make : fix samples glob pattern

* whisper: suppress Windows compiler warnings

This commit disables compiler warnings on window using MSVC.

The motivation for these changes is that some compilers generate
warnings for these conversion, for example Windows MSVC, and
there are quite a few of them. This makes it a little difficult to
spot new warnings that may be introduced and also can be difficult
for users/embedders of ggml where these warnings are hard to separate
from their own warnings.

* squash! whisper: suppress Windows compiler warnings

Move ggml related warnings into ggml. This commit also fixes the
indentation and adds a missing whitespace to the if statement.

.devops/main-cuda.Dockerfile

@@ -13,8 +13,6 @@ WORKDIR /app
 ARG CUDA_DOCKER_ARCH=all
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV GGML_CUDA=1
 
 RUN apt-get update && \
     apt-get install -y build-essential libsdl2-dev wget cmake git \
@@ -25,7 +23,8 @@ ENV CUDA_MAIN_VERSION=12.3
 ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 
 COPY .. .
-RUN make base.en
+# Enable cuBLAS
+RUN make base.en CMAKE_ARGS="-DGGML_CUDA=1"
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 ENV CUDA_MAIN_VERSION=12.3
@@ -37,4 +36,5 @@ RUN apt-get update && \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app
+ENV PATH=/app/build/bin:$PATH
 ENTRYPOINT [ "bash", "-c" ]

.devops/main-musa.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.1
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+WORKDIR /app
+
+RUN apt-get update && \
+    apt-get install -y build-essential libsdl2-dev wget cmake git \
+    && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
+
+COPY .. .
+# Enable muBLAS
+RUN make base.en CMAKE_ARGS="-DGGML_MUSA=1"
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+WORKDIR /app
+
+RUN apt-get update && \
+  apt-get install -y curl ffmpeg wget cmake git \
+  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
+
+COPY --from=build /app /app
+ENV PATH=/app/build/bin:$PATH
+ENTRYPOINT [ "bash", "-c" ]

.devops/main.Dockerfile

@@ -16,4 +16,5 @@ RUN apt-get update && \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app
+ENV PATH=/app/build/bin:$PATH
 ENTRYPOINT [ "bash", "-c" ]

.dockerignore

@@ -0,0 +1,3 @@
+build*/
+.github/
+.devops/

.github/workflows/bindings-ruby.yml

@@ -1,55 +1,11 @@
 name: Bindings Tests (Ruby)
+
 on:
   push:
-    paths:
-      - bindings/ruby/**
-      - src/**/*.c
-      - src/**/*.cpp
-      - src/**/*.h
-      - src/**/*.m
-      - src/**/*.metal
-      - include/**/*.c
-      - include/**/*.cpp
-      - include/**/*.h
-      - include/**/*.m
-      - include/**/*.metal
-      - ggml/**/*.c
-      - ggml/**/*.cpp
-      - ggml/**/*.h
-      - ggml/**/*.m
-      - ggml/**/*.metal
-      - scripts/get-flags.mk
-      - examples/common.h
-      - examples/common.cpp
-      - examples/common-whisper.h
-      - examples/common-whisper.cpp
-      - examples/stb_vorbis.c
-      - examples/miniaudio.h
+    branches:
+      - master
   pull_request:
-    paths:
-      - bindings/ruby/**
-      - src/**/*.c
-      - src/**/*.cpp
-      - src/**/*.h
-      - src/**/*.m
-      - src/**/*.metal
-      - include/**/*.c
-      - include/**/*.cpp
-      - include/**/*.h
-      - include/**/*.m
-      - include/**/*.metal
-      - ggml/**/*.c
-      - ggml/**/*.cpp
-      - ggml/**/*.h
-      - ggml/**/*.m
-      - ggml/**/*.metal
-      - scripts/get-flags.mk
-      - examples/common.h
-      - examples/common.cpp
-      - examples/common-whisper.h
-      - examples/common-whisper.cpp
-      - examples/stb_vorbis.c
-      - examples/miniaudio.h
+    types: [opened, synchronize, reopened]
 
 jobs:
   ubuntu-22:
@@ -60,6 +16,6 @@ jobs:
     steps:
       - uses: ruby/setup-ruby@v1
         with:
-          ruby-version: '3.1'
+          ruby-version: '3.2'
       - uses: actions/checkout@v4
       - run: rake test

.github/workflows/build.yml

@@ -200,23 +200,23 @@ jobs:
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
 
-  freeBSD-latest:
-    runs-on: macos-13
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Build
-        uses: cross-platform-actions/action@v0.27.0
-        with:
-          operating_system: freebsd
-          version: '14.2'
-          run: |
-            sudo pkg update
-            sudo pkg install -y gmake sdl2 cmake git
-            cmake -B build
-            cmake --build build --config Release
+#  freeBSD-latest:
+#    runs-on: macos-13
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Build
+#        uses: cross-platform-actions/action@v0.27.0
+#        with:
+#          operating_system: freebsd
+#          version: '14.2'
+#          run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake sdl2 cmake git
+#            cmake -B build
+#            cmake --build build --config Release
 
   ubuntu-22-gcc:
     if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
@@ -561,6 +561,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          -DBUILD_SHARED_LIBS=ON
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
@@ -572,12 +573,37 @@ jobs:
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
-      - name: Upload dll
+      - name: Upload SDL2.dll
+        if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.jnaPath }}_whisper.dll
+          name: ${{ matrix.s2arc }}_SDL2.dll
+          path: build/bin/${{ matrix.build }}/SDL2.dll
+
+      - name: Upload whisper dll
+        uses: actions/upload-artifact@v4
+        with:
+          name: whisper_${{ matrix.arch }}.dll
           path: build/bin/${{ matrix.build }}/whisper.dll
 
+      - name: Upload ggml dll
+        uses: actions/upload-artifact@v4
+        with:
+          name: ggml_${{ matrix.arch }}.dll
+          path: build/bin/${{ matrix.build }}/ggml.dll
+
+      - name: Upload ggml base dll
+        uses: actions/upload-artifact@v4
+        with:
+          name: ggml_base_${{ matrix.arch }}.dll
+          path: build/bin/${{ matrix.build }}/ggml-base.dll
+
+      - name: Upload ggml cpu dll
+        uses: actions/upload-artifact@v4
+        with:
+          name: ggml_cpu_${{ matrix.arch }}.dll
+          path: build/bin/${{ matrix.build }}/ggml-cpu.dll
+
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v4
@@ -996,38 +1022,88 @@ jobs:
           chmod +x ./gradlew
           ./gradlew assembleRelease
 
-# TODO: disabled because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/9686220096/job/26735899598
-#  java:
-#    needs: [ 'windows' ]
-#    runs-on: windows-latest
-#    steps:
-#      - uses: actions/checkout@v4
-#
-#      - name: Install Java
-#        uses: actions/setup-java@v4
-#        with:
-#          distribution: zulu
-#          java-version: 20
-#
-#      - name: Download Windows lib
-#        uses: actions/download-artifact@v4
-#        with:
-#          name: win32-x86-64_whisper.dll
-#          path: bindings/java/build/generated/resources/main/win32-x86-64
-#
-#      - name: Build
-#        run: |
-#          models\download-ggml-model.cmd tiny.en
-#          cd bindings/java
-#          chmod +x ./gradlew
-#          ./gradlew build
-#
-#      - name: Upload jar
-#        uses: actions/upload-artifact@v4
-#        with:
-#          name: whispercpp.jar
-#          path: bindings/java/build/libs/whispercpp-*.jar
-#
+  bindings-java:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    needs: ['windows']
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Java
+        uses: actions/setup-java@v4
+        with:
+          distribution: zulu
+          java-version: 20
+
+      - name: Download Whisper Windows lib
+        uses: actions/download-artifact@v4
+        with:
+          name: whisper_x64.dll
+
+      - name: Download GGML Windows lib
+        uses: actions/download-artifact@v4
+        with:
+          name: ggml_x64.dll
+
+      - name: Download GGML Base Windows lib
+        uses: actions/download-artifact@v4
+        with:
+          name: ggml_base_x64.dll
+
+      - name: Download GGML CPU Windows lib
+        uses: actions/download-artifact@v4
+        with:
+          name: ggml_cpu_x64.dll
+
+      - name: Download SDL2.dll
+        uses: actions/download-artifact@v4
+        with:
+          name: x64_SDL2.dll
+
+      - name: List downloaded files
+        shell: pwsh
+        run: |
+          Get-ChildItem -Path "." -Recurse -Filter "*.dll"
+
+      - name: Move DLL to correct location
+        shell: pwsh
+        run: |
+          New-Item -Path "build\bin\Release" -ItemType Directory -Force
+
+          Copy-Item -Path "whisper.dll" -Destination "build\bin\Release\whisper.dll" -Force
+          Write-Host "Copied whisper.dll to build\bin\Release\whisper.dll directory"
+
+          Copy-Item -Path "ggml.dll" -Destination "build\bin\Release\ggml.dll" -Force
+          Write-Host "Copied ggml.dll to build\bin\Release\ggml.dll directory"
+
+          Copy-Item -Path "ggml-base.dll" -Destination "build\bin\Release\ggml-base.dll" -Force
+          Write-Host "Copied ggml-base.dll to build\bin\Release\ggml-base.dll directory"
+
+          Copy-Item -Path "ggml-cpu.dll" -Destination "build\bin\Release\ggml-cpu.dll" -Force
+          Write-Host "Copied ggml-cpu.dll to build\bin\Release\ggml-cpu.dll directory"
+
+          Copy-Item -Path "SDL2.dll" -Destination "build\bin\Release\SDL2.dll" -Force
+          Write-Host "Copied SDL2.dll to build\bin\Release\SDL2.dll directory"
+
+      - name: List build release files
+        shell: pwsh
+        run: |
+          Get-ChildItem -Path "build\Release" -Recurse -Filter "*.dll"
+
+      - name: Build
+        run: |
+          models\download-ggml-model.cmd tiny.en models/
+          cd bindings/java
+          chmod +x ./gradlew
+          ./gradlew build --info
+
+      - name: Upload jar
+        uses: actions/upload-artifact@v4
+        with:
+          name: whispercpp.jar
+          path: bindings/java/build/libs/whispercpp-*.jar
+
 #      - name: Publish package
 #        if: ${{ github.ref == 'refs/heads/master' }}
 #        uses: gradle/gradle-build-action@v2.4.2

.github/workflows/docker.yml

@@ -18,6 +18,7 @@ jobs:
       matrix:
         config:
           - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64" }
+          - { tag: "main-musa", dockerfile: ".devops/main-musa.Dockerfile", platform: "linux/amd64" }
           #TODO: the cuda image keeps failing - disable for now
           #      https://github.com/ggerganov/whisper.cpp/actions/runs/11019444428/job/30602020339
           #- { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }

CMakeLists.txt

@@ -135,6 +135,22 @@ if (NOT TARGET ggml)
         add_library(ggml ALIAS ggml::ggml)
     else()
         add_subdirectory(ggml)
+        if(WIN32)
+            # The following adds a _DISABLE_CONSTEXPR_MUTEX_CONSTRUCTOR macro and is a workaround for
+            # the Windows C++ standard library which does not support constexpr mutexes.
+            # From the release notes://github.com/microsoft/STL/wiki/Changelog
+            #  Disable constexpr mutex constructor on Windows
+            #  Fixed mutex's constructor to be constexpr. #3824 #4000 #4339
+            #  Note: Programs that aren't following the documented restrictions on binary compatibility may encounter
+            #  null dereferences in mutex machinery. You must follow this rule:
+            #  When you mix binaries built by different supported versions of the toolset, the Redistributable version
+            #  must be at least as new as the latest toolset used by any app component.
+            #  You can define _DISABLE_CONSTEXPR_MUTEX_CONSTRUCTOR as an escape hatch.
+            #
+            # Specifically to whisper.cpp this would cause a crash when using the Java bindings.
+            # resulting in a Invalid memory access error.
+            target_compile_definitions(ggml-base PRIVATE _DISABLE_CONSTEXPR_MUTEX_CONSTRUCTOR)
+        endif()
     endif()
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -197,3 +213,33 @@ endif ()
 if (WHISPER_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4101  # Unreferenced local variable
+        /wd4005  # Macro redefinition
+        /wd4065  # switch statement contains 'default' but no 'case' labels
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+        /wd4244  # Conversion from one type to another type, possible loss of ata
+        /wd4805  # Unsafe mix of type
+        /wd4305  # Truncation from 'type1' to 'type2' (often double to float)
+        /wd4996  # Function or variable may be unsafe/deprecated
+    )
+    function(disable_msvc_warnings target_name)
+        target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+    endfunction()
+
+    if (WHISPER_BUILD_EXAMPLES)
+        disable_msvc_warnings(common)
+        disable_msvc_warnings(common-sdl)
+        disable_msvc_warnings(lsp)
+        disable_msvc_warnings(wchess-core)
+        disable_msvc_warnings(whisper-command)
+        disable_msvc_warnings(whisper-cli)
+        disable_msvc_warnings(whisper-server)
+        disable_msvc_warnings(whisper-stream)
+        disable_msvc_warnings(whisper-talk-llama)
+        disable_msvc_warnings(whisper-bench)
+        disable_msvc_warnings(quantize)
+    endif()
+endif()

Makefile

@@ -4,7 +4,7 @@
 
 .PHONY: build
 build:
-	cmake -B build
+	cmake -B build $(CMAKE_ARGS)
 	cmake --build build --config Release
 
 # download a few audio samples into folder "./samples":
@@ -41,17 +41,17 @@ samples:
 
 tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo:
 	bash ./models/download-ggml-model.sh $@
-	cmake -B build
+	cmake -B build $(CMAKE_ARGS)
 	cmake --build build --config Release
 	@echo ""
 	@echo "==============================================="
 	@echo "Running $@ on all samples in ./samples ..."
 	@echo "==============================================="
 	@echo ""
-	@for f in samples/*$(.flac .mp3 .ogg .wav); do \
+	@for f in samples/*.{flac,mp3,ogg,wav}; do \
 		echo "----------------------------------------------" ; \
 		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
-	    echo "----------------------------------------------" ; \
+		echo "----------------------------------------------" ; \
 		echo "" ; \
 		./build/bin/whisper-cli -m models/ggml-$@.bin -f $$f ; \
 		echo "" ; \

README.md

@@ -23,6 +23,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - [Efficient GPU support for NVIDIA](#nvidia-gpu-support)
 - [OpenVINO Support](#openvino-support)
 - [Ascend NPU Support](#ascend-npu-support)
+- [Moore Threads GPU Support](#moore-threads-gpu-support)
 - [C-style API](https://github.com/ggml-org/whisper.cpp/blob/master/include/whisper.h)
 
 Supported platforms:
@@ -322,6 +323,12 @@ cmake -B build -DGGML_CUDA=1
 cmake --build build -j --config Release
 ```
 
+or for newer NVIDIA GPU's (RTX 5000 series):
+```
+cmake -B build -DGGML_CUDA=1 -DCMAKE_CUDA_ARCHITECTURES="86"
+cmake --build build -j --config Release
+```
+
 ## Vulkan GPU support
 Cross-vendor solution which allows you to accelerate workload on your GPU.
 First, make sure your graphics card driver provides support for Vulkan API.
@@ -375,6 +382,56 @@ Run the inference examples as usual, for example:
 - If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
 - If you run successfully with your Ascend NPU device, please help update the table `Verified devices`.
 
+## 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 rc3.1.1`: https://developer.mthreads.com/sdk/download/musa?equipment=&os=&driverVersion=&version=rc3.1.1
+
+Now build `whisper.cpp` with MUSA support:
+
+```
+cmake -B build -DGGML_MUSA=1
+cmake --build build -j --config Release
+```
+
+or specify the architecture for your Moore Threads GPU. For example, if you have a MTT S80 GPU, you can specify the architecture as follows:
+
+```
+cmake -B build -DGGML_MUSA=1 -DMUSA_ARCHITECTURES="21"
+cmake --build build -j --config Release
+```
+
+## FFmpeg support (Linux only)
+
+If you want to support more audio formats (such as Opus and AAC), you can turn on the `WHISPER_FFMPEG` build flag to enable FFmpeg integration.
+
+First, you need to install required libraries:
+
+```bash
+# Debian/Ubuntu
+sudo apt install libavcodec-dev libavformat-dev libavutil-dev
+
+# RHEL/Fedora
+sudo dnf install libavcodec-free-devel libavformat-free-devel libavutil-free-devel
+```
+
+Then you can build the project as follows:
+
+```bash
+cmake -B build -D WHISPER_FFMPEG=yes
+cmake --build build
+```
+
+Run the following example to confirm it's working:
+
+```bash
+# Convert an audio file to Opus format
+ffmpeg -i samples/jfk.wav jfk.opus
+
+# Transcribe the audio file
+./build/bin/whisper-cli --model models/ggml-base.en.bin --file jfk.opus
+```
+
 ## Docker
 
 ### Prerequisites
@@ -388,6 +445,7 @@ We have two Docker images available for this project:
 
 1. `ghcr.io/ggml-org/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggml-org/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
+3. `ghcr.io/ggml-org/whisper.cpp:main-musa`: Same as `main` but compiled with MUSA support. (platforms: `linux/amd64`)
 
 ### Usage
 
@@ -400,11 +458,11 @@ docker run -it --rm \
 docker run -it --rm \
   -v path/to/models:/models \
   -v path/to/audios:/audios \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
+  whisper.cpp:main "whisper-cli -m /models/ggml-base.bin -f /audios/jfk.wav"
 # transcribe an audio file in samples folder
 docker run -it --rm \
   -v path/to/models:/models \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
+  whisper.cpp:main "whisper-cli -m /models/ggml-base.bin -f ./samples/jfk.wav"
 ```
 
 ## Installing with Conan

bindings/java/build.gradle

@@ -27,23 +27,41 @@ sourceSets {
 tasks.register('copyLibwhisperDynlib', Copy) {
     from '../../build/src'
     include 'libwhisper.dylib'
-    into 'build/generated/resources/main/darwin'
+    into 'build/generated/resources/main'
 }
 
 tasks.register('copyLibwhisperSo', Copy) {
     from '../../build/src'
     include 'libwhisper.so'
-    into 'build/generated/resources/main/linux-x86-64'
+    into 'build/generated/resources/main'
 }
 
-tasks.register('copyWhisperDll', Copy) {
-    from '../../build/Release'
+tasks.register('copyWhisperDLL', Copy) {
+    from '../../build/bin/Release'
     include 'whisper.dll'
-    into 'build/generated/resources/main/windows-x86-64'
+    into 'build/generated/resources/main'
+}
+
+tasks.register('copyGGML_BASE_DLL', Copy) {
+    from '../../build/bin/Release'
+    include 'ggml-base.dll'
+    into 'build/generated/resources/main'
+}
+
+tasks.register('copyGGML_DLL', Copy) {
+    from '../../build/bin/Release'
+    include 'ggml.dll'
+    into 'build/generated/resources/main'
+}
+
+tasks.register('copyGGML_CPU_DLL', Copy) {
+    from '../../build/bin/Release'
+    include 'ggml-cpu.dll'
+    into 'build/generated/resources/main'
 }
 
 tasks.register('copyLibs') {
-    dependsOn copyLibwhisperDynlib, copyLibwhisperSo, copyWhisperDll
+    dependsOn copyLibwhisperDynlib, copyLibwhisperSo, copyWhisperDLL, copyGGML_BASE_DLL, copyGGML_DLL, copyGGML_CPU_DLL
 }
 
 test {

bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java

@@ -9,6 +9,7 @@ import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 
 public interface WhisperCppJnaLibrary extends Library {
+
     WhisperCppJnaLibrary instance = Native.load("whisper", WhisperCppJnaLibrary.class);
 
     String whisper_print_system_info();

bindings/ruby/.gitignore

@@ -1,3 +1,6 @@
 LICENSE
 pkg/
 lib/whisper.*
+ext/sources/*
+!ext/sources/CMakeGraphVizOptions.cmake
+ext/mkmf.log

bindings/ruby/README.md

@@ -16,6 +16,18 @@ If bundler is not being used to manage dependencies, install the gem by executin
 
     $ gem install whispercpp
 
+You can pass build options for whisper.cpp, for instance:
+
+    $ bundle config build.whispercpp --enable-ggml-cuda
+
+or,
+
+    $ gem install whispercpp -- --enable-ggml-cuda
+
+See whisper.cpp's [README](https://github.com/ggml-org/whisper.cpp/blob/master/README.md) for available options. You need convert options present the README to Ruby-style options.  
+For boolean options like `GGML_CUDA`, the README says `-DGGML_CUDA=1`. You need strip `-D`, prepend `--enable-` for `1` or `ON` (`--disable-` for `0` or `OFF`) and make it kebab-case: `--enable-ggml-cuda`.  
+For options which require arguments like `CMAKE_CUDA_ARCHITECTURES`, the README says `-DCMAKE_CUDA_ARCHITECTURES="86"`. You need strip `-D`, prepend `--`, make it kebab-case, append `=` and append argument: `--cmake-cuda-architectures="86"`.
+
 Usage
 -----
 

bindings/ruby/Rakefile

@@ -3,11 +3,15 @@ require "bundler/gem_tasks"
 require "rake/testtask"
 require_relative "extsources"
 
