
@@ -242,7 +243,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -293,7 +294,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
         int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
 #endif
@@ -316,7 +317,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -366,7 +367,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
         int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
 #endif
@@ -389,7 +390,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -427,7 +428,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
 #endif
@@ -438,7 +439,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -465,8 +466,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * q2 = x[i].qs;
         const  int8_t * q8 = y[i].qs;
         const uint8_t * sc = x[i].scales;
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
         uint8_t *patmp = atmp;
         int vsums;
         int tmp;
@@ -569,8 +570,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8_t *  q8 = y[i].qs;
             const uint8_t * sc = x[i].scales;
 
-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             size_t vl = 16;
 
@@ -644,8 +645,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const uint8_t * q2 = x[i].qs;
             const  int8_t * q8 = y[i].qs;
             const uint8_t * sc = x[i].scales;
-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
             uint8_t *patmp = atmp;
             int vsums;
             int tmp;
@@ -750,8 +751,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -916,7 +917,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             q3 += 32;    q8 += 128;   scale += 8;
         }
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         sumf += d * isum;
     }
 
@@ -1017,7 +1018,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             }
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             sumf += d*sum_t;
 
@@ -1134,7 +1135,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 q3 += 32;    q8 += 128;   scale += 8;
             }
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
             sumf += d * isum;
         }
         break;
@@ -1202,7 +1203,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1239,8 +1240,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int tmp, tmp2, sumi;
         __asm__ __volatile__(
@@ -1361,8 +1362,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             size_t vl = 8;
 
-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
             vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
@@ -1422,8 +1423,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         break;
     case 128:
         for (int i = 0; i < nb; ++i) {
-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             int tmp, tmp2, sumi;
             __asm__ __volatile__(
@@ -1580,9 +1581,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1627,8 +1628,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT hm = x[i].qh;
         const  int8_t * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
 
         vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl);
         vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl);
@@ -1749,9 +1750,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1778,7 +1779,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -1862,7 +1863,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 256:
         for (int i = 0; i < nb; ++i) {
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * GGML_RESTRICT q6 = x[i].ql;
             const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1943,7 +1944,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 128:
         for (int i = 0; i < nb; ++i) {
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * restrict q6 = x[i].ql;
             const uint8_t * restrict qh = x[i].qh;
@@ -2058,7 +2059,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include <cmath>
@@ -90,16 +91,16 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
 
                 // vector version needs Zvfhmin extension
-                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 const float b_scales[8] = {
-                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                 };
                 const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
                 const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
@@ -129,7 +130,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -181,20 +182,20 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
                     // vector version needs Zvfhmin extension
                     const float a_scales[4] = {
-                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
                     };
                     const float b_scales[8] = {
-                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                     };
                     const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
 
@@ -382,7 +383,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -49,7 +50,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f / d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const __vector float v = vec_mul(srcv[j], vec_splats(id));
@@ -94,7 +95,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f / d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         __vector int32_t acc = vec_splats(0);
 
@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             acc = vec_add(acc, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
     }
 #else
     GGML_UNUSED(nb);
@@ -164,7 +165,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
 
         const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
-        const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -185,7 +186,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -219,7 +220,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const uint8x16_t v_x = vec_xl(0, x[ib].qs);
         const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
@@ -231,7 +232,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
         const float32x4_t v_xy = vec_float(v_xy_);
 
-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -252,7 +253,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
         const float32x4_t v_xy = vec_float(v_xy_);
-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -305,7 +306,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -348,7 +349,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sum = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * restrict x0l = x[i].qs;
         const uint8_t * restrict x0h = x[i].hmask;
@@ -497,7 +498,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -537,8 +538,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
         const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -647,9 +648,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -698,8 +699,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
         const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -819,9 +820,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -859,7 +860,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int8x16_t  v_y[4];
 
     for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT x0l = x[i].ql;
         const uint8_t * GGML_RESTRICT x0h = x[i].qh;
@@ -1004,7 +1005,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1071,7 +1072,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 //    float sumf = 0;
 
 //    for (int i = 0; i < nb; ++i) {
-//        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -1121,7 +1122,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 //     float sumf = 0.f;
 //     for (int i = 0; i < nb; ++i) {
-//         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//         const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 //         int32_t bsum = 0;
@@ -1182,12 +1183,12 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
         const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
 
-        sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
     }
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1257,7 +1258,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
     }
 
     *s = sumf;
@@ -1265,7 +1266,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/wasm/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -65,7 +66,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         v128_t accv = wasm_i32x4_splat(0);
 
@@ -126,7 +127,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             accv = wasm_i32x4_add(accv, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(
+        y[i].s = GGML_CPU_FP32_TO_FP16(
                 d * (wasm_i32x4_extract_lane(accv, 0) +
                      wasm_i32x4_extract_lane(accv, 1) +
                      wasm_i32x4_extract_lane(accv, 2) +
@@ -324,8 +325,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         );
 
         // Accumulate results with scaling
-        float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
-        float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d);
+        float scale0 = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
+        float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d);
 
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
@@ -348,7 +349,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -428,7 +429,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                             wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                            wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
     }
 
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -454,7 +455,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -491,7 +492,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
         const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
 
         const v128_t m4b = wasm_i8x16_splat(0x0F);
 
@@ -538,7 +539,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                             wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                            wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
     }
 
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -564,7 +565,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -620,7 +621,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));
 
         // Convert to float and accumulate
-        const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
+        const float scale = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
     }
 
@@ -635,7 +636,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -746,8 +747,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             isum += wasm_i32x4_extract_lane(isum_vec, 0);
         }
 
-        const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf += dall * isum - dmin * summs;
     }
 