+SOURCES_DIR = "ext/sources"
+
 SOURCES = FileList[]
 
 EXTSOURCES.each do |src|
   basename = src.pathmap("%f")
-  dest = basename == "LICENSE" ? basename : src.pathmap("%{../..,ext}p")
+  dest = basename == "LICENSE" ? basename
+                               : src.pathmap("%{\\.\\./\\.\\.,#{SOURCES_DIR}}p")
+                                    .pathmap("%{\\.\\./javascript,#{SOURCES_DIR}/bindings/javascript}p")
   dir = dest.pathmap("%d")
   file src
   directory dir
@@ -18,7 +22,6 @@ EXTSOURCES.each do |src|
 end
 
 CLEAN.include SOURCES
-CLEAN.include FileList["ext/**/*.o", "ext/**/*.metal", "ext/**/*.tmp", "ext/whisper.{so,bundle,dll}"]
 
 SRC = FileList["ext/*.{c,cpp,h}"]
 
@@ -36,6 +39,20 @@ file "ext/Makefile" => SRC + ["ext/extconf.rb"] + SOURCES do |t|
     ruby "extconf.rb"
   end
 end
+if File.exist? "ext/Makefile"
+  task :make_clean do
+    cd "ext" do
+      sh "make", "clean"
+    end
+  end
+  task clean: :make_clean
+  task :make_distclean do
+    cd "ext" do
+      sh "make", "distclean"
+    end
+  end
+  task clobber: :make_distclean
+end
 
 file SO_FILE => "ext/Makefile" do |t|
   chdir "ext" do

bindings/ruby/ext/cpu.mk

@@ -1,13 +0,0 @@
-ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
-	ggml/src/ggml-cpu/ggml-cpu.cpp \
-	ggml/src/ggml-cpu/unary-ops.cpp \
-	ggml/src/ggml-cpu/binary-ops.cpp \
-	ggml/src/ggml-cpu/vec.cpp \
-	ggml/src/ggml-cpu/ops.cpp \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/include/ggml-cpu.h \
-	ggml/src/ggml-impl.h
-	$(CXX) $(CXXFLAGS)   -c $< -o $@

bindings/ruby/ext/dependencies.rb

@@ -0,0 +1,61 @@
+require "tsort"
+
+class Dependencies
+  def initialize(cmake, options)
+    @cmake = cmake
+    @options = options
+
+    generate_dot
+    @libs = parse_dot
+  end
+
+  def to_s
+    @libs.join(" ")
+  end
+
+  private
+
+  def dot_path
+    File.join(__dir__, "build", "whisper.cpp.dot")
+  end
+
+  def generate_dot
+    system @cmake, "-S", "sources", "-B", "build", "--graphviz", dot_path, "-D", "BUILD_SHARED_LIBS=OFF", @options.to_s, exception: true
+  end
+
+  def parse_dot
+    static_lib_shape = nil
+    nodes = {}
+    depends = Hash.new {|h, k| h[k] = []}
+
+    class << depends
+      include TSort
+      alias tsort_each_node each_key
+      def tsort_each_child(node, &block)
+        fetch(node, []).each(&block)
+      end
+    end
+
+    File.open(dot_path).each_line do |line|
+      case line
+      when /\[\s*label\s*=\s*"Static Library"\s*,\s*shape\s*=\s*(?<shape>\w+)\s*\]/
+        static_lib_shape = $~[:shape]
+      when /\A\s*"(?<node>\w+)"\s*\[\s*label\s*=\s*"(?<label>\S+)"\s*,\s*shape\s*=\s*(?<shape>\w+)\s*\]\s*;\s*\z/
+        node = $~[:node]
+        label = $~[:label]
+        shape = $~[:shape]
+        nodes[node] = [label, shape]
+      when /\A\s*"(?<depender>\w+)"\s*->\s*"(?<dependee>\w+)"/
+        depender = $~[:depender]
+        dependee = $~[:dependee]
+        depends[depender] ||= []
+        depends[depender] << dependee
+      end
+    end
+    depends.tsort.filter_map {|node|
+      label, shape = nodes[node]
+      shape == static_lib_shape ? label : nil
+    }.collect {|lib| "lib#{lib}.a"}
+      .reverse
+  end
+end

bindings/ruby/ext/extconf.rb

@@ -1,212 +1,22 @@
-require 'mkmf'
+require "mkmf"
+require_relative "options"
+require_relative "dependencies"
 
-# need to use c++ compiler flags
-$CXXFLAGS << ' -std=c++17'
+cmake = find_executable("cmake") || abort
+options = Options.new
+have_library("gomp") rescue nil
+libs = Dependencies.new(cmake, options)
 
-$LDFLAGS << ' -lstdc++'
+$INCFLAGS << " -Isources/include -Isources/ggml/include -Isources/examples"
+$LOCAL_LIBS << " #{libs}"
+$cleanfiles << " build #{libs}"
 
-# Set to true when building binary gems
-if enable_config('static-stdlib', false)
-  $LDFLAGS << ' -static-libgcc -static-libstdc++'
-end
-
-if enable_config('march-tune-native', false)
-  $CFLAGS << ' -march=native -mtune=native'
-  $CXXFLAGS << ' -march=native -mtune=native'
-end
-
-if ENV['WHISPER_METAL']
-  $GGML_METAL ||= true
-  $DEPRECATE_WARNING ||= true
-end
-
-$UNAME_S = `uname -s`.chomp
-$UNAME_P = `uname -p`.chomp
-$UNAME_M = `uname -m`.chomp
-
-if $UNAME_S == 'Darwin'
-  unless ENV['GGML_NO_METAL']
-    $GGML_METAL ||= true
-  end
-  $GGML_NO_OPENMP ||= true
-end
-
-if $GGML_METAL
-  $GGML_METAL_EMBED_LIBRARY = true
-end
-
-$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples -DGGML_USE_CPU'
-$MK_CFLAGS   = '-std=c11   -fPIC'
-$MK_CXXFLAGS = '-std=c++17 -fPIC'
-$MK_NVCCFLAGS = '-std=c++17'
-$MK_LDFLAGS = ''
-
-$OBJ_GGML = []
-$OBJ_WHISPER = []
-$OBJ_COMMON = []
-$OBJ_SDL = []
-
-$MK_CPPFLAGS << ' -D_XOPEN_SOURCE=600'
-
-if $UNAME_S == 'Linux'
-  $MK_CPPFLAGS << ' -D_GNU_SOURCE'
-end
-
-if $UNAME_S == 'Darwin'
-  $MK_CPPFLAGS << ' -D_DARWIN_C_SOURCE'
-end
-
-if ENV['WHISPER_DEBUG']
-  $MK_CFLAGS    << ' -O0 -g'
-  $MK_CXXFLAGS  << ' -O0 -g'
-  $MK_LDFLAGS   << ' -g'
-  $MK_NVCCFLAGS << ' -O0 -g'
-else
-  $MK_CPPFLAGS   << ' -DNDEBUG'
-  $MK_CFLAGS     << ' -O3'
-  $MK_CXXFLAGS   << ' -O3'
-  $MK_NVCCFLAGS  << ' -O3'
-end
-
-$WARN_FLAGS =
-  ' -Wall' <<
-  ' -Wextra' <<
-  ' -Wpedantic' <<
-  ' -Wcast-qual' <<
-  ' -Wno-unused-function'
-
-$MK_CFLAGS <<
-  $WARN_FLAGS <<
-  ' -Wshadow' <<
-  ' -Wstrict-prototypes' <<
-  ' -Wpointer-arith' <<
-  ' -Wmissing-prototypes' <<
-  ' -Werror=implicit-int' <<
-  ' -Werror=implicit-function-declaration'
-
-$MK_CXXFLAGS <<
-  $WARN_FLAGS <<
-  ' -Wmissing-declarations' <<
-  ' -Wmissing-noreturn'
-
-unless `#{cc_command} #{$LDFLAGS} -Wl,-v 2>&1`.chomp.include? 'dyld-1015.7'
-  $MK_CPPFLAGS << ' -DHAVE_BUGGY_APPLE_LINKER'
-end
-
-if %w[Linux Darwin FreeBSD NetBSD OpenBSD Haiku].include? $UNAME_S
-  $MK_CFLAGS   << ' -pthread'
-  $MK_CXXFLAGS << ' -pthread'
-end
-
-unless $_WIN32
-  $DSO_EXT = '.so'
-else
-  $DSO_EXT = '.dll'
-end
-
-unless ENV['RISCV']
-  if %w[x86_64 i686 amd64].include? $UNAME_M
-    $HOST_CXXFLAGS ||= ''
-
-    $MK_CFLAGS     << ' -march=native -mtune=native'
-    $HOST_CXXFLAGS << ' -march=native -mtune=native'
-  end
-else
-  $MK_CFLAGS   << ' -march=rv64gcv -mabi=lp64d'
-  $MK_CXXFLAGS << ' -march=rv64gcv -mabi=lp64d'
-end
-
-unless ENV['GGML_NO_ACCELERATE']
-  if $UNAME_S == 'Darwin'
-    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE'
-    $MK_CPPFLAGS << ' -DACCELERATE_NEW_LAPACK'
-    $MK_CPPFLAGS << ' -DACCELERATE_LAPACK_ILP64'
-    $MK_LDFLAGS  << ' -framework Accelerate'
-    $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
-  end
-end
-
-if ENV['GGML_OPENBLAS']
-  $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas`.chomp}"
-  $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas)`.chomp}"
-  $MK_LDFLAGS  << " #{`pkg-config --libs openblas`}"
-  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
-end
-
-if ENV['GGML_OPENBLAS64']
-  $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas64`.chomp}"
-  $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas64)`.chomp}"
-  $MK_LDFLAGS  << " #{`pkg-config --libs openblas64`}"
-  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
-end
-
-if $GGML_METAL
-  $MK_CPPFLAGS << ' -DGGML_USE_METAL'
-  $MK_LDFLAGS  << ' -framework Foundation -framework Metal -framework MetalKit'
-  $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal.o'
-
-  if ENV['GGML_METAL_NDEBUG']
-    $MK_CPPFLAGS << ' -DGGML_METAL_NDEBUG'
-  end
-
-  if $GGML_METAL_EMBED_LIBRARY
-    $MK_CPPFLAGS << ' -DGGML_METAL_EMBED_LIBRARY'
-    $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal-embed.o'
-  end
-end
-
-$OBJ_GGML <<
-  'ggml/src/ggml.o' <<
-  'ggml/src/ggml-alloc.o' <<
-  'ggml/src/ggml-backend.o' <<
-  'ggml/src/ggml-backend-reg.o' <<
-  'ggml/src/ggml-opt.o' <<
-  'ggml/src/ggml-quants.o' <<
-  'ggml/src/ggml-threading.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-cpp.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-aarch64.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-hbm.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-quants.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-traits.o' <<
-  'ggml/src/ggml-cpu/unary-ops.o' <<
-  'ggml/src/ggml-cpu/binary-ops.o' <<
-  'ggml/src/ggml-cpu/vec.o' <<
-  'ggml/src/ggml-cpu/ops.o'
-
-$OBJ_WHISPER <<
-  'src/whisper.o' <<
-  'examples/common.o' <<
-  'examples/common-whisper.o'
-
-$objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
-$objs <<
-  "ruby_whisper.o" <<
-  "ruby_whisper_context.o" <<
-  "ruby_whisper_transcribe.o" <<
-  "ruby_whisper_params.o" <<
-  "ruby_whisper_error.o" <<
-  "ruby_whisper_segment.o" <<
-  "ruby_whisper_model.o"
-
-$CPPFLAGS  = "#{$MK_CPPFLAGS} #{$CPPFLAGS}"
-$CFLAGS    = "#{$CPPFLAGS} #{$MK_CFLAGS} #{$GF_CFLAGS} #{$CFLAGS}"
-$BASE_CXXFLAGS = "#{$MK_CXXFLAGS} #{$CXXFLAGS}"
-$CXXFLAGS  = "#{$BASE_CXXFLAGS} #{$HOST_CXXFLAGS} #{$GF_CXXFLAGS} #{$CPPFLAGS}"
-$NVCCFLAGS = "#{$MK_NVCCFLAGS} #{$NVCCFLAGS}"
-$LDFLAGS   = "#{$MK_LDFLAGS} #{$LDFLAGS}"
-
-create_makefile('whisper')
-
-File.open 'Makefile', 'a' do |file|
-  file.puts 'include scripts/get-flags.mk'
-  file.puts 'include cpu.mk'
-
-  if $GGML_METAL
-    file.puts 'include metal.mk'
-
-    if $GGML_METAL_EMBED_LIBRARY
-      file.puts 'include metal-embed.mk'
-    end
-  end
+create_makefile "whisper" do |conf|
+  conf << <<~EOF
+    $(TARGET_SO): #{libs}
+    #{libs}: cmake-targets
+    cmake-targets:
+    #{"\t"}#{cmake} -S sources -B build -D BUILD_SHARED_LIBS=OFF -D CMAKE_ARCHIVE_OUTPUT_DIRECTORY=#{__dir__} -D CMAKE_POSITION_INDEPENDENT_CODE=ON #{options}
+    #{"\t"}#{cmake} --build build --config Release --target common whisper
+  EOF
 end

bindings/ruby/ext/metal-embed.mk

@@ -1,17 +0,0 @@
-ggml/src/ggml-metal/ggml-metal-embed.o: \
-	ggml/src/ggml-metal/ggml-metal.metal \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
-	ggml/src/ggml-common.h
-	@echo "Embedding Metal library"
-	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
-	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
-	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
-	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
-	@rmdir ${TEMP_ASSEMBLY}

bindings/ruby/ext/metal.mk

@@ -1,6 +0,0 @@
-ggml/src/ggml-metal/ggml-metal.o: \
-	ggml/src/ggml-metal/ggml-metal.m \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
-	ggml/include/ggml-metal.h \
-	ggml/include/ggml.h
-	$(CC) $(CFLAGS) -c $< -o $@

bindings/ruby/ext/options.rb

@@ -0,0 +1,219 @@
+class Options
+  def initialize
+    @options = {}
+    @pending_options = []
+    @ignored_options = []
+
+    configure
+  end
+
+  def help
+    @options
+      .collect_concat {|name, (type, value)|
+        option = option_name(name)
+        if type == :bool
+          ["--enable-#{option}", "--disable-#{option}"]
+        else
+          "--#{option}=#{type.upcase}"
+        end
+      }
+      .join($/)
+  end
+
+  def to_s
+    @options
+      .reject {|name, (type, value)| value.nil?}
+      .collect {|name, (type, value)| "-D #{name}=#{value == true ? "ON" : value == false ? "OFF" : value.shellescape}"}
+      .join(" ")
+  end
+
+  def cmake_options
+    return @cmake_options if @cmake_options
+
+    output = nil
+    Dir.chdir __dir__ do
+      output = `cmake -S sources -B build -L`
+    end
+    started = false
+    @cmake_options = output.lines.filter_map {|line|
+      if line.chomp == "-- Cache values"
+        started = true
+        next
+      end
+      next unless started
+      option, value = line.chomp.split("=", 2)
+      name, type = option.split(":", 2)
+      [name, type, value]
+    }
+  end
+
+  def missing_options
+    cmake_options.collect {|name, type, value| name} -
+      @options.keys - @pending_options - @ignored_options
+  end
+
+  def extra_options
+    @options.keys + @pending_options - @ignored_options -
+      cmake_options.collect {|name, type, value| name}
+  end
+
+  private
+
+  def configure
+    filepath "ACCELERATE_FRAMEWORK"
+    ignored "BUILD_SHARED_LIBS"
+    ignored "BUILD_TESTING"
+    ignored "CMAKE_BUILD_TYPE"
+    ignored "CMAKE_INSTALL_PREFIX"
+    string "CMAKE_OSX_ARCHITECTURES"
+    ignored "CMAKE_OSX_DEPLOYMENT_TARGET"
+    string "CMAKE_OSX_SYSROOT"
+    filepath "FOUNDATION_LIBRARY"
+    bool "GGML_ACCELERATE"
+    bool "GGML_ALL_WARNINGS_3RD_PARTY"
+    bool "GGML_AMX_BF16"
+    bool "GGML_AMX_INT8"
+    bool "GGML_AMX_TILE"
+    bool "GGML_AVX"
+    bool "GGML_AVX2"
+    bool "GGML_AVX512"
+    bool "GGML_AVX512_BF16"
+    bool "GGML_AVX512_VBMI"
+    bool "GGML_AVX512_VNNI"
+    bool "GGML_AVX_VNNI"
+    ignored "GGML_BACKEND_DL"
+    ignored "GGML_BIN_INSTALL_DIR"
+    bool "GGML_BLAS"
+    string "GGML_BLAS_VENDOR"
+    bool "GGML_BMI2"
+    ignored "GGML_BUILD_EXAMPLES"
+    ignored "GGML_BUILD_TESTS"
+    filepath "GGML_CCACHE_FOUND"
+    bool "GGML_CPU"
+    bool "GGML_CPU_AARCH64"
+    ignored "GGML_CPU_ALL_VARIANTS"
+    string "GGML_CPU_ARM_ARCH"
+    bool "GGML_CPU_HBM"
+    bool "GGML_CPU_KLEIDIAI"
+    string "GGML_CPU_POWERPC_CPUTYPE"
+    bool "GGML_CUDA"
+    string "GGML_CUDA_COMPRESSION_MODE"
+    bool "GGML_CUDA_F16"
+    bool "GGML_CUDA_FA"
+    bool "GGML_CUDA_FA_ALL_QUANTS"
+    bool "GGML_CUDA_FORCE_CUBLAS"
+    bool "GGML_CUDA_FORCE_MMQ"
+    ignored "GGML_CUDA_GRAPHS"
+    bool "GGML_CUDA_NO_PEER_COPY"
+    bool "GGML_CUDA_NO_VMM"
+    string "GGML_CUDA_PEER_MAX_BATCH_SIZE"
+    bool "GGML_F16C"
+    bool "GGML_FMA"
+    bool "GGML_GPROF"
+    bool "GGML_HIP"
+    bool "GGML_HIP_GRAPHS"
+    bool "GGML_HIP_NO_VMM"
+    bool "GGML_HIP_ROCWMMA_FATTN"
+    ignored "GGML_INCLUDE_INSTALL_DIR"
+    bool "GGML_KOMPUTE"
+    bool "GGML_LASX"
+    ignored "GGML_LIB_INSTALL_DIR"
+    ignored "GGML_LLAMAFILE"
+    bool "GGML_LSX"
+    bool "GGML_LTO"
+    bool "GGML_METAL"
+    bool "GGML_METAL_EMBED_LIBRARY"
+    string "GGML_METAL_MACOSX_VERSION_MIN"
+    bool "GGML_METAL_NDEBUG"
+    bool "GGML_METAL_SHADER_DEBUG"
+    string "GGML_METAL_STD"
+    bool "GGML_METAL_USE_BF16"
+    bool "GGML_MUSA"
+    bool "GGML_NATIVE"
+    bool "GGML_OPENCL"
+    bool "GGML_OPENCL_EMBED_KERNELS"
+    bool "GGML_OPENCL_PROFILING"
+    string "GGML_OPENCL_TARGET_VERSION"
+    bool "GGML_OPENCL_USE_ADRENO_KERNELS"
+    bool "GGML_OPENMP"
+    bool "GGML_RPC"
+    bool "GGML_RVV"
+    bool "GGML_RV_ZFH"
+    pending "GGML_SCCACHE_FOUND"
+    string "GGML_SCHED_MAX_COPIES"
+    bool "GGML_SSE42"
+    ignored "GGML_STATIC"
+    bool "GGML_SYCL"
+    string "GGML_SYCL_DEVICE_ARCH"
+    bool "GGML_SYCL_F16"
+    bool "GGML_SYCL_GRAPH"
+    string "GGML_SYCL_TARGET"
+    bool "GGML_VULKAN"
+    bool "GGML_VULKAN_CHECK_RESULTS"
+    bool "GGML_VULKAN_DEBUG"
+    bool "GGML_VULKAN_MEMORY_DEBUG"
+    bool "GGML_VULKAN_PERF"
+    ignored "GGML_VULKAN_RUN_TESTS"
+    filepath "GGML_VULKAN_SHADERS_GEN_TOOLCHAIN"
+    bool "GGML_VULKAN_SHADER_DEBUG_INFO"
+    pending "GGML_VULKAN_VALIDATE"
+    bool "GGML_VXE"
+    filepath "GIT_EXE"
+    filepath "MATH_LIBRARY"
+    filepath "METALKIT_FRAMEWORK"
+    filepath "METAL_FRAMEWORK"
+    bool "WHISPER_ALL_WARNINGS"
+    bool "WHISPER_ALL_WARNINGS_3RD_PARTY"
+    ignored "WHISPER_BIN_INSTALL_DIR"
+    ignored "WHISPER_BUILD_EXAMPLES"
+    ignored "WHISPER_BUILD_SERVER"
+    ignored"WHISPER_BUILD_TESTS"
+    bool "WHISPER_CCACHE"
+    bool "WHISPER_COREML"
+    bool "WHISPER_COREML_ALLOW_FALLBACK"
+    ignored "WHISPER_CURL"
+    bool "WHISPER_FATAL_WARNINGS"
+    ignored "WHISPER_FFMPEG"
+    ignored "WHISPER_INCLUDE_INSTALL_DIR"
+    ignored "WHISPER_LIB_INSTALL_DIR"
+    bool "WHISPER_OPENVINO"
+    bool "WHISPER_SANITIZE_ADDRESS"
+    bool "WHISPER_SANITIZE_THREAD"
+    bool "WHISPER_SANITIZE_UNDEFINED"
+    ignored "WHISPER_SDL2"
+    pending "WHISPER_USE_SYSTEM_GGML"
+  end
+
+  def option_name(name)
+    name.downcase.gsub("_", "-")
+  end
+
+  def bool(name)
+    option = option_name(name)
+    value = enable_config(option)
+    @options[name] = [:bool, value]
+  end
+
+  def string(name, type=:string)
+    option = "--#{option_name(name)}"
+    value = arg_config(option)
+    raise "String expected for #{option}" if value == true || value&.empty?
+    @options[name] = [type, value]
+  end
+
+  def path(name)
+    string(name, :path)
+  end
+
+  def filepath(name)
+    string(name, :filepath)
+  end
+
+  def pending(name)
+    @pending_options << name
+  end
+
+  def ignored(name)
+    @ignored_options << name
+  end
+end

bindings/ruby/ext/ruby_whisper.h

@@ -19,6 +19,7 @@ typedef struct {
   bool diarize;
   ruby_whisper_callback_container *new_segment_callback_container;
   ruby_whisper_callback_container *progress_callback_container;
+  ruby_whisper_callback_container *encoder_begin_callback_container;
   ruby_whisper_callback_container *abort_callback_container;
 } ruby_whisper_params;
 

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 30
+#define RUBY_WHISPER_PARAMS_PARAM_NAMES_COUNT 32
 
 extern VALUE cParams;
 
@@ -63,6 +63,8 @@ static ID id_new_segment_callback;
 static ID id_new_segment_callback_user_data;
 static ID id_progress_callback;
 static ID id_progress_callback_user_data;
+static ID id_encoder_begin_callback;
+static ID id_encoder_begin_callback_user_data;
 static ID id_abort_callback;
 static ID id_abort_callback_user_data;
 
@@ -126,6 +128,33 @@ static void progress_callback(struct whisper_context *ctx, struct whisper_state
   }
 }
 
+static bool encoder_begin_callback(struct whisper_context *ctx, struct whisper_state *state, void *user_data) {
+  const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
+  bool is_aborted = false;
+  VALUE result;
+
+  // Currently, doesn't support state because
+  // those require to resolve GC-related problems.
+  if (!NIL_P(container->callback)) {
+    result = rb_funcall(container->callback, id_call, 3, *container->context, Qnil, container->user_data);
+    if (result == Qfalse) {
+      is_aborted = true;
+    }
+  }
+  const long callbacks_len = RARRAY_LEN(container->callbacks);
+  if (0 == callbacks_len) {
+    return !is_aborted;
+  }
+  for (int j = 0; j < callbacks_len; j++) {
+    VALUE cb = rb_ary_entry(container->callbacks, j);
+    result = rb_funcall(cb, id_call, 0);
+    if (result == Qfalse) {
+      is_aborted = true;
+    }
+  }
+  return !is_aborted;
+}
+
 static bool abort_callback(void * user_data) {
   const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
   if (!NIL_P(container->callback)) {
@@ -161,6 +190,12 @@ void register_callbacks(ruby_whisper_params * rwp, VALUE * context) {
     rwp->params.progress_callback_user_data = rwp->progress_callback_container;
   }
 
+  if (!NIL_P(rwp->encoder_begin_callback_container->callback) || 0 != RARRAY_LEN(rwp->encoder_begin_callback_container->callbacks)) {
+    rwp->encoder_begin_callback_container->context = context;
+    rwp->params.encoder_begin_callback = encoder_begin_callback;
+    rwp->params.encoder_begin_callback_user_data = rwp->encoder_begin_callback_container;
+  }
+
   if (!NIL_P(rwp->abort_callback_container->callback) || 0 != RARRAY_LEN(rwp->abort_callback_container->callbacks)) {
     rwp->abort_callback_container->context = context;
     rwp->params.abort_callback = abort_callback;
@@ -173,6 +208,7 @@ rb_whisper_params_mark(ruby_whisper_params *rwp)
 {
   rb_whisper_callbcack_container_mark(rwp->new_segment_callback_container);
   rb_whisper_callbcack_container_mark(rwp->progress_callback_container);
+  rb_whisper_callbcack_container_mark(rwp->encoder_begin_callback_container);
   rb_whisper_callbcack_container_mark(rwp->abort_callback_container);
 }
 
@@ -198,6 +234,7 @@ ruby_whisper_params_allocate(VALUE klass)
   rwp->diarize = false;
   rwp->new_segment_callback_container = rb_whisper_callback_container_allocate();
   rwp->progress_callback_container = rb_whisper_callback_container_allocate();
+  rwp->encoder_begin_callback_container = rb_whisper_callback_container_allocate();
   rwp->abort_callback_container = rb_whisper_callback_container_allocate();
   return Data_Wrap_Struct(klass, rb_whisper_params_mark, rb_whisper_params_free, rwp);
 }
@@ -849,6 +886,57 @@ ruby_whisper_params_set_progress_callback_user_data(VALUE self, VALUE value)
   rwp->progress_callback_container->user_data = value;
   return value;
 }
+
+static VALUE
+ruby_whisper_params_get_encoder_begin_callback(VALUE self)
+{
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return rwp->encoder_begin_callback_container->callback;
+}
+
+/*
+ * Sets encoder begin callback, called when the encoder starts.
+ *
+ *   params.encoder_begin_callback = ->(context, _, user_data) {
+ *     # ...
+ *   }
+ *
+ * call-seq:
+ *   encoder_begin_callback = callback -> callback
+ */
+static VALUE
+ruby_whisper_params_set_encoder_begin_callback(VALUE self, VALUE value)
+{
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->encoder_begin_callback_container->callback = value;
+  return value;
+}
+
+static VALUE
+ruby_whisper_params_get_encoder_begin_callback_user_data(VALUE self)
+{
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return rwp->encoder_begin_callback_container->user_data;
+}
+
+/*
+ * Sets user data passed to the last argument of encoder begin callback.
+ *
+ * call-seq:
+ *   encoder_begin_callback_user_data = user_data -> use_data
+ */
+static VALUE
+ruby_whisper_params_set_encoder_begin_callback_user_data(VALUE self, VALUE value)
+{
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->encoder_begin_callback_container->user_data = value;
+  return value;
+}
+
 static VALUE
 ruby_whisper_params_get_abort_callback(VALUE self)
 {
@@ -958,6 +1046,8 @@ ruby_whisper_params_initialize(int argc, VALUE *argv, VALUE self)
       SET_PARAM_IF_SAME(new_segment_callback_user_data)
       SET_PARAM_IF_SAME(progress_callback)
       SET_PARAM_IF_SAME(progress_callback_user_data)
+      SET_PARAM_IF_SAME(encoder_begin_callback)
+      SET_PARAM_IF_SAME(encoder_begin_callback_user_data)
       SET_PARAM_IF_SAME(abort_callback)
       SET_PARAM_IF_SAME(abort_callback_user_data)
     }
@@ -1008,6 +1098,26 @@ ruby_whisper_params_on_progress(VALUE self)
   return Qnil;
 }
 
+/*
+ * Hook called when the encoder starts.
+ *
+ *   whisper.on_encoder_begin do
+ *     # ...
+ *   end
+ *
+ * call-seq:
+ *   on_encoder_begin { ... }
+ */
+static VALUE
+ruby_whisper_params_on_encoder_begin(VALUE self)
+{
+  ruby_whisper_params *rws;
+  Data_Get_Struct(self, ruby_whisper_params, rws);
+  const VALUE blk = rb_block_proc();
+  rb_ary_push(rws->encoder_begin_callback_container->callbacks, blk);
+  return Qnil;
+}
+
 /*
  * Call block to determine whether abort or not. Return +true+ when you want to abort.
  *
@@ -1068,10 +1178,13 @@ init_ruby_whisper_params(VALUE *mWhisper)
   DEFINE_PARAM(new_segment_callback_user_data, 25)
   DEFINE_PARAM(progress_callback, 26)
   DEFINE_PARAM(progress_callback_user_data, 27)
-  DEFINE_PARAM(abort_callback, 28)
-  DEFINE_PARAM(abort_callback_user_data, 29)
+  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)
 
   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);
+  rb_define_method(cParams, "on_encoder_begin", ruby_whisper_params_on_encoder_begin, 0);
   rb_define_method(cParams, "abort_on", ruby_whisper_params_abort_on, 0);
 }

bindings/ruby/ext/ruby_whisper_transcribe.cpp

@@ -50,15 +50,16 @@ ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
     fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
     return self;
   }
-  {
-    static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+  // Commented out because it is work in progress
+  // {
+  //   static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
-      bool is_aborted = *(bool*)user_data;
-      return !is_aborted;
-    };
-    rwp->params.encoder_begin_callback_user_data = &is_aborted;
-  }
+  //   rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+  //     bool is_aborted = *(bool*)user_data;
+  //     return !is_aborted;
+  //   };
+  //   rwp->params.encoder_begin_callback_user_data = &is_aborted;
+  // }
 
   register_callbacks(rwp, &self);
 

bindings/ruby/ext/sources/CMakeGraphVizOptions.cmake

@@ -0,0 +1,8 @@
+set(GRAPHVIZ_EXECUTABLES FALSE)
+set(GRAPHVIZ_STATIC_LIBS TRUE)
+set(GRAPHVIZ_SHARED_LIBS FALSE)
+set(GRAPHVIZ_MODULE_LIBS FALSE)
+set(GRAPHVIZ_INTERFACE_LIBS FALSE)
+set(GRAPHVIZ_OBJECT_LIBS FALSE)
+set(GRAPHVIZ_UNKNOWN_LIBS FALSE)
+set(GRAPHVIZ_GENERATE_DEPENDERS FALSE)

bindings/ruby/extsources.rb

@@ -1,6 +1,34 @@
-require "yaml"
+ignored_dirs = %w[
+  .devops
+  examples/wchess/wchess.wasm
+  examples/whisper.android
+  examples/whisper.android.java
+  examples/whisper.objc
+  examples/whisper.swiftui
+  grammars
+  models
+  samples
+  scripts
+]
+ignored_files = %w[
+  AUTHORS
+  Makefile
+  README.md
+  README_sycl.md
+  .gitignore
+  .gitmodules
+  whisper.nvim
+  twitch.sh
+  yt-wsp.sh
+]
 
-sources = `git ls-files -z ../..`.split("\x0")
-paths = YAML.load_file("../../.github/workflows/bindings-ruby.yml")[true]["push"]["paths"]
-paths.delete "bindings/ruby/**"
-EXTSOURCES = (Dir.glob(paths, base: "../..").collect {|path| "../../#{path}"} << "../../LICENSE") & sources
+EXTSOURCES =
+  `git ls-files -z ../..`.split("\x0")
+    .select {|file|
+      basename = File.basename(file)
+
+      ignored_dirs.all? {|dir| !file.start_with?("../../#{dir}")} &&
+        !ignored_files.include?(basename) &&
+        (file.start_with?("../..") || file.start_with?("../javascript")) &&
+        (!file.start_with?("../../.github/") || basename == "bindings-ruby.yml")
+    }

bindings/ruby/lib/whisper/model/uri.rb

@@ -53,7 +53,7 @@ module Whisper
           http.request request do |response|
             case response
             when Net::HTTPNotModified
-            # noop
+              # noop
             when Net::HTTPOK
               download response
             when Net::HTTPRedirection
@@ -68,7 +68,7 @@ module Whisper
       rescue => err
         if cache_path.exist?
           warn err
-        # Use cache file
+          # Use cache file
         else
           raise
         end

bindings/ruby/sig/whisper.rbs

@@ -7,6 +7,7 @@ module Whisper
   type log_callback = ^(Integer level, String message, Object user_data) -> void
   type new_segment_callback = ^(Whisper::Context, void, Integer n_new, Object user_data) -> void
   type progress_callback = ^(Whisper::Context, void, Integer progress, Object user_data) -> void
+  type encoder_begin_callback = ^(Whisper::Context, void, Object user_data) -> void
   type abort_callback = ^(Whisper::Context, void, Object user_data) -> boolish
 
   LOG_LEVEL_NONE: Integer
@@ -23,9 +24,20 @@ module Whisper
   def self.log_set: (log_callback, Object? user_data) -> log_callback
 
   class Context
-    def self.new: (string | _ToPath | ::URI::HTTP) -> instance
+    def self.new: (path | ::URI::HTTP) -> instance
+
+    # transcribe a single file
+    # can emit to a block results
+    #
+    #   params = Whisper::Params.new
+    #   params.duration = 60_000
+    #   whisper.transcribe "path/to/audio.wav", params do |text|
+    #     puts text
+    #   end
+    #
     def transcribe: (string, Params) -> self
                   | (string, Params) { (String) -> void } -> self
+
     def model_n_vocab: () -> Integer
     def model_n_audio_ctx: () -> Integer
     def model_n_audio_state: () -> Integer
@@ -34,19 +46,72 @@ module Whisper
     def model_n_mels: () -> Integer
     def model_ftype: () -> Integer
     def model_type: () -> String
+
+    # Yields each Whisper::Segment:
+    #
+    #   whisper.transcribe("path/to/audio.wav", params)
+    #   whisper.each_segment do |segment|
+    #     puts segment.text
+    #   end
+    #
+    # Returns an Enumerator if no block given:
+    #
+    #   whisper.transcribe("path/to/audio.wav", params)
+    #   enum = whisper.each_segment
+    #   enum.to_a # => [#<Whisper::Segment>, ...]
+    #
     def each_segment: { (Segment) -> void } -> void
                     | () -> Enumerator[Segment]
+
     def model: () -> Model
     def full_get_segment: (Integer nth) -> Segment
     def full_n_segments: () -> Integer
+
+    # Language ID, which can be converted to string by Whisper.lang_str and Whisper.lang_str_full.
+    #
     def full_lang_id: () -> Integer
+
+    # Start time of a segment indexed by +segment_index+ in centiseconds (10 times milliseconds).
+    #
+    #   full_get_segment_t0(3) # => 1668 (16680 ms)
+    #
     def full_get_segment_t0: (Integer) -> Integer
+
+    # End time of a segment indexed by +segment_index+ in centiseconds (10 times milliseconds).
+    #
+    #   full_get_segment_t1(3) # => 1668 (16680 ms)
+    #
     def full_get_segment_t1: (Integer) -> Integer
+
+    # Whether the next segment indexed by +segment_index+ is predicated as a speaker turn.
+    #
+    #   full_get_segment_speacker_turn_next(3) # => true
+    #
     def full_get_segment_speaker_turn_next: (Integer) -> (true | false)
+
+    # Text of a segment indexed by +segment_index+.
+    #
+    #   full_get_segment_text(3) # => "ask not what your country can do for you, ..."
+    #
     def full_get_segment_text: (Integer) -> String
+
     def full_get_segment_no_speech_prob: (Integer) -> Float
+
+    # Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+    # Not thread safe for same context
+    # Uses the specified decoding strategy to obtain the text.
+    #
+    # The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
+    #
     def full: (Params, Array[Float] samples, ?Integer n_samples) -> self
             | (Params, _Samples, ?Integer n_samples) -> self
+
+    # Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+    # Result is stored in the default state of the context
+    # Not thread safe if executed in parallel on the same context.
+    # It seems this approach can offer some speedup in some cases.
+    # However, the transcription accuracy can be worse at the beginning and end of each chunk.
+    #
     def full_parallel: (Params, Array[Float], ?Integer n_samples) -> self
                      | (Params, _Samples, ?Integer n_samples) -> self
                      | (Params, _Samples, ?Integer? n_samples, Integer n_processors) -> self
@@ -82,71 +147,223 @@ module Whisper
       ?new_segment_callback_user_data: Object,
       ?progress_callback: progress_callback,
       ?progress_callback_user_data: Object,
+      ?encoder_begin_callback: encoder_begin_callback,
+      ?encoder_begin_callback_user_data: Object,
       ?abort_callback: abort_callback,
       ?abort_callback_user_data: Object
     ) -> instance
+
+    # params.language = "auto" | "en", etc...
+    #
     def language=: (String) -> String # TODO: Enumerate lang names
+
     def language: () -> String
     def translate=: (boolish) -> boolish
     def translate: () -> (true | false)
     def no_context=: (boolish) -> boolish
+
+    # If true, does not use past transcription (if any) as initial prompt for the decoder.
+    #
     def no_context: () -> (true | false)
+
     def single_segment=: (boolish) -> boolish
+
+    # If true, forces single segment output (useful for streaming).
+    #
     def single_segment: () -> (true | false)
+
     def print_special=: (boolish) -> boolish
+
+    # If true, prints special tokens (e.g. <SOT>, <EOT>, <BEG>, etc.).
+    #
     def print_special: () -> (true | false)
+
     def print_progress=: (boolish) -> boolish
+
+    # If true, prints progress information.
+    #
     def print_progress: () -> (true | false)
+
     def print_realtime=: (boolish) -> boolish
+
+    # If true, prints results from within whisper.cpp. (avoid it, use callback instead)
+    #
     def print_realtime: () -> (true | false)
+
+    # If true, prints timestamps for each text segment when printing realtime.
+    #
     def print_timestamps=: (boolish) -> boolish
+
     def print_timestamps: () -> (true | false)
+
     def suppress_blank=: (boolish) -> boolish
+
+    # If true, suppresses blank outputs.
+    #
     def suppress_blank: () -> (true | false)
+
     def suppress_nst=: (boolish) -> boolish
+
+    # If true, suppresses non-speech-tokens.
+    #
     def suppress_nst: () -> (true | false)
+
     def token_timestamps=: (boolish) -> boolish
+
+    # If true, enables token-level timestamps.
+    #
     def token_timestamps: () -> (true | false)
+
     def split_on_word=: (boolish) -> boolish
+
+    # If true, split on word rather than on token (when used with max_len).
+    #
     def split_on_word: () -> (true | false)
+
     def initial_prompt=: (_ToS) -> _ToS
+
+    # Tokens to provide to the whisper decoder as initial prompt
+    # these are prepended to any existing text context from a previous call
+    # use whisper_tokenize() to convert text to tokens.
+    # Maximum of whisper_n_text_ctx()/2 tokens are used (typically 224).
+    #
     def initial_prompt: () -> (String | nil)
+
     def diarize=: (boolish) -> boolish
+
+    # If true, enables diarization.
+    #
     def diarize: () -> (true | false)
+
     def offset=: (Integer) -> Integer
+
+    # Start offset in ms.
+    #
     def offset: () -> Integer
+
     def duration=: (Integer) -> Integer
+
+    # Audio duration to process in ms.
+    #
     def duration: () -> Integer
+
     def max_text_tokens=: (Integer) -> Integer
+
+    # Max tokens to use from past text as prompt for the decoder.
+    #
     def max_text_tokens: () -> Integer
+
     def temperature=: (Float) -> Float
     def temperature: () -> Float
     def max_initial_ts=: (Float) -> Float
+
+    # See https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L97
+    #
     def max_initial_ts: () -> Float
+
     def length_penalty=: (Float) -> Float
     def length_penalty: () -> Float
     def temperature_inc=: (Float) -> Float
     def temperature_inc: () -> Float
     def entropy_thold=: (Float) -> Float
+
+    # Similar to OpenAI's "compression_ratio_threshold"
+    #
     def entropy_thold: () -> Float
+
     def logprob_thold=: (Float) -> Float
     def logprob_thold: () -> Float
     def no_speech_thold=: (Float) -> Float
     def no_speech_thold: () -> Float
+
+    # Sets new segment callback, called for every newly generated text segment.
+    #
+    #   params.new_segment_callback = ->(context, _, n_new, user_data) {
+    #     # ...
+    #   }
+    #
     def new_segment_callback=: (new_segment_callback) -> new_segment_callback
     def new_segment_callback: () -> (new_segment_callback | nil)
+
+    # Sets user data passed to the last argument of new segment callback.
+    #
     def new_segment_callback_user_data=: (Object) -> Object
+
     def new_segment_callback_user_data: () -> Object
+
+    # Sets progress callback, called on each progress update.
+    #
+    #   params.new_segment_callback = ->(context, _, progress, user_data) {
+    #     # ...
+    #   }
+    #
+    # +progress+ is an Integer between 0 and 100.
+    #
     def progress_callback=: (progress_callback) -> progress_callback
+
     def progress_callback: () -> (progress_callback | nil)
+
+    # Sets user data passed to the last argument of progress callback.
+    #
     def progress_callback_user_data=: (Object) -> Object
+
     def progress_callback_user_data: () -> Object
+
+    # Sets encoder begin callback, called when the encoder starts.
+    #
+    def encoder_begin_callback=: (encoder_begin_callback) -> encoder_begin_callback
+
+    def encoder_begin_callback: () -> (encoder_begin_callback | nil)
+
+    # Sets user data passed to the last argument of encoder begin callback.
+    #
+    def encoder_begin_callback_user_data=: (Object) -> Object
+
+    def encoder_begin_callback_user_data: () -> Object
+
+    # Sets abort callback, called to check if the process should be aborted.
+    #
+    #   params.abort_callback = ->(user_data) {
+    #     # ...
+    #   }
+    #
+    #
     def abort_callback=: (abort_callback) -> abort_callback
+
     def abort_callback: () -> (abort_callback | nil)
+
+    # Sets user data passed to the last argument of abort callback.
+    #
     def abort_callback_user_data=: (Object) -> Object
+
     def abort_callback_user_data: () -> Object
+
+    # Hook called on new segment. Yields each Whisper::Segment.
+    #
+    #   whisper.on_new_segment do |segment|
+    #     # ...
+    #   end
+    #
     def on_new_segment: { (Segment) -> void } -> void
+
+    # Hook called on progress update. Yields each progress Integer between 0 and 100.
+    #
     def on_progress: { (Integer progress) -> void } -> void
+
+    # Hook called on encoder starts.
+    #
+    def on_encoder_begin: { () -> void } -> void
+
+    # Call block to determine whether abort or not. Return +true+ when you want to abort.
+    #
+    #   params.abort_on do
+    #     if some_condition
+    #       true # abort
+    #     else
+    #       false # continue
+    #     end
+    #   end
+    #
     def abort_on: { (Object user_data) -> boolish } -> void
   end
 