@@ -768,8 +769,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -880,7 +881,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         // Accumulate results
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const v128_t v_d = wasm_f32x4_splat(d);
         v128_t v_sum = wasm_f32x4_add(
             wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
@@ -957,7 +958,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -991,8 +992,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Corrected sign
 
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1136,9 +1137,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1170,8 +1171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Fixed sign
 
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1331,9 +1332,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1420,7 +1421,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         wasm_v128_store(&aux32[0], acc0);
         wasm_v128_store(&aux32[4], acc1);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) {
             sums[l] += d * aux32[l];
         }
@@ -1470,7 +1471,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -256,9 +257,9 @@ static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_
 
 // quad fp16 delta calculation
 static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
-    // GGML_FP16_TO_FP32 is faster than Intel F16C
-    return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0)));
+    // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -305,7 +306,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -401,7 +402,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -425,7 +426,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__AVX2__)
         // Compute the sum of the quants and set y[i].s
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
 
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -455,7 +456,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
         const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
 
         // Convert int32 to int16
         ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -552,7 +553,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
@@ -613,7 +614,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
 
@@ -631,7 +632,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
 
@@ -680,7 +681,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -711,10 +712,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = _mm256_set1_ps( d0 );
         const __m256 d1v = _mm256_set1_ps( d1 );
@@ -752,7 +753,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -783,7 +784,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -807,7 +808,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
         const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -851,7 +852,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -883,16 +884,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
         qx = _mm256_or_si256(qx, bxhi);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -910,9 +911,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
         const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -926,7 +927,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         bxh = _mm_or_si128(bxh, bxhih);
         bx_0 = MM256_SET_M128I(bxh, bxl);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@@ -956,7 +957,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -986,7 +987,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
         __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
         __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
@@ -1025,7 +1026,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1144,7 +1145,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         }
 
         const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
 
         sumi0 = _mm256_sub_epi16(sumi0, ysum);
         sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
@@ -1190,7 +1191,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -1244,7 +1245,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         }
 
         const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
 
         sumi0 = _mm256_add_epi16(sumi0, sumi1);
         sumi0 = _mm256_sub_epi16(sumi0, ysum);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -1299,8 +1300,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1366,8 +1367,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1533,7 +1534,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1638,7 +1639,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1824,7 +1825,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1862,8 +1863,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1928,8 +1929,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,9 +2050,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2092,8 +2093,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -2170,8 +2171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2311,9 +2312,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2344,7 +2345,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2422,7 +2423,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2555,7 +2556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2622,7 +2623,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = _mm256_setzero_si256();
@@ -2663,7 +2664,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m128i sumi1_0 = _mm_setzero_si128();
@@ -2717,7 +2718,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -2792,7 +2793,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -2913,7 +2914,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -3035,7 +3036,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3104,7 +3105,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3177,7 +3178,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3253,7 +3254,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -3313,7 +3314,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3358,7 +3359,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3414,7 +3415,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3480,7 +3481,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3565,7 +3566,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3648,7 +3649,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3753,7 +3754,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
         accum1 += d * sumi1;
 
@@ -3801,7 +3802,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
         accum1 += d * sumi1;
 
@@ -3835,7 +3836,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -3947,7 +3948,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 8; qh += 4;
         }
 
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
 
         accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
         accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
@@ -4033,7 +4034,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 8; qh += 4;
         }
 
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
 
         accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
         accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
@@ -4083,7 +4084,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qh += 2;
         }
 
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -4129,9 +4130,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
         const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                 _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                 _mm256_cvtepi32_ps(p_2), accum2);
     }
 
@@ -4164,7 +4165,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4219,7 +4220,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi1 = _mm256_add_epi32(p_1, sumi1);
             sumi2 = _mm256_add_epi32(p_2, sumi2);
         }
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
     }
 
@@ -4267,7 +4268,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
         }
         __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
         __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
     }
 
@@ -4276,7 +4277,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include <cmath>
@@ -39,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
     float tmp[16];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     for (int i = 0; i < 8; i++) {
-        tmp[i + 8] = GGML_FP16_TO_FP32(y[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -54,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
     _mm_storeu_si128((__m128i*)tmphalf, x);
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -67,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -76,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -88,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang
 
     _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -211,7 +212,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
 
             // Store the scale for the individual block
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
 
             // Store the values in blocks of eight values - Aim is to use these later for block interleaving
             srcv[row_iter][0] = v0;
@@ -297,7 +298,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -647,7 +648,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
 
                 // Load and convert to FP32 scale from block_q8_0
-                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
 
                 // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
                 __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
@@ -706,7 +707,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -972,13 +973,13 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                 }
             }
             for (int sb = 0; sb < 8; sb++) {
                 uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -1755,7 +1756,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -3259,7 +3260,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi2 = sumi2 * scales_1[j];
                                 sumi += sumi1 + sumi2;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -3268,7 +3269,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     for(int m = 0; m < 4; m++) {
                         const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                         for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                         }
                     }
                 }

ggml/src/ggml-cpu/common.h

@@ -4,6 +4,7 @@
 #include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
+#include "simd-mappings.h"
 
 #ifdef __cplusplus
 
@@ -12,11 +13,11 @@
 // convenience functions/macros for use in template calls
 // note: these won't be required after the 'traits' lookup table is used.
 static inline ggml_fp16_t f32_to_f16(float x) {
-    return GGML_FP32_TO_FP16(x);
+    return GGML_CPU_FP32_TO_FP16(x);
 }
 
 static inline float f16_to_f32(ggml_fp16_t x) {
-    return GGML_FP16_TO_FP32(x);
+    return GGML_CPU_FP16_TO_FP32(x);
 }
 
 static inline ggml_bf16_t f32_to_bf16(float x) {