@@ -167,16 +384,24 @@ module Whisper
     def type: () -> String
 
     class URI
-      def self.new: (string | ::URI::HTTP) -> self
+      def self.new: (string | ::URI::HTTP) -> instance
       def to_path: -> String
       def clear_cache: -> void
     end
   end
 
   class Segment
+    # Start time in milliseconds.
+    #
     def start_time: () -> Integer
+
+    # End time in milliseconds.
+    #
     def end_time: () -> Integer
+
+    # Whether the next segment is predicted as a speaker turn.
     def speaker_next_turn?: () -> (true | false)
+
     def text: () -> String
     def no_speech_prob: () -> Float
   end

bindings/ruby/tests/helper.rb

@@ -6,9 +6,9 @@ class TestBase < Test::Unit::TestCase
   AUDIO = File.join(__dir__, "..", "..", "..", "samples", "jfk.wav")
 
   class << self
-    attr_reader :whisper
+    def whisper
+      return @whisper if @whisper
 
-    def startup
       @whisper = Whisper::Context.new("base.en")
       params = Whisper::Params.new
       params.print_timestamps = false
@@ -21,4 +21,15 @@ class TestBase < Test::Unit::TestCase
   def whisper
     self.class.whisper
   end
+
+  module BuildOptions
+    load "ext/options.rb", self
+    Options.include self
+
+    def enable_config(name)
+    end
+
+    def arg_config(name)
+    end
+  end
 end

bindings/ruby/tests/test_callback.rb

@@ -111,6 +111,48 @@ class TestCallback < TestBase
     assert_equal 100, last
   end
 
+  def test_encoder_begin_callback
+    i = 0
+    @params.encoder_begin_callback = ->(context, state, user_data) {
+      i += 1
+    }
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+
+  def test_encoder_begin_callback_abort
+    logs = []
+    Whisper.log_set -> (level, buffer, user_data) {
+      logs << buffer if level == Whisper::LOG_LEVEL_ERROR
+    }, logs
+    @params.encoder_begin_callback = ->(context, state, user_data) {
+      return false
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_match(/encoder_begin_callback returned false - aborting/, logs.join)
+    Whisper.log_set ->(level, buffer, user_data) {}, nil
+  end
+
+  def test_encoder_begin_callback_user_data
+    udata = Object.new
+    @params.encoder_begin_callback_user_data = udata
+    yielded = nil
+    @params.encoder_begin_callback = ->(context, state, user_data) {
+      yielded = user_data
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_same udata, yielded
+  end
+
+  def test_on_encoder_begin
+    i = 0
+    @params.on_encoder_begin do
+      i += 1
+    end
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+
   def test_abort_callback
     i = 0
     @params.abort_callback = ->(user_data) {

bindings/ruby/tests/test_package.rb

@@ -21,11 +21,26 @@ class TestPackage < TestBase
       match_data = `rake -Tbuild`.match(/(whispercpp-(.+)\.gem)/)
       filename = match_data[1]
       version = match_data[2]
-      basename = "whisper.#{RbConfig::CONFIG["DLEXT"]}"
       Dir.mktmpdir do |dir|
         system "gem", "install", "--install-dir", dir.shellescape, "--no-document", "pkg/#{filename.shellescape}", exception: true
-        assert_path_exist File.join(dir, "gems/whispercpp-#{version}/lib", basename)
+        assert_installed dir, version
       end
     end
+
+    private
+
+    def assert_installed(dir, version)
+      assert_path_exist File.join(dir, "gems/whispercpp-#{version}/lib", "whisper.#{RbConfig::CONFIG["DLEXT"]}")
+      assert_path_exist File.join(dir, "gems/whispercpp-#{version}/LICENSE")
+      assert_path_not_exist File.join(dir, "gems/whispercpp-#{version}/ext/build")
+    end
+  end
+
+  def test_build_options
+    options = BuildOptions::Options.new
+    assert_empty options.missing_options
+    unless ENV["CI"]
+      assert_empty options.extra_options
+    end
   end
 end

bindings/ruby/whispercpp.gemspec

@@ -3,8 +3,8 @@ require_relative "extsources"
 Gem::Specification.new do |s|
   s.name    = "whispercpp"
   s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
-  s.version = '1.3.1'
-  s.date    = '2024-12-19'
+  s.version = '1.3.2'
+  s.date    = '2025-04-25'
   s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
   s.email   = 'todd.fisher@gmail.com'
   s.extra_rdoc_files = ['LICENSE', 'README.md']
@@ -15,7 +15,8 @@ Gem::Specification.new do |s|
                 if s.extra_rdoc_files.include?(basename)
                   basename
                 else
-                  file.sub("../..", "ext")
+                  file.sub("../..", "ext/sources")
+                      .sub("../javascript", "ext/sources/bindings/javascript")
                 end
               }
 

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

@@ -19,6 +19,12 @@ const whisperParamsMock = {
   no_timestamps: false,
   audio_ctx: 0,
   max_len: 0,
+  prompt: "",
+  print_progress: false,
+  progress_callback: (progress) => {
+    console.log(`Progress: ${progress}`);
+  },
+  max_context: -1
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -368,6 +368,12 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   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>();
   
+  // Add support for max_context
+  int32_t max_context = -1;
+  if (whisper_params.Has("max_context") && whisper_params.Get("max_context").IsNumber()) {
+    max_context = whisper_params.Get("max_context").As<Napi::Number>();
+  }
+  
   // support prompt
   std::string prompt = "";
   if (whisper_params.Has("prompt") && whisper_params.Get("prompt").IsString()) {
@@ -407,6 +413,7 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   params.pcmf32 = pcmf32_vec;
   params.comma_in_time = comma_in_time;
   params.max_len = max_len;
+  params.max_context = max_context;
   params.print_progress = print_progress;
   params.prompt = prompt;
 

examples/ffmpeg-transcode.cpp

@@ -194,7 +194,7 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 	AVIOContext *avio_ctx;
 	AVStream *stream;
 	AVCodecContext *codec;
-	AVPacket packet;
+	AVPacket *packet;
 	AVFrame *frame;
 	struct SwrContext *swr;
 	u8 *avio_ctx_buffer;
@@ -249,6 +249,20 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 	/* prepare resampler */
 	swr = swr_alloc();
 
+#if LIBAVCODEC_VERSION_MAJOR > 60
+	AVChannelLayout in_ch_layout = codec->ch_layout;
+	AVChannelLayout out_ch_layout = AV_CHANNEL_LAYOUT_MONO;
+
+	/* Set the source audio layout as-is */
+	av_opt_set_chlayout(swr, "in_chlayout", &in_ch_layout, 0);
+	av_opt_set_int(swr, "in_sample_rate", codec->sample_rate, 0);
+	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
+
+	/* Convert it into 16khz Mono */
+	av_opt_set_chlayout(swr, "out_chlayout", &out_ch_layout, 0);
+	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
+	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
+#else
 	av_opt_set_int(swr, "in_channel_count", codec->channels, 0);
 	av_opt_set_int(swr, "out_channel_count", 1, 0);
 	av_opt_set_int(swr, "in_channel_layout", codec->channel_layout, 0);
@@ -257,6 +271,7 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
 	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
 	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
+#endif
 
 	swr_init(swr);
 	if (!swr_is_initialized(swr)) {
@@ -264,7 +279,11 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 		return -1;
 	}
 
-	av_init_packet(&packet);
+	packet=av_packet_alloc();
+	if (!packet) {
+		LOG("Error allocating the packet\n");
+		return -1;
+	}
 	frame = av_frame_alloc();
 	if (!frame) {
         LOG("Error allocating the frame\n");
@@ -274,8 +293,8 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 	/* iterate through frames */
 	*data = NULL;
 	*size = 0;
-	while (av_read_frame(fmt_ctx, &packet) >= 0) {
-		avcodec_send_packet(codec, &packet);
+	while (av_read_frame(fmt_ctx, packet) >= 0) {
+		avcodec_send_packet(codec, packet);
 
 		err = avcodec_receive_frame(codec, frame);
 		if (err == AVERROR(EAGAIN))
@@ -286,10 +305,11 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
 	/* Flush any remaining conversion buffers... */
 	convert_frame(swr, codec, frame, data, size, true);
 
+	av_packet_free(&packet);
 	av_frame_free(&frame);
 	swr_free(&swr);
     //avio_context_free(); // todo?
-	avcodec_close(codec);
+	avcodec_free_context(&codec);
 	avformat_close_input(&fmt_ctx);
 	avformat_free_context(fmt_ctx);
 

examples/server/server.cpp

@@ -79,6 +79,7 @@ struct whisper_params {
     bool use_gpu         = true;
     bool flash_attn      = false;
     bool suppress_nst    = false;
+    bool no_context      = false;
 
     std::string language        = "en";
     std::string prompt          = "";
@@ -140,6 +141,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server\n", sparams.ffmpeg_converter ? "true" : "false");
     fprintf(stderr, "  -sns,      --suppress-nst      [%-7s] suppress non-speech tokens\n", params.suppress_nst ? "true" : "false");
     fprintf(stderr, "  -nth N,    --no-speech-thold N [%-7.2f] no speech threshold\n",   params.no_speech_thold);
+    fprintf(stderr, "  -nc,       --no-context        [%-7s] do not use previous audio context\n", params.no_context ? "true" : "false");
     fprintf(stderr, "\n");
 }
 
@@ -186,6 +188,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
         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; }
 
         // server params
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
@@ -506,6 +509,10 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.suppress_nst = parse_str_to_bool(req.get_file_value("suppress_nst").content);
     }
+    if (req.has_file("no_context"))
+    {
+        params.no_context = parse_str_to_bool(req.get_file_value("no_context").content);
+    }
 }
 
 }  // namespace
@@ -818,6 +825,7 @@ int main(int argc, char ** argv) {
 
             wparams.no_timestamps    = params.no_timestamps;
             wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
+            wparams.no_context       = params.no_context;
 
             wparams.suppress_nst     = params.suppress_nst;
 
@@ -918,14 +926,26 @@ int main(int argc, char ** argv) {
             res.set_content(ss.str(), "text/vtt");
         } else if (params.response_format == vjson_format) {
             /* try to match openai/whisper's Python format */
-            std::string results = output_str(ctx, params, pcmf32s);
+            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()}
+                {"segments", json::array()},
+                {"detected_language", whisper_lang_str_full(detected_lang_id)},
+                {"detected_language_probability", lang_probs[detected_lang_id]},
+                {"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);
             for (int i = 0; i < n_segments; ++i)
             {

examples/talk-llama/CMakeLists.txt

@@ -12,9 +12,12 @@ if (WHISPER_SDL2)
         llama-context.cpp
         llama-cparams.cpp
         llama-grammar.cpp
+        llama-graph.cpp
         llama-hparams.cpp
         llama-impl.cpp
+        llama-io.cpp
         llama-kv-cache.cpp
+        llama-memory.cpp
         llama-mmap.cpp
         llama-model-loader.cpp
         llama-model.cpp

examples/talk-llama/llama-adapter.cpp

@@ -4,14 +4,13 @@
 #include "llama-mmap.h"
 #include "llama-model.h"
 
-#include <algorithm>
 #include <map>
 #include <cassert>
 #include <stdexcept>
 
 // vec
 
-struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
+ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
         return nullptr;
     }
@@ -19,7 +18,7 @@ struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     return tensors[il];
 }
 
-struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const {
     ggml_tensor * layer_dir = tensor_for(il);
     if (layer_dir != nullptr) {
         cur = ggml_add(ctx, cur, layer_dir);
@@ -40,7 +39,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -91,7 +90,7 @@ bool llama_adapter_cvec::init(const llama_model & model) {
     return true;
 }
 
-int32_t llama_adapter_cvec::apply(
+bool llama_adapter_cvec::apply(
         const llama_model & model,
         const float * data,
         size_t len,
@@ -104,17 +103,17 @@ int32_t llama_adapter_cvec::apply(
         // disable the current control vector (but leave allocated for later)
         layer_start = -1;
         layer_end   = -1;
-        return 0;
+        return true;
     }
 
     if (n_embd != (int) hparams.n_embd) {
         LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
-        return 1;
+        return false;
     }
 
     if (tensors.empty()) {
         if (!init(model)) {
-            return 1;
+            return false;
         }
     }
 
@@ -130,12 +129,12 @@ int32_t llama_adapter_cvec::apply(
         }
     }
 
-    return 0;
+    return true;
 }
 
 // lora
 
-llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) {
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     const std::string name(w->name);
 
     const auto pos = ab_map.find(name);
@@ -146,11 +145,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor *
     return nullptr;
 }
 
-static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
     ggml_context * ctx_init;
-    struct gguf_init_params meta_gguf_params = {
+    gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
         /* .ctx      = */ &ctx_init,
     };
@@ -201,7 +200,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             // add a new context
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -248,6 +247,26 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         }
     }
 
+    // get extra buffer types of the CPU
+    // TODO: a more general solution for non-CPU extra buft should be imlpemented in the future
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/12593#pullrequestreview-2718659948
+    std::vector<ggml_backend_buffer_type_t> buft_extra;
+    {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_extra.emplace_back(*extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    }
+
     // add tensors
     for (auto & it : ab_map) {
         const std::string & name = it.first;
@@ -264,7 +283,23 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
             throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
         }
 
-        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer);
+
+        // do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case
+        for (auto & ex : buft_extra) {
+            if (ex == buft) {
+                LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+                auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                buft = ggml_backend_dev_buffer_type(cpu_dev);
+
+                break;
+            }
+        }
+
+        LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+        ggml_context * dev_ctx = ctx_for_buft(buft);
         // validate tensor shape
         if (is_token_embd) {
             // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
@@ -281,8 +316,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
         }
 
         // save tensor to adapter
-        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
-        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
         ggml_set_name(tensor_a, w.a->name);
         ggml_set_name(tensor_b, w.b->name);
         adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
@@ -308,7 +343,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     {
         llama_file gguf_file(path_lora, "rb");
         std::vector<uint8_t> read_buf;
-        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+        auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) {
             size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
             size_t size = ggml_nbytes(orig);
             read_buf.resize(size);
@@ -327,8 +362,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
-struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) {
-    struct llama_adapter_lora * adapter = new llama_adapter_lora();
+llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
+    llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
         llama_adapter_lora_init_impl(*model, path_lora, *adapter);
@@ -342,6 +377,6 @@ struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model,
     return nullptr;
 }
 
-void llama_adapter_lora_free(struct llama_adapter_lora * adapter) {
+void llama_adapter_lora_free(llama_adapter_lora * adapter) {
     delete adapter;
 }

examples/talk-llama/llama-adapter.h

@@ -15,11 +15,11 @@
 //
 
 struct llama_adapter_cvec {
-    struct ggml_tensor * tensor_for(int il) const;
+    ggml_tensor * tensor_for(int il) const;
 
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+    ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const;
 
-    int32_t apply(
+    bool apply(
             const llama_model & model,
             const float * data,
             size_t len,
@@ -36,7 +36,7 @@ private:
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
-    std::vector<struct ggml_tensor *> tensors; // per layer
+    std::vector<ggml_tensor *> tensors; // per layer
 };
 
 //
@@ -44,8 +44,8 @@ private:
 //
 
 struct llama_adapter_lora_weight {
-    struct ggml_tensor * a = nullptr;
-    struct ggml_tensor * b = nullptr;
+    ggml_tensor * a = nullptr;
+    ggml_tensor * b = nullptr;
 
     // get actual scale based on rank and alpha
     float get_scale(float alpha, float adapter_scale) const {
@@ -55,12 +55,12 @@ struct llama_adapter_lora_weight {
     }
 
     llama_adapter_lora_weight() = default;
-    llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+    llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
 };
 
 struct llama_adapter_lora {
     // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
+    std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
@@ -70,5 +70,7 @@ struct llama_adapter_lora {
     llama_adapter_lora() = default;
     ~llama_adapter_lora() = default;
 
-    llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
+    llama_adapter_lora_weight * get_weight(ggml_tensor * w);
 };
+
+using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

examples/talk-llama/llama-arch.cpp

@@ -6,6 +6,7 @@
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,            "llama"            },
+    { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
     { LLM_ARCH_FALCON,           "falcon"           },
     { LLM_ARCH_GROK,             "grok"             },
@@ -25,6 +26,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2,            "qwen2"            },
     { LLM_ARCH_QWEN2MOE,         "qwen2moe"         },
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
+    { LLM_ARCH_QWEN3,            "qwen3"            },
+    { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -36,6 +39,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MINICPM3,         "minicpm3"         },
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
+    { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
     { LLM_ARCH_XVERSE,           "xverse"           },
@@ -50,6 +54,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK,         "deepseek"         },
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
+    { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -58,10 +63,14 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_EXAONE,           "exaone"           },
     { 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_CHAMELEON,        "chameleon"        },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
+    { LLM_ARCH_PLM,              "plm"              },
+    { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -70,6 +79,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
     { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
     { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
     { LLM_KV_GENERAL_NAME,                 "general.name"                          },
     { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
     { LLM_KV_GENERAL_VERSION,              "general.version"                       },
@@ -108,23 +118,30 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
     { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
+    { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
 
-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
-    { LLM_KV_ATTENTION_GROUPNORM_EPS,          "%s.attention.group_norm_epsilon"     },
-    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,       "%s.attention.group_norm_groups"      },
-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
-    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
-    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
-    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
+    { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,               "%s.attention.max_alibi_bias"               },
+    { LLM_KV_ATTENTION_CLAMP_KQV,                    "%s.attention.clamp_kqv"                    },
+    { LLM_KV_ATTENTION_KEY_LENGTH,                   "%s.attention.key_length"                   },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,                 "%s.attention.value_length"                 },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,                "%s.attention.layer_norm_epsilon"           },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            "%s.attention.layer_norm_rms_epsilon"       },
+    { LLM_KV_ATTENTION_GROUPNORM_EPS,                "%s.attention.group_norm_epsilon"           },
+    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,             "%s.attention.group_norm_groups"            },
+    { LLM_KV_ATTENTION_CAUSAL,                       "%s.attention.causal"                       },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,                  "%s.attention.q_lora_rank"                  },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,                 "%s.attention.kv_lora_rank"                 },
+    { LLM_KV_ATTENTION_DECAY_LORA_RANK,              "%s.attention.decay_lora_rank"              },
+    { LLM_KV_ATTENTION_ICLR_LORA_RANK,               "%s.attention.iclr_lora_rank"               },
+    { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
+    { LLM_KV_ATTENTION_GATE_LORA_RANK,               "%s.attention.gate_lora_rank"               },
+    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
+    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+    { 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_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -223,6 +240,35 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_LLAMA4,
+        {
+            { 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_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_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { 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_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { 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_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_DECI,
         {
@@ -554,6 +600,45 @@ 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_QWEN3,
+        {
+            { 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_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_QWEN3MOE,
+        {
+            { 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_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" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -766,6 +851,27 @@ 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_GEMMA3,
+        {
+            { 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_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_ARCH_STARCODER2,
         {
@@ -999,6 +1105,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
             { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
             { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_K_B,           "blk.%d.attn_k_b" },
+            { LLM_TENSOR_ATTN_V_B,           "blk.%d.attn_v_b" },
             { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
@@ -1015,6 +1123,22 @@ 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_PLM,
+        {
+            { 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_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
+            { LLM_TENSOR_ATTN_KV_A_NORM,     "blk.%d.attn_kv_a_norm" },
+            { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { 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_ARCH_CHATGLM,
         {
@@ -1033,6 +1157,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
         },
     },
+    {
+        LLM_ARCH_GLM4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { 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_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -1217,6 +1360,74 @@ 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_RWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_NORM_2,               "blk.%d.attn_norm_2" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_CHANNEL_MIX_LERP_K,        "blk.%d.channel_mix_lerp_k" },
+            { LLM_TENSOR_CHANNEL_MIX_KEY,           "blk.%d.channel_mix_key" },
+            { LLM_TENSOR_CHANNEL_MIX_VALUE,         "blk.%d.channel_mix_value" },
+        },
+    },
+    {
+        LLM_ARCH_ARWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_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_GRANITE,
         {
@@ -1296,6 +1507,29 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_POS_NET_ATTN_OUT,  "posnet.%d.attn_output" },
         },
     },
+    {
+        LLM_ARCH_BAILINGMOE,
+        {
+            { 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_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_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_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { 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_UNKNOWN,
         {
@@ -1333,23 +1567,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1376,6 +1595,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W1,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W2,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_KEY,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1394,6 +1619,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {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}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_R,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_A,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_R_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_W,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_K,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_V,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
@@ -1401,6 +1629,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LERP_G,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_FUSED,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_DECAY,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_W0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_A0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_V0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_FIRST,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
     {LLM_TENSOR_ATTN_NORM,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_NORM_2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

examples/talk-llama/llama-arch.h

@@ -10,6 +10,7 @@
 
 enum llm_arch {
     LLM_ARCH_LLAMA,
+    LLM_ARCH_LLAMA4,
     LLM_ARCH_DECI,
     LLM_ARCH_FALCON,
     LLM_ARCH_BAICHUAN,
@@ -29,6 +30,8 @@ enum llm_arch {
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
     LLM_ARCH_QWEN2VL,
+    LLM_ARCH_QWEN3,
+    LLM_ARCH_QWEN3MOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -40,6 +43,7 @@ enum llm_arch {
     LLM_ARCH_MINICPM3,
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
+    LLM_ARCH_GEMMA3,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
@@ -54,6 +58,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -62,10 +67,14 @@ enum llm_arch {
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
+    LLM_ARCH_RWKV7,
+    LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_WAVTOKENIZER_DEC,
+    LLM_ARCH_PLM,
+    LLM_ARCH_BAILINGMOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -74,6 +83,7 @@ enum llm_kv {
     LLM_KV_GENERAL_ARCHITECTURE,
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_FILE_TYPE,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -112,6 +122,7 @@ enum llm_kv {
     LLM_KV_RESIDUAL_SCALE,
     LLM_KV_EMBEDDING_SCALE,
     LLM_KV_TOKEN_SHIFT_COUNT,
+    LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -126,9 +137,15 @@ enum llm_kv {
     LLM_KV_ATTENTION_CAUSAL,
     LLM_KV_ATTENTION_Q_LORA_RANK,
     LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_DECAY_LORA_RANK,
+    LLM_KV_ATTENTION_ICLR_LORA_RANK,
+    LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
+    LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -242,6 +259,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -249,8 +268,20 @@ enum llm_tensor {
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_A0,
+    LLM_TENSOR_TIME_MIX_A1,
+    LLM_TENSOR_TIME_MIX_A2,
+    LLM_TENSOR_TIME_MIX_V0,
+    LLM_TENSOR_TIME_MIX_V1,
+    LLM_TENSOR_TIME_MIX_V2,
+    LLM_TENSOR_TIME_MIX_G1,
+    LLM_TENSOR_TIME_MIX_G2,
+    LLM_TENSOR_TIME_MIX_K_K,
+    LLM_TENSOR_TIME_MIX_K_A,
+    LLM_TENSOR_TIME_MIX_R_K,
     LLM_TENSOR_TIME_MIX_LERP_X,
     LLM_TENSOR_TIME_MIX_LERP_W,
     LLM_TENSOR_TIME_MIX_LERP_K,
@@ -277,6 +308,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,

examples/talk-llama/llama-batch.h

@@ -42,9 +42,9 @@ struct llama_sbatch {
     bool logits_all; // TODO: remove once lctx.logits_all is removed too
 
     // sorted indices into the batch
-    std::vector<size_t> ids;
+    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<size_t> out_ids;
+    std::vector<int64_t> out_ids;
     std::vector<llama_sbatch_seq> seq;
 
     const llama_batch * batch = nullptr;

examples/talk-llama/llama-chat.cpp

@@ -4,6 +4,7 @@
 
 #include <map>
 #include <sstream>
+#include <algorithm>
 
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)
@@ -58,6 +59,10 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+    { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
+    { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
+    { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -77,7 +82,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     if (tmpl_contains("<|im_start|>")) {
         return tmpl_contains("<|im_sep|>")
             ? LLM_CHAT_TEMPLATE_PHI_4
-            : LLM_CHAT_TEMPLATE_CHATML;
+            : tmpl_contains("<end_of_utterance>")
+                ? LLM_CHAT_TEMPLATE_SMOLVLM // SmolVLM uses <|im_start|> as BOS, but it is NOT chatml
+                : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -117,6 +124,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_PHI_3;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
+    } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
+        return LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -167,6 +176,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GIGACHAT;
     } else if (tmpl_contains("<|role_start|>")) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
+    } else if (tmpl_contains(" Ассистент:")) {
+        return LLM_CHAT_TEMPLATE_YANDEX;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
+        return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA4;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -566,6 +581,66 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|role_start|>assistant<|role_end|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
+        // Yandex template ("\n\n" is defined as EOT token)
+
+        ss << "<s>";
+
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << " Пользователь: " << chat[i]->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << " Ассистент: " << chat[i]->content << "\n\n";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << " Ассистент:[SEP]";
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
+        // Bailing (Ling) template
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content;
+        }
+
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA4) {
+        // Llama 4
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|header_start|>" << role << "<|header_end|>\n\n" << trim(message->content) << "<|eot|>";
+        }
+        if (add_ass) {
+            ss << "<|header_start|>assistant<|header_end|>\n\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SMOLVLM) {
+        // SmolVLM
+        ss << "<|im_start|>"; // uses <|im_start|> as BOS, but the actual content is NOT chatml
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "<end_of_utterance>\n";
+            } else {
+                ss << "Assistant: " << message->content << "<end_of_utterance>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
     } else {
         // template not supported
         return -1;
@@ -584,4 +659,3 @@ int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
     }
     return (int32_t) LLM_CHAT_TEMPLATES.size();
 }
-

examples/talk-llama/llama-chat.h

@@ -38,6 +38,10 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_YANDEX,
+    LLM_CHAT_TEMPLATE_BAILING,
+    LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

examples/talk-llama/llama-context.h

@@ -3,66 +3,213 @@
 #include "llama.h"
 #include "llama-batch.h"
 #include "llama-cparams.h"
-#include "llama-model.h"
-#include "llama-kv-cache.h"
+#include "llama-graph.h"
 #include "llama-adapter.h"
 
 #include "ggml-cpp.h"
 
 #include <map>
-#include <unordered_map>
 #include <vector>
-#include <set>
+
+struct llama_model;
+struct llama_kv_cache;
+
+class llama_io_read_i;
+class llama_io_write_i;
 
 struct llama_context {
-    llama_context(const llama_model & model)
-        : model(model)
-        , t_start_us(model.t_start_us)
-        , t_load_us(model.t_load_us) {}
+    // init scheduler and compute buffers, reserve worst-case graphs
+    llama_context(
+            const llama_model & model,
+                  llama_context_params params);
 
-    const struct llama_model & model;
+    ~llama_context();
 
-    struct llama_cparams      cparams;
-    struct llama_sbatch       sbatch;  // TODO: revisit if needed
-    struct llama_kv_cache     kv_self;
-    struct llama_adapter_cvec cvec;
+    void synchronize();
 
-    std::unordered_map<struct llama_adapter_lora *, float> lora;
+    const llama_model & get_model() const;
 
-    std::vector<ggml_backend_ptr> backends;
-    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+    uint32_t n_ctx()         const;
+    uint32_t n_ctx_per_seq() const;
+    uint32_t n_batch()       const;
+    uint32_t n_ubatch()      const;
+    uint32_t n_seq_max()     const;
 
-    ggml_backend_t backend_cpu = nullptr;
+    uint32_t n_threads()       const;
+    uint32_t n_threads_batch() const;
 
-    ggml_threadpool_t threadpool       = nullptr;
-    ggml_threadpool_t threadpool_batch = nullptr;
+          llama_kv_cache * get_kv_self();
+    const llama_kv_cache * get_kv_self() const;
 
-    bool has_evaluated_once = false;
+    void kv_self_update();
 
-    mutable int64_t t_start_us;
-    mutable int64_t t_load_us;
-    mutable int64_t t_p_eval_us = 0;
-    mutable int64_t t_eval_us   = 0;
+    enum llama_pooling_type pooling_type() const;
 
-    mutable int64_t t_compute_start_us = 0;
-    mutable int64_t n_queued_tokens = 0;
+    float * get_logits();
+    float * get_logits_ith(int32_t i);
 
-    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
+    float * get_embeddings();
+    float * get_embeddings_ith(int32_t i);
+    float * get_embeddings_seq(llama_seq_id seq_id);
 
-    // host buffer for the model output (logits and embeddings)
-    ggml_backend_buffer_ptr buf_output;
+    void attach_threadpool(
+            ggml_threadpool_t threadpool,
+            ggml_threadpool_t threadpool_batch);
+
+    void detach_threadpool();
+
+    void set_n_threads(int32_t n_threads, int32_t n_threads_batch);
+
+    void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
+
+    void set_embeddings (bool value);
+    void set_causal_attn(bool value);
+    void set_warmup(bool value);
+
+    void set_adapter_lora(
+            llama_adapter_lora * adapter,
+            float scale);
+
+    bool rm_adapter_lora(
+            llama_adapter_lora * adapter);
+
+    void clear_adapter_lora();
+
+    bool apply_adapter_cvec(
+            const float * data,
+                 size_t   len,
+                int32_t   n_embd,
+                int32_t   il_start,
+                int32_t   il_end);
+
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
+
+    //
+    // state save/load
+    //
+
+    size_t state_get_size();
+    size_t state_get_data(      uint8_t * dst, size_t size);
+    size_t state_set_data(const uint8_t * src, size_t size);
+
+    size_t state_seq_get_size(llama_seq_id seq_id);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+
+    bool state_load_file(
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    bool state_save_file(
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    size_t state_seq_load_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    size_t state_seq_save_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    //
+    // perf
+    //
+
+    llama_perf_context_data perf_get_data() const;
+    void perf_reset();
+
+private:
+    //
+    // output
+    //
+
+    // Make sure enough space is available for outputs.
+    // Returns max number of outputs for which space was reserved.
+    int32_t output_reserve(int32_t n_outputs);
+
+    // make the outputs have the same order they had in the user-provided batch
+    // TODO: maybe remove this
+    void output_reorder();
+
+    //
+    // graph
+    //
+
+    int32_t graph_max_nodes() const;
+
+    // zero-out inputs and create the ctx_compute for the compute graph
+    ggml_cgraph * graph_init();
+
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
+    // returns the result of ggml_backend_sched_graph_compute_async execution
+    ggml_status graph_compute(
+            ggml_cgraph * gf,
+                   bool   batched);
+
+    llm_graph_cb graph_get_cb() const;
+
+    // used by kv_self_update()
+    ggml_tensor * build_rope_shift(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * shift,
+        ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale,
+        ggml_backend_buffer * bbuf) const;
+
+    llm_graph_result_ptr build_kv_self_shift(
+            ggml_context * ctx0,
+            ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_kv_self_defrag(
+            ggml_context * ctx0,
+            ggml_cgraph * gf) const;
+
+    // TODO: read/write lora adapters and cvec
+    size_t state_write_data(llama_io_write_i & io);
+    size_t state_read_data (llama_io_read_i  & io);
+
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+
+    //
+    // members
+    //
+
+    const llama_model & model;
+
+    llama_cparams       cparams;
+    llama_adapter_cvec  cvec;
+    llama_adapter_loras loras;
+    llama_sbatch        sbatch;
+
+    llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
+
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
 
-    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
-    size_t  output_size = 0; // capacity (of tokens positions) for the output buffers
-    int32_t n_outputs   = 0; // number of actually-used outputs in the current ubatch or last logical batch
-
-    bool logits_all = false;
-
     // embeddings output (2-dimensional array: [n_outputs][n_embd])
     // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
     size_t  embd_size = 0; // capacity (of floats) for embeddings
@@ -72,57 +219,47 @@ struct llama_context {
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    // whether we are computing encoder output or decoder output
-    bool is_encoding = false;
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
 
-    // TODO: find a better way to accommodate mutli-dimension position encoding methods
-    // number of position id each token get, 1 for each token in most cases.
-    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
-    int n_pos_per_token = 1;
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 
-    // output of the encoder part of the encoder-decoder models
-    std::vector<float> embd_enc;
-    std::vector<std::set<llama_seq_id>> seq_ids_enc;
-
-    // memory buffers used to evaluate the model
-    std::vector<uint8_t> buf_compute_meta;
     ggml_backend_sched_ptr sched;
 
+    ggml_backend_t backend_cpu = nullptr;
+    std::vector<ggml_backend_ptr> backends;
+
+    ggml_context_ptr ctx_compute;
+
+    ggml_threadpool_t threadpool       = nullptr;
+    ggml_threadpool_t threadpool_batch = nullptr;
+
     ggml_abort_callback abort_callback      = nullptr;
     void *              abort_callback_data = nullptr;
 
-    // input tensors
-    struct ggml_tensor * inp_tokens;        // I32 [n_batch]
-    struct ggml_tensor * inp_embd;          // F32 [n_embd, n_batch]
-    struct ggml_tensor * inp_pos;           // I32 [n_batch]
-    struct ggml_tensor * inp_out_ids;       // I32 [n_outputs]
-    struct ggml_tensor * inp_KQ_mask;       // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_KQ_mask_swa;   // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_K_shift;       // I32 [kv_size]
-    struct ggml_tensor * inp_mean;          // F32 [n_batch, n_batch]
-    struct ggml_tensor * inp_cls;           // I32 [n_batch]
-    struct ggml_tensor * inp_s_copy;        // I32 [kv_size]
-    struct ggml_tensor * inp_s_mask;        // F32 [1, n_kv]
-    struct ggml_tensor * inp_s_seq;         // I32 [n_kv, n_batch]
-    struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
-    struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
-    struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
+    // buffer types used for the compute buffer of each backend
+    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;
+
+    // host buffer for the model output (logits and embeddings)
+    ggml_backend_buffer_ptr buf_output;
+
+    bool has_evaluated_once = false;
+
+    // perf
+    mutable int64_t t_start_us  = 0;
+    mutable int64_t t_load_us   = 0;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens    = 0;
+
+    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
 };
-
-// TODO: make these methods of llama_context
-void llama_set_k_shift(struct llama_context & lctx);
-
-void llama_set_s_copy(struct llama_context & lctx);
-
-void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch);
-
-// Make sure enough space is available for outputs.
-// Returns max number of outputs for which space was reserved.
-size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs);
-
-// make the outputs have the same order they had in the user-provided batch
-void llama_output_reorder(struct llama_context & ctx);
-
-// For internal test use
-// TODO: remove
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

examples/talk-llama/llama-cparams.h

@@ -29,6 +29,7 @@ struct llama_cparams {
     bool offload_kqv;
     bool flash_attn;
     bool no_perf;
+    bool warmup;
 
     enum llama_pooling_type pooling_type;
 

examples/talk-llama/llama-grammar.cpp

@@ -345,194 +345,194 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-        auto handle_repetitions = [&](int min_times, int max_times) {
+    auto handle_repetitions = [&](int min_times, int max_times) {
 
-            if (last_sym_start == rule.size()) {
-                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
-            }
+        if (last_sym_start == rule.size()) {
+            throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
+        }
 
-            // apply transformation to previous symbol (last_sym_start to end) according to
-            // the following rewrite rules:
-            // S{m,n} --> S S S (m times) S'(n-m)
-            //            S'(x)   ::= S S'(x-1) |
-            //            (... n-m definitions of these S' rules ...)
-            //            S'(1)   ::= S |
-            // S{m,} -->  S S S (m times) S'
-            //            S'     ::= S S' |
-            // S*     --> S{0,}
-            //        --> S'     ::= S S' |
-            // S+     --> S{1,}
-            //        --> S S'
-            //            S'     ::= S S' |
-            // S?     --> S{0,1}
-            //        --> S'
-            //            S'     ::= S |
+        // apply transformation to previous symbol (last_sym_start to end) according to
+        // the following rewrite rules:
+        // S{m,n} --> S S S (m times) S'(n-m)
+        //            S'(x)   ::= S S'(x-1) |
+        //            (... n-m definitions of these S' rules ...)
+        //            S'(1)   ::= S |
+        // S{m,} -->  S S S (m times) S'
+        //            S'     ::= S S' |
+        // S*     --> S{0,}
+        //        --> S'     ::= S S' |
+        // S+     --> S{1,}
+        //        --> S S'
+        //            S'     ::= S S' |
+        // S?     --> S{0,1}
+        //        --> S'
+        //            S'     ::= S |
 
-            llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end());
-            if (min_times == 0) {
-                rule.resize(last_sym_start);
-            } else {
-                // Repeat the previous elements (min_times - 1) times
-                for (int i = 1; i < min_times; i++) {
-                    rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
-                }
-            }
-
-            uint32_t last_rec_rule_id = 0;
-            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
-
-            llama_grammar_rule rec_rule(prev_rule);
-            for (int i = 0; i < n_opt; i++) {
-                rec_rule.resize(prev_rule.size());
-                uint32_t rec_rule_id = generate_symbol_id( rule_name);
-                if (i > 0 || max_times < 0) {
-                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
-                }
-                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
-                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
-                add_rule( rec_rule_id, rec_rule);
-                last_rec_rule_id = rec_rule_id;
-            }
-            if (n_opt > 0) {
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
-            }
-        };
-
-        while (*pos) {
-            if (*pos == '"') { // literal string
-                pos++;
-                last_sym_start = rule.size();
-                while (*pos != '"') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '[') { // char range(s)
-                pos++;
-                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
-                if (*pos == '^') {
-                    pos++;
-                    start_type = LLAMA_GRETYPE_CHAR_NOT;
-                }
-                last_sym_start = rule.size();
-                while (*pos != ']') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    enum llama_gretype type = last_sym_start < rule.size()
-                        ? LLAMA_GRETYPE_CHAR_ALT
-                        : start_type;
-
-                    rule.push_back({type, char_pair.first});
-                    if (pos[0] == '-' && pos[1] != ']') {
-                        if (!pos[1]) {
-                            throw std::runtime_error("unexpected end of input");
-                        }
-                        auto endchar_pair = parse_char(pos + 1);
-                             pos          = endchar_pair.second;
-                        rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
-                    }
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (is_word_char(*pos)) { // rule reference
-                const char * name_end    = parse_name(pos);
-                uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
-                pos = parse_space(name_end, is_nested);
-                last_sym_start = rule.size();
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
-            } else if (*pos == '(') { // grouping
-                // parse nested alternates into synthesized rule
-                pos = parse_space(pos + 1, true);
-                uint32_t sub_rule_id = generate_symbol_id(rule_name);
-                pos = parse_alternates(pos, rule_name, sub_rule_id, true);
-                last_sym_start = rule.size();
-                // output reference to synthesized rule
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
-                if (*pos != ')') {
-                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '.') { // any char
-                last_sym_start = rule.size();
-                rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, -1);
-            } else if (*pos == '+') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(1, -1);
-            } else if (*pos == '?') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, 1);
-            } else if (*pos == '{') {
-                pos = parse_space(pos + 1, is_nested);
-
-                if (!is_digit_char(*pos)) {
-                    throw std::runtime_error(std::string("expecting an int at ") + pos);
-                }
-                const char * int_end = parse_int(pos);
-                int min_times = std::stoul(std::string(pos, int_end - pos));
-                pos = parse_space(int_end, is_nested);
-
-                int max_times = -1;
-
-                if (*pos == '}') {
-                    max_times = min_times;
-                    pos = parse_space(pos + 1, is_nested);
-                } else if (*pos == ',') {
-                    pos = parse_space(pos + 1, is_nested);
-
-                    if (is_digit_char(*pos)) {
-                        const char * int_end = parse_int(pos);
-                        max_times = std::stoul(std::string(pos, int_end - pos));
-                        pos = parse_space(int_end, is_nested);
-                    }
-
-                    if (*pos != '}') {
-                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
-                    }
-                    pos = parse_space(pos + 1, is_nested);
-                } else {
-                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
-                }
-                handle_repetitions(min_times, max_times);
-            } else {
-                break;
+        llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end());
+        if (min_times == 0) {
+            rule.resize(last_sym_start);
+        } else {
+            // Repeat the previous elements (min_times - 1) times
+            for (int i = 1; i < min_times; i++) {
+                rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
-        return pos;
+
+        uint32_t last_rec_rule_id = 0;
+        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+
+        llama_grammar_rule rec_rule(prev_rule);
+        for (int i = 0; i < n_opt; i++) {
+            rec_rule.resize(prev_rule.size());
+            uint32_t rec_rule_id = generate_symbol_id( rule_name);
+            if (i > 0 || max_times < 0) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            }
+            rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            rec_rule.push_back({LLAMA_GRETYPE_END, 0});
+            add_rule( rec_rule_id, rec_rule);
+            last_rec_rule_id = rec_rule_id;
+        }
+        if (n_opt > 0) {
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
+        }
+    };
+
+    while (*pos) {
+        if (*pos == '"') { // literal string
+            pos++;
+            last_sym_start = rule.size();
+            while (*pos != '"') {
+                if (!*pos) {
+                    throw std::runtime_error("unexpected end of input");
+                }
+                auto char_pair = parse_char(pos);
+                     pos       = char_pair.second;
+                rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '[') { // char range(s)
+            pos++;
+            enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+            if (*pos == '^') {
+                pos++;
+                start_type = LLAMA_GRETYPE_CHAR_NOT;
+            }
+            last_sym_start = rule.size();
+            while (*pos != ']') {
+                if (!*pos) {
+                    throw std::runtime_error("unexpected end of input");
+                }
+                auto char_pair = parse_char(pos);
+                     pos       = char_pair.second;
+                enum llama_gretype type = last_sym_start < rule.size()
+                    ? LLAMA_GRETYPE_CHAR_ALT
+                    : start_type;
+
+                rule.push_back({type, char_pair.first});
+                if (pos[0] == '-' && pos[1] != ']') {
+                    if (!pos[1]) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
+                    auto endchar_pair = parse_char(pos + 1);
+                         pos          = endchar_pair.second;
+                    rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                }
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (is_word_char(*pos)) { // rule reference
+            const char * name_end    = parse_name(pos);
+            uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
+            pos = parse_space(name_end, is_nested);
+            last_sym_start = rule.size();
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+        } else if (*pos == '(') { // grouping
+            // parse nested alternates into synthesized rule
+            pos = parse_space(pos + 1, true);
+            uint32_t sub_rule_id = generate_symbol_id(rule_name);
+            pos = parse_alternates(pos, rule_name, sub_rule_id, true);
+            last_sym_start = rule.size();
+            // output reference to synthesized rule
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+            if (*pos != ')') {
+                throw std::runtime_error(std::string("expecting ')' at ") + pos);
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '.') { // any char
+            last_sym_start = rule.size();
+            rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '*') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(0, -1);
+        } else if (*pos == '+') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(1, -1);
+        } else if (*pos == '?') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(0, 1);
+        } else if (*pos == '{') {
+            pos = parse_space(pos + 1, is_nested);
+
+            if (!is_digit_char(*pos)) {
+                throw std::runtime_error(std::string("expecting an int at ") + pos);
+            }
+            const char * int_end = parse_int(pos);
+            int min_times = std::stoul(std::string(pos, int_end - pos));
+            pos = parse_space(int_end, is_nested);
+
+            int max_times = -1;
+
+            if (*pos == '}') {
+                max_times = min_times;
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == ',') {
+                pos = parse_space(pos + 1, is_nested);
+
+                if (is_digit_char(*pos)) {
+                    const char * int_end = parse_int(pos);
+                    max_times = std::stoul(std::string(pos, int_end - pos));
+                    pos = parse_space(int_end, is_nested);
+                }
+
+                if (*pos != '}') {
+                    throw std::runtime_error(std::string("expecting '}' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                throw std::runtime_error(std::string("expecting ',' at ") + pos);
+            }
+            handle_repetitions(min_times, max_times);
+        } else {
+            break;
+        }
     }
+    return pos;
+}
 
 const char * llama_grammar_parser::parse_rule(const char * src) {
-        const char * name_end = parse_name(src);
-        const char * pos      = parse_space(name_end, false);
-        size_t       name_len = name_end - src;
-        uint32_t     rule_id  = get_symbol_id(src, name_len);
-        const std::string name(src, name_len);
+    const char * name_end = parse_name(src);
+    const char * pos      = parse_space(name_end, false);
+    size_t       name_len = name_end - src;
+    uint32_t     rule_id  = get_symbol_id(src, name_len);
+    const std::string name(src, name_len);
 
-        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
-            throw std::runtime_error(std::string("expecting ::= at ") + pos);
-        }
-        pos = parse_space(pos + 3, true);
-
-        pos = parse_alternates(pos, name, rule_id, false);
-
-        if (*pos == '\r') {
-            pos += pos[1] == '\n' ? 2 : 1;
-        } else if (*pos == '\n') {
-            pos++;
-        } else if (*pos) {
-            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
-        }
-        return parse_space(pos, true);
+    if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+        throw std::runtime_error(std::string("expecting ::= at ") + pos);
     }
+    pos = parse_space(pos + 3, true);
+
+    pos = parse_alternates(pos, name, rule_id, false);
+
+    if (*pos == '\r') {
+        pos += pos[1] == '\n' ? 2 : 1;
+    } else if (*pos == '\n') {
+        pos++;
+    } else if (*pos) {
+        throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+    }
+    return parse_space(pos, true);
+}
 
 bool llama_grammar_parser::parse(const char * src) {
     try {
@@ -969,7 +969,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .awaiting_trigger = */ false,
         /* .trigger_buffer = */   "",
         /* .trigger_tokens   = */ {},
-        /* .trigger_words    = */ {},
+        /* .trigger_patterns    = */ {},
     };
 }
 
@@ -978,19 +978,15 @@ struct llama_grammar * llama_grammar_init_impl(
                       const char * grammar_str,
                       const char * grammar_root,
                               bool lazy,
-                     const char ** trigger_words,
-                            size_t num_trigger_words,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
     llama_grammar_parser parser;
 
     // if there is a grammar, parse it
-    if (!parser.parse(grammar_str)) {
-        return nullptr;
-    }
-
-    // will be empty (default) if there are parse errors
-    if (parser.rules.empty()) {
+    // rules will be empty (default) if there are parse errors
+    if (!parser.parse(grammar_str) || parser.rules.empty()) {
         fprintf(stderr, "%s: failed to parse grammar\n", __func__);
         return nullptr;
     }
@@ -1054,14 +1050,16 @@ struct llama_grammar * llama_grammar_init_impl(
     } while (true);
 
     std::vector<llama_token>    vec_trigger_tokens;
-    std::vector<std::string> vec_trigger_words;
+    std::vector<llama_grammar_trigger_pattern> vec_trigger_patterns;
     for (size_t i = 0; i < num_trigger_tokens; i++) {
         GGML_ASSERT(trigger_tokens != nullptr);
         vec_trigger_tokens.push_back(trigger_tokens[i]);
     }
-    for (size_t i = 0; i < num_trigger_words; i++) {
-        GGML_ASSERT(trigger_words != nullptr);
-        vec_trigger_words.push_back(trigger_words[i]);
+    for (size_t i = 0; i < num_trigger_patterns; i++) {
+        GGML_ASSERT(trigger_patterns != nullptr);
+        auto & trigger = vec_trigger_patterns.emplace_back();
+        trigger.pattern = trigger_patterns[i];
+        trigger.regex = std::regex(trigger.pattern);
     }
 
     // Important: vec_rules has to be moved here, not copied, because stacks contains
@@ -1076,7 +1074,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .awaiting_trigger = */ lazy,
         /* .trigger_buffer = */   "",
         std::move(vec_trigger_tokens),
-        std::move(vec_trigger_words),
+        std::move(vec_trigger_patterns),
     };
 }
 
@@ -1089,7 +1087,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
 }
 
 struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
-    llama_grammar * result = new llama_grammar {
+    auto * result = new llama_grammar {
         grammar.vocab,
         grammar.rules,
         grammar.stacks,
@@ -1098,7 +1096,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.awaiting_trigger,
         grammar.trigger_buffer,
         grammar.trigger_tokens,
-        grammar.trigger_words,
+        grammar.trigger_patterns,
     };
 
     // redirect elements in stacks to point to new rules
@@ -1173,20 +1171,22 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
             LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
             return;
         } else {
-            // TODO: consider a smarter incremental substring search algorithm (store last position to search from).
             grammar.trigger_buffer += piece;
-            for (const auto & word : grammar.trigger_words) {
-                auto pos = grammar.trigger_buffer.find(word);
-                if (pos != std::string::npos) {
+
+            std::smatch match;
+            for (const auto & trigger_pattern : grammar.trigger_patterns) {
+                if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
                     grammar.awaiting_trigger = false;
-                    auto constrained_str = grammar.trigger_buffer.substr(pos);
+                    // get from the first match to the end of the string
+                    auto constrained_str = grammar.trigger_buffer.substr(match.position(1));
+                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
                     llama_grammar_accept_str(grammar, constrained_str);
-                    LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str());
+                    LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
                     return;
                 }
             }
-            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`) (buffer: `%s`)\n", token, piece.c_str(), grammar.trigger_buffer.c_str());
+            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`)\n", token, piece.c_str());
             return;
         }
     }

examples/talk-llama/llama-grammar.h

@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <map>
+#include <regex>
 #include <string>
 #include <vector>
 
@@ -105,6 +106,11 @@ struct llama_grammar_parser {
     void print(FILE * file);
 };
 
+struct llama_grammar_trigger_pattern {
+    std::string pattern;
+    std::regex  regex;
+};
+
 struct llama_grammar {
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
@@ -116,13 +122,16 @@ struct llama_grammar {
     llama_partial_utf8 partial_utf8;
 
     // lazy grammars wait for trigger words or tokens before constraining the sampling.
-    // we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
     // (useful e.g. for tool_choice=required)
     bool                     lazy             = false;
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
-    std::vector<std::string> trigger_words;
+    std::vector<llama_grammar_trigger_pattern>
+                             trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
+                                                       // string, and the grammar will be given the string from the first match group onwards.
+
 };
 
 //
@@ -141,8 +150,8 @@ struct llama_grammar * llama_grammar_init_impl(
                       const char * grammar_str,
                       const char * grammar_root,
                               bool lazy,
-                     const char ** trigger_words,
-                            size_t num_trigger_words,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens);
 

examples/talk-llama/llama-graph.h

@@ -0,0 +1,596 @@
+#pragma once
+
+#include "llama-arch.h"
+#include "llama-hparams.h"
+#include "llama-adapter.h"
+
+#include <cstdint>
+#include <vector>
+#include <memory>
+#include <set>
+#include <functional>
+
+struct ggml_cgraph;
+struct ggml_context;
+struct ggml_tensor;
+
+struct llama_ubatch;
+struct llama_cparams;
+
+class llama_memory_i;
+class llama_kv_cache_unified;
+
+// certain models (typically multi-modal) can produce different types of graphs
+enum llm_graph_type {
+    LLM_GRAPH_TYPE_DEFAULT,
+    LLM_GRAPH_TYPE_ENCODER,
+    LLM_GRAPH_TYPE_DECODER,
+};
+
+enum llm_ffn_op_type {
+    LLM_FFN_SILU,
+    LLM_FFN_GELU,
+    LLM_FFN_RELU,
+    LLM_FFN_RELU_SQR,
+    LLM_FFN_SWIGLU,
+};
+
+enum llm_ffn_gate_type {
+    LLM_FFN_SEQ,
+    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up
+};
+
+enum llm_norm_type {
+    LLM_NORM,
+    LLM_NORM_RMS,
+    LLM_NORM_GROUP,
+};
+
+// TODO: tmp - need something better to pass the data from the encoder to the decoder
+struct llama_cross {
+    // the output embeddings from the encoder as a ggml tensor
+    // TODO: this needs more work to be correct, for now copy the embeddings data to host memory
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524
+    //ggml_tensor * t_embd = nullptr;
+
+    int64_t n_embd = 0;
+    int64_t n_enc  = 0;
+
+    // embeddings data copied to host memory (tmp)
+    std::vector<float> v_embd;
+
+    // needed to construct the cross-attention mask in the decoder
+    std::vector<std::set<llama_seq_id>> seq_ids_enc;
+};
+
+//
+// llm_graph_input
+//
+
+class llm_graph_input_i {
+public:
+    virtual ~llm_graph_input_i() = default;
+
+    virtual void set_input(const llama_ubatch * ubatch) = 0;
+};
+
+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;
+    virtual ~llm_graph_input_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * tokens = nullptr; // I32 [n_batch]
+    ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+};
+
+class llm_graph_input_pos : public llm_graph_input_i {
+public:
+    llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {}
+    virtual ~llm_graph_input_pos() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos = nullptr; // I32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+};
+
+// temperature tuning, used by llama4
+class llm_graph_input_attn_temp : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+        : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    virtual ~llm_graph_input_attn_temp() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * attn_scale = nullptr; // F32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+
+    const uint32_t n_attn_temp_floor_scale;
+    const float    f_attn_temp_scale;
+};
+
+class llm_graph_input_pos_bucket : public llm_graph_input_i {
+public:
+    llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {}
+    virtual ~llm_graph_input_pos_bucket() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
+
+    const llama_hparams & hparams;
+};
+
+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 * kv_self) : hparams(hparams), kv_self(kv_self) {}
+    virtual ~llm_graph_input_pos_bucket_kv() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_out_ids : public llm_graph_input_i {
+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) {}
+    virtual ~llm_graph_input_out_ids() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * out_ids; // I32 [n_outputs]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const int32_t n_outputs;
+};
+
+class llm_graph_input_mean : public llm_graph_input_i {
+public:
+    llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_mean() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * mean; // F32 [n_batch, n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_cls : public llm_graph_input_i {
+public:
+    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_cls() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cls; // I32 [n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_s_copy : public llm_graph_input_i {
+public:
+    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_copy() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_copy; // I32 [kv_size]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_s_mask : public llm_graph_input_i {
+public:
+    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_mask() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_mask; // F32 [1, n_kv]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_cross_embd : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_embd(
+            const llama_cross * cross) : cross(cross) {}
+    virtual ~llm_graph_input_cross_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc]
+
+    const llama_cross * cross;
+};
+
+class llm_graph_input_attn_no_cache : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) :
+        hparams(hparams),
+        cparams(cparams) {
+    }
+    ~llm_graph_input_attn_no_cache() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    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]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_kv_unified(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_unified * kv_self) :
+        hparams(hparams),
+        cparams(cparams),
+        kv_self(kv_self) {
+    }
+    ~llm_graph_input_attn_kv_unified() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    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]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_attn_cross : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {}
+    ~llm_graph_input_attn_cross() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    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]
+
+    const llama_cross * cross = nullptr;
+};
+
+//
+// llm_graph_result
+//
+
+// these objects deliver the result from the graph build process back to the llama_context
+// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their
+//   specific data, by calling the set_inputs() method
+// 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;
+
+    virtual ggml_tensor * get_logits()      = 0;
+    virtual ggml_tensor * get_embd()        = 0;
+    virtual ggml_tensor * get_embd_pooled() = 0;
+
+    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
+};
+
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
+
+
+class llm_graph_result : public llm_graph_result_i {
+public:
+    virtual ~llm_graph_result() = default;
+
+    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; }
+
+    void set_inputs(const llama_ubatch * ubatch) override {
+        for (auto & input : inputs) {
+            input->set_input(ubatch);
+        }
+    }
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
+        inputs.emplace_back(std::move(input));
+        return inputs.back().get();
+    }
+
+    // important graph nodes
+    ggml_tensor * t_logits      = nullptr;
+    ggml_tensor * t_embd        = nullptr;
+    ggml_tensor * t_embd_pooled = nullptr;
+
+    std::vector<llm_graph_input_ptr> inputs;
+};
+
+//
+// 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 * sched;
+    ggml_backend * backend_cpu;
+
+    const llama_adapter_cvec  * cvec;
+    const llama_adapter_loras * loras;
+    const llama_memory_i      * memory;
+    const llama_cross         * cross;
+
+    int32_t n_outputs;
+
+    const llm_graph_cb & cb;
+};
+
+struct llm_graph_context {
+    const llm_arch arch;
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+    const llama_ubatch  & ubatch;
+
+    const int64_t n_embd;
+    const int64_t n_layer;
+    const int64_t n_rot;
+    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
+    const int64_t n_ctx_per_seq;
+    const int64_t n_head;
+    const int64_t n_head_kv;
+    const int64_t n_embd_head_k;
+    const int64_t n_embd_k_gqa;
+    const int64_t n_embd_head_v;
+    const int64_t n_embd_v_gqa;
+    const int64_t n_expert;
+    const int64_t n_expert_used;
+
+    const float freq_base;
+    const float freq_scale;
+    const float ext_factor;
+    const float attn_factor;
+    const float beta_fast;
+    const float beta_slow;
+    const float norm_eps;
+    const float norm_rms_eps;
+
+    const int32_t n_tokens;
+    const int32_t n_outputs;
+    const int32_t n_ctx_orig; // yarn
+
+    const enum llama_pooling_type pooling_type;
+    const enum llama_rope_type    rope_type;
+
+    ggml_context * ctx0 = nullptr;
+
+    ggml_backend_sched * sched;
+
+    ggml_backend * 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_i      * memory;
+    const llama_cross         * cross;
+
+    const llm_graph_cb & cb_func;
+
+    std::unique_ptr<llm_graph_result> res;
+
+    llm_graph_context(const llm_graph_params & params);
+
+    int64_t n_pos_per_token() const;
+
+    void cb(ggml_tensor * cur, const char * name, int il) const;
+
+    //
+    // common
+    //
+
+    ggml_tensor * build_cvec(
+             ggml_tensor * cur,
+                     int   il) const;
+
+    // do mat_mul, while optionally apply lora
+    ggml_tensor * build_lora_mm(
+              ggml_tensor * w,
+              ggml_tensor * cur) const;
+
+    // do mat_mul_id, while optionally apply lora
+    ggml_tensor * build_lora_mm_id(
+              ggml_tensor * w,   // ggml_tensor * as
+              ggml_tensor * cur, // ggml_tensor * b
+              ggml_tensor * ids) const;
+
+    ggml_tensor * build_norm(
+             ggml_tensor * cur,
+             ggml_tensor * mw,
+             ggml_tensor * mb,
+           llm_norm_type   type,
+                     int   il) const;
+
+    ggml_tensor * build_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * up,
+             ggml_tensor * up_b,
+             ggml_tensor * up_s,
+             ggml_tensor * gate,
+             ggml_tensor * gate_b,
+             ggml_tensor * gate_s,
+             ggml_tensor * down,
+             ggml_tensor * down_b,
+             ggml_tensor * down_s,
+             ggml_tensor * act_scales,
+         llm_ffn_op_type   type_op,
+       llm_ffn_gate_type   type_gate,
+                     int   il) const;
+
+    ggml_tensor * build_moe_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * gate_inp,
+             ggml_tensor * up_exps,
+             ggml_tensor * gate_exps,
+             ggml_tensor * down_exps,
+             ggml_tensor * exp_probs_b,
+                 int64_t   n_expert,
+                 int64_t   n_expert_used,
+         llm_ffn_op_type   type_op,
+                    bool   norm_w,
+                    bool   scale_w,
+                   float   w_scale,
+            llama_expert_gating_func_type gating_op,
+                     int   il) const;
+
+    //
+    // inputs
+    //
+
+    ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const;
+    ggml_tensor * build_inp_pos() const;
+    ggml_tensor * build_inp_attn_scale() const;
+    ggml_tensor * build_inp_out_ids() const;
+    ggml_tensor * build_inp_mean() const;
+    ggml_tensor * build_inp_cls() const;
+    ggml_tensor * build_inp_s_copy() const;
+    ggml_tensor * build_inp_s_mask() const;
+
+    ggml_tensor * build_inp_cross_embd() const;
+    ggml_tensor * build_inp_pos_bucket_enc() const;
+    ggml_tensor * build_inp_pos_bucket_dec() const;
+    ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
+
+    //
+    // attention
+    //
+
+    ggml_tensor * build_attn_mha(
+             ggml_cgraph * gf,
+             ggml_tensor * q,     // [n_embd_head_q, n_tokens, n_head_q]
+             ggml_tensor * k,     // [n_embd_head_k, n_tokens, n_head_k]
+             ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
+             ggml_tensor * kq_b,
+             ggml_tensor * kq_mask,
+             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                    bool   v_trans,
+                   float   kq_scale) const;
+
+    llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
+
+    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]
+            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;
+
+    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+
+    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]
+            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;
+
+    llm_graph_input_attn_cross * build_attn_inp_cross() const;
+
+    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]
+            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
+    //
+
+    ggml_tensor * build_copy_mask_state(
+             ggml_cgraph * gf,
+             ggml_tensor * s,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+                 int32_t   n_state,
+                 int32_t   n_seqs) const;
+
+    ggml_tensor * build_rwkv_token_shift_load(
+             ggml_cgraph * gf,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    ggml_tensor * build_rwkv_token_shift_store(
+             ggml_tensor * token_shift,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    //
+    // pooling
+    //
+
+    void build_pooling(
+            ggml_cgraph * gf,
+            ggml_tensor * cls,
+            ggml_tensor * cls_b,
+            ggml_tensor * cls_out,
+            ggml_tensor * cls_out_b) const;
+};

examples/talk-llama/llama-hparams.cpp

@@ -69,3 +69,11 @@ uint32_t llama_hparams::n_embd_v_s() const {
     // corresponds to Mamba's ssm_states size
     return ssm_d_state * ssm_d_inner;
 }
+
+bool llama_hparams::is_swa(uint32_t il) const {
+    if (il < n_layer) {
+        return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+    }
+
+    GGML_ABORT("fatal error");
+}

examples/talk-llama/llama-hparams.h

@@ -36,12 +36,17 @@ struct llama_hparams {
     uint32_t n_layer;
     uint32_t n_rot;
     uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
     uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 0;
+
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;
@@ -75,10 +80,16 @@ struct llama_hparams {
     uint32_t time_decay_extra_dim   = 0;
     uint32_t wkv_head_size          = 0;
     uint32_t token_shift_count      = 2;
+    uint32_t n_lora_decay           = 0;
+    uint32_t n_lora_iclr            = 0;
+    uint32_t n_lora_value_res_mix   = 0;
+    uint32_t n_lora_gate            = 0;
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul;
 
@@ -105,6 +116,14 @@ struct llama_hparams {
     bool use_alibi     = false;
     bool attn_soft_cap = false;
 
+    uint32_t n_moe_layer_step        = 0;
+    bool     use_kq_norm             = true;
+    uint32_t n_attn_chunk            = 0;
+    // values below seems to be fixed on llama4
+    uint32_t n_no_rope_layer_step    = 4;
+    uint32_t n_attn_temp_floor_scale = 8192;
+    float    f_attn_temp_scale       = 0.1;
+
     // 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;
@@ -133,6 +152,8 @@ struct llama_hparams {
 
     // dimension of the recurrent state embeddings
     uint32_t n_embd_v_s() const;
+
+    bool is_swa(uint32_t il) const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

examples/talk-llama/llama-impl.h

@@ -6,13 +6,13 @@
 #include <vector>
 
 #ifdef __GNUC__
-#ifdef __MINGW32__
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
 #else
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-#else
-#define LLAMA_ATTRIBUTE_FORMAT(...)
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
 #endif
 
 //

examples/talk-llama/llama-io.cpp

@@ -0,0 +1,15 @@
+#include "llama-io.h"
+
+void llama_io_write_i::write_string(const std::string & str) {
+    uint32_t str_size = str.size();
+
+    write(&str_size,  sizeof(str_size));
+    write(str.data(), str_size);
+}
+
+void llama_io_read_i::read_string(std::string & str) {
+    uint32_t str_size;
+    read_to(&str_size, sizeof(str_size));
+
+    str.assign((const char *) read(str_size), str_size);
+}

examples/talk-llama/llama-io.h

@@ -0,0 +1,35 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+struct ggml_tensor;
+
+class llama_io_write_i {
+public:
+    llama_io_write_i() = default;
+    virtual ~llama_io_write_i() = default;
+
+    virtual void write(const void * src, size_t size) = 0;
+    virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0;
+
+    // bytes written so far
+    virtual size_t n_bytes() = 0;
+
+    void write_string(const std::string & str);
+};
+
+class llama_io_read_i {
+public:
+    llama_io_read_i() = default;
+    virtual ~llama_io_read_i() = default;
+
+    virtual const uint8_t * read(size_t size) = 0;
+    virtual void read_to(void * dst, size_t size) = 0;
+
+    // bytes read so far
+    virtual size_t n_bytes() = 0;
+
+    void read_string(std::string & str);
+};

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

@@ -1,15 +1,51 @@
 #pragma once
 
 #include "llama.h"
+#include "llama-io.h"
+#include "llama-memory.h"
 
 #include "ggml-cpp.h"
 
+#include <functional>
 #include <set>
 #include <vector>
 
+struct llama_cparams;
+struct llama_hparams;
+struct llama_ubatch;
+
+struct llama_kv_cache : public llama_memory_i {
+    using llama_memory_i::llama_memory_i;
+
+    virtual void restore() = 0; // call if batch processing fails - restores the cache state
+    virtual void commit() = 0;  // call after successful batch processing - clears any pending state
+
+    virtual int32_t get_n_tokens()   const = 0;
+    virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+
+    virtual bool get_can_shift() const = 0;
+
+    bool get_can_edit() const override { return get_can_shift(); }
+};
+
+struct llama_kv_cache_guard {
+    llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
+
+    ~llama_kv_cache_guard() {
+        kv->restore();
+    }
+
+    void commit() {
+        kv->commit();
+    }
+
+private:
+    llama_kv_cache * kv;
+};
+
 struct llama_kv_cell {
     llama_pos pos   = -1;
-    llama_pos delta = 0;
+    llama_pos delta =  0;
     int32_t   src   = -1; // used by recurrent state models to copy states
     int32_t   tail  = -1;
 
@@ -29,15 +65,112 @@ struct llama_kv_cell {
 };
 
 // ring-buffer of cached KV data
-struct llama_kv_cache {
+// TODO: pimpl
+// TODO: add notion of max sequences
+class llama_kv_cache_unified : public llama_kv_cache {
+public:
+    // can be used to query data from the model if needed
+    struct callbacks {
+        std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
+    };
+
+    llama_kv_cache_unified(
+            const llama_hparams & hparams,
+            callbacks             cbs);
+
+    virtual ~llama_kv_cache_unified() = default;
+
+    // TODO: become constructor
+    bool init(
+            const llama_model & model,   // TODO: do not reference the model
+          const llama_cparams & cparams,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                     uint32_t   kv_size,
+                         bool   offload);
+
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
+
+    size_t total_size() const;
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos pos_max() const;
+
+    void clear() override;
+    void defrag() override;
+
+    virtual void restore() override;
+    virtual void commit() override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id) override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    bool get_can_shift() const override;
+
+    // find an empty slot of size "n_tokens" in the cache
+    // updates the cache head
+    // Note: On success, it's important that cache.head points
+    // to the first cell of the slot.
+    bool find_slot(const llama_ubatch & batch);
+
+    // TODO: maybe not needed
+    uint32_t get_padding(const llama_cparams & cparams) const;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    // defrag
+
+    struct {
+        std::vector<uint32_t> ids;
+    } defrag_info;
+
+    // return true if cells have been moved
+    bool defrag_prepare(int32_t n_max_nodes);
+
+    // commit/restore cache
+
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
+
+    // members
+
+    const llama_hparams & hparams;
+
+    callbacks cbs;
+
     bool has_shift = false;
     bool do_defrag = false;
+
+    // TODO: remove this and implement llama_kv_cache_recurrent instead
     bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
+
     bool v_trans   = true;  // the value tensor is transposed
     bool can_shift = false;
 
     // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_internal also uses it, so it
+    // for a free KV slot. llama_decode_impl also uses it, so it
     // cannot be freely changed after a slot has been allocated.
     uint32_t head = 0;
     uint32_t size = 0;
@@ -46,173 +179,35 @@ struct llama_kv_cache {
     // computed before each graph build
     uint32_t n = 0;
 
+    std::vector<llama_kv_cell> cells;
+
+    std::vector<ggml_tensor *> k_l; // per layer
+    std::vector<ggml_tensor *> v_l;
+
+private:
     ggml_type type_k = GGML_TYPE_F16;
     ggml_type type_v = GGML_TYPE_F16;
 
-    std::vector<llama_kv_cell> cells;
-
-    std::vector<struct ggml_tensor *> k_l; // per layer
-    std::vector<struct ggml_tensor *> v_l;
-
-    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
-    size_t total_size() const {
-        size_t size = 0;
-        for (const auto & buf : bufs) {
-            size += ggml_backend_buffer_get_size(buf.get());
-        }
+    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;
 
-        return size;
-    }
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos max_pos() const {
-        llama_pos max_pos = -1;
-        for (const auto & cell : cells) {
-            max_pos = std::max(max_pos, cell.pos);
-        }
-
-        return max_pos;
-    }
+    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);
 };
 
-// a structure holds information about the slot found in llama_kv_cache_find_slot
-struct llama_kv_cache_slot_info {
-    std::pair<uint32_t, uint32_t> boundaries; // slot boundaries [begin, end)
-    bool found = false;                       // the slot was found
-
-    explicit llama_kv_cache_slot_info(bool found_) : found{found_} {}
-    llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {}
-
-    operator bool() const { return found; }
-};
-
-// TODO: maybe not needed
-uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams);
-
-bool llama_kv_cache_init(
-        struct llama_kv_cache & cache,
-            const llama_model & model,
-          const llama_cparams & cparams,
-                    ggml_type   type_k,
-                    ggml_type   type_v,
-                     uint32_t   kv_size,
-                         bool   offload);
-
-// find an empty slot of size "n_tokens" in the cache
-// updates the cache head
-// returns a structure holding information about the slot found
-// Note: On success, it's important that cache.head points
-// to the first cell of the slot.
-struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
-           struct llama_kv_cache & cache,
-       const struct llama_ubatch & batch);
-
-// find how many cells are currently in use
-uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache);
-
-void llama_kv_cache_clear(struct llama_kv_cache & cache);
-
-bool llama_kv_cache_seq_rm(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id,
-                    llama_pos   p0,
-                    llama_pos   p1);
-
-void llama_kv_cache_seq_cp(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id_src,
-                 llama_seq_id   seq_id_dst,
-                    llama_pos   p0,
-                    llama_pos   p1);
-
-void llama_kv_cache_seq_keep(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id);
-
-void llama_kv_cache_seq_add(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id,
-                    llama_pos   p0,
-                    llama_pos   p1,
-                    llama_pos   delta);
-
-void llama_kv_cache_seq_div(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id,
-                    llama_pos   p0,
-                    llama_pos   p1,
-                          int   d);
-
-llama_pos llama_kv_cache_seq_pos_max(
-        struct llama_kv_cache & cache,
-                 llama_seq_id   seq_id);
-
-void llama_kv_cache_defrag(struct llama_kv_cache & cache);
-
-int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv);
-
-int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv);
-
-bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv);
+// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
+//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
+//public:
+//    using llama_kv_cache_unified::llama_kv_cache_unified;
+//};
 
 //
 // kv cache view
 //
 
-struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max);
-
-void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv);
-
-//
-// kv cache restore
-//
-
-// saves the kv_cache state for future recovery.
-// used to rollback llama_kv_cache_find_slot changes.
-struct llama_kv_slot_restorer {
-    struct llama_kv_cache_state {
-        uint32_t head = 0;
-        uint32_t n    = 0;
-    } old_state;
-
-    // for non-recurrent models only
-    // list of slots to restore
-    std::vector<std::pair<uint32_t, uint32_t>> slot_boundaries;
-
-    bool do_restore = false;
-
-    explicit llama_kv_slot_restorer(const struct llama_kv_cache & cache) {
-        old_state.head = cache.head;
-        old_state.n    = cache.n;
-    }
-
-    // saves a slot information for future restoration
-    void save(const struct llama_kv_cache_slot_info & slot) {
-        if (slot) {
-            do_restore = true;
-            if (slot.boundaries.first != slot.boundaries.second) {
-                slot_boundaries.push_back(slot.boundaries);
-            }
-        }
-    }
-
-    // must be explicitly called to restore the kv_cache state
-    // and rollback changes from all llama_kv_cache_find_slot calls
-    void restore(struct llama_kv_cache & cache) {
-        if (do_restore) {
-            cache.head = old_state.head;
-            cache.n    = old_state.n;
-
-            if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased
-                llama_kv_cache_seq_rm(cache, -1, -1, -1);
-            } else {
-                for (auto & slot : slot_boundaries) {
-                    llama_kv_cache_seq_rm(cache, -1, slot.first, slot.second);
-                }
-            }
-        }
-    }
-};
+llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max);
 
+void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv);

examples/talk-llama/llama-memory.cpp

@@ -0,0 +1 @@
+#include "llama-memory.h"

examples/talk-llama/llama-memory.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include "llama.h"
+
+// general concept of LLM memory
+// the KV cache is a type of LLM memory, but there can be other types
+class llama_memory_i {
+public:
+    virtual void clear() = 0;
+    virtual void defrag() = 0;
+
+    virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_keep(llama_seq_id seq_id) = 0;
+    virtual void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) = 0;
+    virtual void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) = 0;
+
+    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
+
+    virtual bool get_can_edit() const = 0;
+};

examples/talk-llama/llama-mmap.cpp

@@ -8,6 +8,7 @@
 #include <climits>
 #include <stdexcept>
 #include <cerrno>
+#include <algorithm>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -34,6 +35,10 @@
     #include <io.h>
 #endif
 
+#if defined(__APPLE__)
+#include <TargetConditionals.h>
+#endif
+
 // TODO: consider moving to llama-impl.h if needed in more places
 #if defined(_WIN32)
 static std::string llama_format_win_err(DWORD err) {
@@ -471,7 +476,11 @@ struct llama_mlock::impl {
 
         char* errmsg = std::strerror(errno);
         bool suggest = (errno == ENOMEM);
-
+#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
+        // visionOS/tvOS dont't support RLIMIT_MEMLOCK
+        // Skip resource limit checks on visionOS/tvOS
+        suggest = false;
+#else
         struct rlimit lock_limit;
         if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
             suggest = false;
@@ -479,6 +488,7 @@ struct llama_mlock::impl {
         if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
             suggest = false;
         }
+#endif
 
         LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
                 size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");

examples/talk-llama/llama-mmap.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <cstdint>
 #include <memory>
 #include <vector>
 

examples/talk-llama/llama-model-loader.cpp

@@ -445,7 +445,8 @@ llama_model_loader::llama_model_loader(
         std::vector<std::string> & splits,
         bool use_mmap,
         bool check_tensors,
-        const struct llama_model_kv_override * param_overrides_p) {
+        const llama_model_kv_override * param_overrides_p,
+        const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
     int trace = 0;
     if (getenv("LLAMA_TRACE")) {
         trace = atoi(getenv("LLAMA_TRACE"));
@@ -457,6 +458,8 @@ llama_model_loader::llama_model_loader(
         }
     }
 
+    tensor_buft_overrides = param_tensor_buft_overrides_p;
+
     // Load the main GGUF
     struct ggml_context * ctx = NULL;
     struct gguf_init_params params = {
@@ -600,7 +603,9 @@ llama_model_loader::llama_model_loader(
 
             if (trace > 0) {
                 const uint16_t sid = w.idx;
-                LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ]\n", __func__, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str());
+                LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ] %8.2f MiB\n", __func__,
+                        sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str(),
+                        ggml_nbytes(tensor)/1024.0f/1024.0f);
             }
         }
 
@@ -640,9 +645,9 @@ llama_model_loader::llama_model_loader(
         ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED);
 
         {
-            const int kid = gguf_find_key(meta.get(), "general.file_type"); // TODO: use LLM_KV
-            if (kid >= 0) {
-                ftype = (llama_ftype) gguf_get_val_u32(meta.get(), kid);
+            uint32_t ftype_val = 0;
+            if (get_key(LLM_KV_GENERAL_FILE_TYPE, ftype_val, false)) {
+                ftype = (llama_ftype) ftype_val;
             }
         }
 

examples/talk-llama/llama-model-loader.h

@@ -77,8 +77,9 @@ struct llama_model_loader {
 
     llama_mmaps mappings;
 
-    std::map<std::string, struct llama_tensor_weight, weight_name_comparer> weights_map;
-    std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
+    std::map<std::string, llama_tensor_weight, weight_name_comparer> weights_map;
+    std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
+    const llama_model_tensor_buft_override * tensor_buft_overrides;
 
     gguf_context_ptr meta;
     std::vector<ggml_context_ptr> contexts;
@@ -95,7 +96,8 @@ struct llama_model_loader {
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
         bool check_tensors,
-        const struct llama_model_kv_override * param_overrides_p);
+        const llama_model_kv_override * param_overrides_p,
+        const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
 
     template<typename T>
     typename std::enable_if<std::is_integral<T>::value, bool>::type

examples/talk-llama/llama-model.h

@@ -2,7 +2,9 @@
 
 #include "llama.h"
 #include "llama-arch.h"
+#include "llama-graph.h"
 #include "llama-hparams.h"
+#include "llama-memory.h"
 #include "llama-vocab.h"
 
 #include <memory>
@@ -10,6 +12,8 @@
 #include <unordered_map>
 #include <vector>
 
+struct llama_cparams;
+struct llama_ubatch;
 struct llama_model_loader;
 
 // available models
@@ -25,6 +29,7 @@ enum llm_type {
     LLM_TYPE_109M,
     LLM_TYPE_137M,
     LLM_TYPE_160M,
+    LLM_TYPE_190M,
     LLM_TYPE_220M,
     LLM_TYPE_250M,
     LLM_TYPE_270M,
@@ -39,8 +44,10 @@ enum llm_type {
     LLM_TYPE_1_4B,
     LLM_TYPE_1_5B,
     LLM_TYPE_1_6B,
+    LLM_TYPE_1_8B,
     LLM_TYPE_2B,
     LLM_TYPE_2_8B,
+    LLM_TYPE_2_9B,
     LLM_TYPE_3B,
     LLM_TYPE_4B,
     LLM_TYPE_6B,
@@ -78,6 +85,9 @@ enum llm_type {
     LLM_TYPE_10B_128x3_66B,
     LLM_TYPE_57B_A14B,
     LLM_TYPE_27B,
+    LLM_TYPE_290B,
+    LLM_TYPE_17B_16E, // llama4 Scout
+    LLM_TYPE_17B_128E, // llama4 Maverick
 };
 
 struct llama_layer_posnet {
@@ -161,6 +171,8 @@ struct llama_layer {
     struct ggml_tensor * wq_b      = nullptr;
     struct ggml_tensor * wkv_a_mqa = nullptr;
     struct ggml_tensor * wkv_b     = nullptr;
+    struct ggml_tensor * wk_b      = nullptr;
+    struct ggml_tensor * wv_b      = nullptr;
     struct ggml_tensor * wq_cross  = nullptr;
     struct ggml_tensor * wk_cross  = nullptr;
     struct ggml_tensor * wv_cross  = nullptr;
@@ -256,6 +268,20 @@ struct llama_layer {
     struct ggml_tensor * time_mix_receptance_b = nullptr;
     struct ggml_tensor * time_mix_gate         = nullptr;
 
+    // rwkv7
+    struct ggml_tensor * time_mix_w0         = nullptr;
+    struct ggml_tensor * time_mix_a0         = nullptr;
+    struct ggml_tensor * time_mix_a1         = nullptr;
+    struct ggml_tensor * time_mix_a2         = nullptr;
+    struct ggml_tensor * time_mix_v0         = nullptr;
+    struct ggml_tensor * time_mix_v1         = nullptr;
+    struct ggml_tensor * time_mix_v2         = nullptr;
+    struct ggml_tensor * time_mix_g1         = nullptr;
+    struct ggml_tensor * time_mix_g2         = nullptr;
+    struct ggml_tensor * time_mix_k_k        = nullptr;
+    struct ggml_tensor * time_mix_k_a        = nullptr;
+    struct ggml_tensor * time_mix_r_k        = nullptr;
+
     struct ggml_tensor * time_mix_ln     = nullptr;
     struct ggml_tensor * time_mix_ln_b   = nullptr;
     struct ggml_tensor * time_mix_output = nullptr;
@@ -347,7 +373,7 @@ struct llama_model {
     std::string desc() const;
 
     size_t size() const;
-    size_t max_nodes() const;
+    size_t n_tensors() const;
     size_t n_devices() const;
 
     // total number of parameters in the model
@@ -360,11 +386,26 @@ struct llama_model {
 
     ggml_backend_buffer_type_t select_buft(int il) const;
 
+    bool has_tensor_overrides() const;
+
     const struct ggml_tensor * get_tensor(const char * name) const;
 
+    // TODO: move this to new llm_arch_model_i interface
+    llama_memory_i * create_memory() const; // TODO: params
+
+    // 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;
+
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;
 };
 
 const char * llm_type_name(llm_type type);
+
+// For internal test use
+// TODO: remove
+const std::vector<std::pair<std::string, ggml_tensor *>> & llama_internal_get_tensor_map(const llama_model * model);

examples/talk-llama/llama-quant.cpp

@@ -10,6 +10,7 @@
 #include <cinttypes>
 #include <fstream>
 #include <mutex>
+#include <regex>
 #include <thread>
 #include <unordered_map>
 
@@ -47,8 +48,14 @@ struct quantize_state_impl {
         {}
 };
 
+// changes to this struct must be replicated in quantize.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
 static void llama_tensor_dequantize_impl(
-    struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
+    ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
     const size_t nelements, const int nthread
 ) {
     if (output.size() < nelements) {
@@ -527,7 +534,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());
@@ -536,7 +543,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     model.load_hparams(ml);
     model.load_stats  (ml);
 
-    struct quantize_state_impl qs(model, params);
+    quantize_state_impl qs(model, params);
 
     if (params->only_copy) {
         ftype = ml.ftype;
@@ -661,7 +668,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     // populate the original tensors so we get an initial meta data
     for (const auto * it : tensors) {
         uint16_t i_split = params->keep_split ? it->idx : 0;
-        struct ggml_tensor * tensor = it->tensor;
+        ggml_tensor * tensor = it->tensor;
         if (!ctx_outs[i_split]) {
             ctx_outs[i_split].reset(gguf_init_empty());
         }
@@ -710,7 +717,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     new_ofstream(0);
     for (const auto * it : tensors) {
         const auto & weight = *it;
-        struct ggml_tensor * tensor = weight.tensor;
+        ggml_tensor * tensor = weight.tensor;
         if (weight.idx != cur_split && params->keep_split) {
             close_ofstream();
             new_ofstream(weight.idx);
@@ -756,10 +763,19 @@ 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("ssm_conv1d.weight") == std::string::npos;
 
-        // do not quantize RWKV's time_mix_first tensors
+        // do not quantize RWKV's small yet 2D weights
         quantize &= name.find("time_mix_first.weight") == std::string::npos;
+        quantize &= name.find("time_mix_w0.weight") == std::string::npos;
         quantize &= name.find("time_mix_w1.weight") == std::string::npos;
         quantize &= name.find("time_mix_w2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_v0.weight") == std::string::npos;
+        quantize &= name.find("time_mix_v1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_v2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_a0.weight") == std::string::npos;
+        quantize &= name.find("time_mix_a1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_a2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_g1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_g2.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
         quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos;
@@ -767,7 +783,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
 
-        enum ggml_type new_type;
+        ggml_type new_type;
         void * new_data;
         size_t new_size;
 
@@ -777,6 +793,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
                 new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+                // unless the user specifies a type
+                if (params->tensor_types) {
+                    const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
+                    for (const auto & [tname, qtype] : tensor_types) {
+                        if (std::regex pattern(tname); std::regex_search(tensor->name, pattern)) {
+                            if (qtype != new_type) {
+                                LLAMA_LOG_DEBUG("(overriding %s -> %s), ", ggml_type_name(new_type), ggml_type_name(qtype));
+                            }
+                            new_type = qtype;
+                            break;
+                        }
+                    }
+                }
             }
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
@@ -901,8 +930,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 // interface implementation
 //
 
-struct llama_model_quantize_params llama_model_quantize_default_params() {
-    struct llama_model_quantize_params result = {
+llama_model_quantize_params llama_model_quantize_default_params() {
+    llama_model_quantize_params result = {
         /*.nthread                     =*/ 0,
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
         /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
@@ -914,6 +943,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.keep_split                  =*/ false,
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
+        /*.tensor_type                 =*/ nullptr,
     };
 
     return result;

examples/talk-llama/llama-sampling.cpp

@@ -316,6 +316,13 @@ static uint32_t get_rng_seed(uint32_t seed) {
 
 // llama_sampler API
 
+struct llama_sampler * llama_sampler_init(const struct llama_sampler_i * iface, llama_sampler_context_t ctx) {
+    return new llama_sampler {
+        /* .iface = */ iface,
+        /* .ctx   = */ ctx,
+    };
+}
+
 const char * llama_sampler_name(const struct llama_sampler * smpl) {
     if (!smpl->iface) {
         return "(null)";
@@ -347,10 +354,10 @@ struct llama_sampler * llama_sampler_clone(const struct llama_sampler * smpl) {
     }
 
     if (smpl->ctx == nullptr) {
-        return new llama_sampler {
+        return llama_sampler_init(
             /* .iface = */ smpl->iface,
-            /* .ctx   = */ nullptr,
-        };
+            /* .ctx   = */ nullptr
+        );
     }
 
     GGML_ABORT("the sampler does not support cloning");
@@ -472,15 +479,15 @@ static struct llama_sampler_i llama_sampler_chain_i = {
 };
 
 struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_chain_i,
         /* .ctx   = */ new llama_sampler_chain {
             /* .params      = */ params,
             /* .samplers    = */ {},
             /* .t_sample_us = */ 0,
             /* .n_sample    = */ 0,
-        },
-    };
+        }
+    );
 }
 
 void llama_sampler_chain_add(struct llama_sampler * chain, struct llama_sampler * smpl) {
@@ -546,10 +553,10 @@ static struct llama_sampler_i llama_sampler_greedy_i = {
 };
 
 struct llama_sampler * llama_sampler_init_greedy() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_greedy_i,
-        /* .ctx   = */ nullptr,
-    };
+        /* .ctx   = */ nullptr
+    );
 }
 
 // dist
@@ -608,14 +615,14 @@ static struct llama_sampler_i llama_sampler_dist_i = {
 
 struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dist_i,
         /* .ctx   = */ new llama_sampler_dist {
             /* .seed     = */ seed,
             /* .seed_cur = */ seed_cur,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // softmax
@@ -638,10 +645,10 @@ static struct llama_sampler_i llama_sampler_softmax_i = {
 };
 
 struct llama_sampler * llama_sampler_init_softmax() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_softmax_i,
-        /* .ctx   = */ nullptr,
-    };
+        /* .ctx   = */ nullptr
+    );
 }
 
 // top-k
@@ -678,12 +685,12 @@ static struct llama_sampler_i llama_sampler_top_k_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_k_i,
         /* .ctx   = */ new llama_sampler_top_k {
             /* .k = */ k,
-        },
-    };
+        }
+    );
 }
 
 // top-p
@@ -744,13 +751,13 @@ static struct llama_sampler_i llama_sampler_top_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_p_i,
         /* .ctx   = */ new llama_sampler_top_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // min-p
@@ -840,13 +847,13 @@ static struct llama_sampler_i llama_sampler_min_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_min_p_i,
         /* .ctx   = */ new llama_sampler_min_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // typical
@@ -939,13 +946,13 @@ static struct llama_sampler_i llama_sampler_typical_i = {
 };
 
 struct llama_sampler * llama_sampler_init_typical(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_typical_i,
         /* .ctx   = */ new llama_sampler_typical {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // temp
@@ -983,12 +990,12 @@ static struct llama_sampler_i llama_sampler_temp_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp(float temp) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_i,
         /* .ctx   = */ new llama_sampler_temp {
             /*.temp = */ temp,
-        },
-    };
+        }
+    );
 }
 
 // temp-ext
@@ -1093,14 +1100,14 @@ static struct llama_sampler_i llama_sampler_temp_ext_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, float exponent) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_ext_i,
         /* .ctx   = */ new llama_sampler_temp_ext {
             /* .temp     = */ temp,
             /* .delta    = */ delta,
             /* .exponent = */ exponent,
-        },
-    };
+        }
+    );
 }
 
 // xtc
@@ -1185,7 +1192,7 @@ static struct llama_sampler_i llama_sampler_xtc_i = {
 
 struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_xtc_i,
         /* .ctx   = */ new llama_sampler_xtc {
             /* .probability   = */ p,
@@ -1194,8 +1201,8 @@ struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep,
             /* .seed          = */ seed,
             /* .seed_cur      = */ seed_cur,
             /* .rng           = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // mirostat
@@ -1292,7 +1299,7 @@ static struct llama_sampler_i llama_sampler_mirostat_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t seed, float tau, float eta, int32_t m) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_i,
         /* .ctx   = */ new llama_sampler_mirostat {
             /* .n_vocab  = */ n_vocab,
@@ -1303,8 +1310,8 @@ struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t see
             /* .m        = */ m,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // mirostat v2
@@ -1391,7 +1398,7 @@ static struct llama_sampler_i llama_sampler_mirostat_v2_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, float eta) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_v2_i,
         /* .ctx   = */ new llama_sampler_mirostat_v2 {
             /* .seed     = */ seed,
@@ -1400,8 +1407,8 @@ struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau,
             /* .eta      = */ eta,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // grammar
@@ -1442,7 +1449,9 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                      const char ** trigger_words,
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
-                            size_t num_trigger_tokens);
+                            size_t num_trigger_tokens,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns);
 
 static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_grammar *) smpl->ctx;
@@ -1450,12 +1459,14 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
         return;
     }
 
-    std::vector<const char *>  trigger_words;
-    for (auto & word : ctx->grammar->trigger_words) {
-        trigger_words.push_back(word.c_str());
+    std::vector<const char *>  trigger_patterns_c;
+    trigger_patterns_c.reserve(ctx->grammar->trigger_patterns.size());
+    for (auto & trigger_pattern : ctx->grammar->trigger_patterns) {
+        trigger_patterns_c.push_back(trigger_pattern.pattern.c_str());
     }
+
     auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(),
-                                                 ctx->grammar->lazy, trigger_words.data(), trigger_words.size(),
+                                                 ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                  ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
 
     llama_grammar_free_impl(ctx->grammar);
@@ -1465,7 +1476,8 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_grammar *) smpl->ctx;
 
-    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0);
+    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0, nullptr, 0);
+    GGML_ASSERT(result);
 
     // copy the state
     {
@@ -1509,16 +1521,38 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                      const char ** trigger_words,
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
-                            size_t num_trigger_tokens) {
+                            size_t num_trigger_tokens,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns) {
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        // TODO: remove trigger_words support.
+        if (trigger_words != nullptr && num_trigger_words > 0) {
+            GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
+            std::string trigger_pattern("[\\s\\S]*?(");
+            for (size_t i = 0; i < num_trigger_words; ++i) {
+                static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
+                if (i > 0) {
+                    trigger_pattern += "|";
+                }
+                trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
+            }
+            trigger_pattern += ")[\\s\\S]*";
+            auto trigger_pattern_c = trigger_pattern.c_str();
+            trigger_patterns = &trigger_pattern_c;
+            num_trigger_patterns = 1;
+        }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
         };
+        if (!ctx->grammar) {
+            delete ctx;
+            return nullptr;
+        }
     } else {
         *ctx = {
             /* .vocab        = */ vocab,
@@ -1528,17 +1562,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
         };
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_grammar_i,
-        /* .ctx   = */ ctx,
-    };
+        /* .ctx   = */ ctx
+    );
 }
 
 struct llama_sampler * llama_sampler_init_grammar(
         const struct llama_vocab * vocab,
                       const char * grammar_str,
                       const char * grammar_root) {
-    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0);
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0, nullptr, 0);
 }
 
 struct llama_sampler * llama_sampler_init_grammar_lazy(
@@ -1549,7 +1583,18 @@ struct llama_sampler * llama_sampler_init_grammar_lazy(
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
-    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens);
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens, nullptr, 0);
+}
+
+struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, nullptr, 0, trigger_tokens, num_trigger_tokens, trigger_patterns, num_trigger_patterns);
 }
 
 // penalties
@@ -1678,7 +1723,7 @@ struct llama_sampler * llama_sampler_init_penalties(
         float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_penalties_i,
         /* .ctx   = */ new llama_sampler_penalties {
             /* .penalty_last_n  = */ penalty_last_n,
@@ -1687,8 +1732,75 @@ struct llama_sampler * llama_sampler_init_penalties(
             /* .penalty_present = */ penalty_present,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
             /* .token_count     = */ {},
-        },
-    };
+        }
+    );
+}
+
+// top-n-sigma
+
+struct llama_sampler_top_n_sigma {
+    const float n;
+};
+
+static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler * /*smpl*/) {
+    return "top-n-sigma";
+}
+
+static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
+
+    // find max logit and calculate mean
+    float max = cur_p->data[0].logit;
+    float logits_sum = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit > max) {
+            max = cur_p->data[i].logit;
+        }
+        logits_sum += cur_p->data[i].logit;
+    }
+    float mean = logits_sum/cur_p->size;
+
+    // calculate standard deviation
+    float acc = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        acc += pow(cur_p->data[i].logit - mean, 2);
+    }
+    float std = sqrt(acc/cur_p->size);
+
+    //apply mask
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit < max - (ctx->n * std)) {
+            cur_p->data[i].logit = -INFINITY;
+        }
+    }
+    llama_sampler_softmax_impl(cur_p);
+}
+
+static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_top_n_sigma *) smpl->ctx;
+    return llama_sampler_init_top_n_sigma(ctx->n);
+}
+
+static void llama_sampler_top_n_sigma_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_top_n_sigma *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_top_n_sigma_i = {
+    /* .name   = */ llama_sampler_top_n_sigma_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_top_n_sigma_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_top_n_sigma_clone,
+    /* .free   = */ llama_sampler_top_n_sigma_free,
+};
+
+struct llama_sampler * llama_sampler_init_top_n_sigma(float n) {
+    return llama_sampler_init(
+        /* .iface = */ &llama_sampler_top_n_sigma_i,
+        /* .ctx   = */ new llama_sampler_top_n_sigma {
+            /* .n = */ n,
+        }
+    );
 }
 
 // DRY
@@ -2041,7 +2153,7 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
         }
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dry_i,
         /* .ctx   = */ new llama_sampler_dry {
             /* .total_context_size     = */ context_size,
@@ -2053,8 +2165,8 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
             /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
             /* .dry_max_token_repeat   = */ {},
             /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
-        },
-    };
+        }
+    );
 }
 
 // wrapper for test-sampling.cpp
@@ -2155,14 +2267,14 @@ struct llama_sampler * llama_sampler_init_logit_bias(
                          int32_t   n_vocab,
                          int32_t   n_logit_bias,
           const llama_logit_bias * logit_bias) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_logit_bias_i,
         /* .ctx   = */ new llama_sampler_logit_bias {
             /* .n_vocab    = */ n_vocab,
             /* .logit_bias = */ std::vector<llama_logit_bias>(logit_bias, logit_bias + n_logit_bias),
             /* .to_search  = */ {},
-        },
-    };
+        }
+    );
 }
 
 // infill
@@ -2377,14 +2489,14 @@ static struct llama_sampler_i llama_sampler_infill_i = {
 };
 
 struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_infill_i,
         /* .ctx   = */ new llama_sampler_infill {
             /* .vocab = */ vocab,
             /* .buf0  = */ std::vector<char>(512),
             /* .buf1  = */ std::vector<char>(512),
-        },
-    };
+        }
+    );
 }
 
 // utils

examples/talk-llama/llama-vocab.cpp

@@ -16,6 +16,7 @@
 #include <queue>
 #include <set>
 #include <unordered_map>
+#include <cctype>
 
 //
 // helpers
@@ -341,6 +342,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_MPT:
             case LLAMA_VOCAB_PRE_TYPE_OLMO:
             case LLAMA_VOCAB_PRE_TYPE_JAIS:
+            case LLAMA_VOCAB_PRE_TYPE_TRILLION:
                 regex_exprs = {
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                 };
@@ -392,6 +394,27 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?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{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?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+",
+                    "[^\\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_SUPERBPE:
+                regex_exprs = {
+                    "\\p{N}+",
+                    "(?=(\\d{3})+(?!\\d))",
+                };
+                break;
+            case LLAMA_VOCAB_PRE_TYPE_BAILINGMOE:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "'(?i:[sdmt]|ll|ve|re)|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+"
+                    // FIXME? Changed possessive quantifiers (?+ and ++) to greedy to avoid errors and imatrix hanging (tried atomic grouping but it's not supported?)
+                    "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1483,7 +1506,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama3"   ||
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe"||
-                    tokenizer_pre == "falcon3") {
+                    tokenizer_pre == "falcon3"  ||
+                    tokenizer_pre == "pixtral") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;
@@ -1549,6 +1573,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_PORO;
                 clean_spaces = false;
             } else if (
+                tokenizer_pre == "glm4" ||
                 tokenizer_pre == "chatglm-bpe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_CHATGLM4;
                 special_bos_id = LLAMA_TOKEN_NULL;
@@ -1592,6 +1617,23 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                 tokenizer_pre == "megrez") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                    tokenizer_pre == "gpt-4o" ||
+                    tokenizer_pre == "llama4") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "superbpe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "trillion") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_TRILLION;
+                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "bailingmoe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -1769,6 +1811,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<end_of_turn>"
                         || t.first == "<|endoftext|>"
                         || t.first == "<EOT>"
+                        || t.first == "_<EOT>"
                         || t.first == "<｜end▁of▁sentence｜>" // DeepSeek
                    ) {
                     special_eot_id = t.second;
@@ -1799,8 +1842,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_prefix|>"  // Qwen
                         || t.first == "<fim-prefix>"
+                        || t.first == "<fim_prefix>"    // Granite
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
+                        || t.first == "▁<PRE>"          // CodeLlama
                         ) {
                     special_fim_pre_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1816,8 +1861,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_suffix|>" // Qwen
                         || t.first == "<fim-suffix>"
+                        || t.first == "<fim_suffix>"   // Granite
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
+                        || t.first == "▁<SUF>"         // CodeLlama
                         ) {
                     special_fim_suf_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1833,8 +1880,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_middle|>" // Qwen
                         || t.first == "<fim-middle>"
+                        || t.first == "<fim_middle>"   // Granite
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
+                        || t.first == "▁<MID>"         // CodeLlama
                         ) {
                     special_fim_mid_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1850,6 +1899,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_pad|>" // Qwen
                         || t.first == "<fim-pad>"
+                        || t.first == "<fim_pad>"   // Granite
                         || t.first == "<PAD>"
                         ) {
                     special_fim_pad_id = t.second;
@@ -1868,6 +1918,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|repo_name|>"
                         || t.first == "<fim-repo>"
                         || t.first == "<REPO>"
+                        || t.first == "<reponame>"    // Granite
                         ) {
                     special_fim_rep_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1919,6 +1970,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|endoftext|>"
                     || t.first == "<|eom_id|>"
                     || t.first == "<EOT>"
+                    || t.first == "_<EOT>"
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2177,14 +2229,12 @@ void llama_vocab::impl::tokenizer_st_partition(std::forward_list<fragment_buffer
                     // find the first occurrence of a given special token in this fragment
                     //  passing offset argument only limit the "search area" but match coordinates
                     //  are still relative to the source full raw_text
-                    auto match = raw_text.find(text, raw_text_base_offset);
+                    //  string_view begins at pos 0 for the same reason
+                    auto match = std::string_view(raw_text.data(), raw_text_base_offset + raw_text_base_length).find(text, raw_text_base_offset);
 
                     // no occurrences found, stop processing this fragment for a given special token
                     if (match == std::string::npos) break;
 
-                    // check if match is within bounds of offset <-> length
-                    if (match + text.length() > raw_text_base_offset + raw_text_base_length) break;
-
 #ifdef PRETOKENIZERDEBUG
                     LLAMA_LOG_WARN("FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
 #endif

examples/talk-llama/llama.h

@@ -60,6 +60,7 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
+    struct llama_kv_cache;
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -105,6 +106,12 @@ extern "C" {
         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,
     };
 
     enum llama_rope_type {
@@ -213,7 +220,7 @@ extern "C" {
         LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
 
-    // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
+    // TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979)
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -275,10 +282,18 @@ extern "C" {
         };
     };
 
+    struct llama_model_tensor_buft_override {
+        const char * pattern;
+        ggml_backend_buffer_type_t buft;
+    };
+
     struct llama_model_params {
         // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
         ggml_backend_dev_t * devices;
 
+        // NULL-terminated list of buffer types to use for tensors that match a pattern
+        const struct llama_model_tensor_buft_override * tensor_buft_overrides;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
@@ -307,7 +322,7 @@ extern "C" {
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
-    //       https://github.com/ggerganov/llama.cpp/pull/7544
+    //       https://github.com/ggml-org/llama.cpp/pull/7544
     struct llama_context_params {
         uint32_t n_ctx;             // text context, 0 = from model
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
@@ -320,7 +335,7 @@ extern "C" {
         enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
         enum llama_attention_type    attention_type;    // attention type to use for embeddings
 
-        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        // ref: https://github.com/ggml-org/llama.cpp/pull/2054
         float    rope_freq_base;   // RoPE base frequency, 0 = from model
         float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
         float    yarn_ext_factor;  // YaRN extrapolation mix factor, negative = from model
@@ -353,17 +368,18 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // token embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;                      // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;               // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;    // output tensor type
+        enum ggml_type token_embedding_type;  // token embeddings tensor type
+        bool allow_requantize;                // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;          // quantize output.weight
+        bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                            // quantize all tensors to the default type
+        bool keep_split;                      // quantize to the same number of shards
+        void * imatrix;                       // pointer to importance matrix data
+        void * kv_overrides;                  // pointer to vector containing overrides
+        void * tensor_types;                  // pointer to vector containing tensor types
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -385,7 +401,7 @@ extern "C" {
     struct llama_adapter_lora;
 
     // Helpers for getting default parameters
-    // TODO: update API to start accepting pointers to params structs (https://github.com/ggerganov/llama.cpp/discussions/9172)
+    // TODO: update API to start accepting pointers to params structs (https://github.com/ggml-org/llama.cpp/discussions/9172)
     LLAMA_API struct llama_model_params          llama_model_default_params(void);
     LLAMA_API struct llama_context_params        llama_context_default_params(void);
     LLAMA_API struct llama_sampler_chain_params  llama_sampler_chain_default_params(void);
@@ -468,7 +484,8 @@ extern "C" {
     DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
 
     LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
-    LLAMA_API enum llama_pooling_type    llama_pooling_type(const struct llama_context * ctx);
+    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
@@ -477,6 +494,7 @@ extern "C" {
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
@@ -584,7 +602,7 @@ extern "C" {
     // KV cache
     //
 
-    // TODO: remove llama_kv_cache_view_* API
+    // TODO: start using struct llama_kv_cache
 
     // Information associated with an individual cell in the KV cache view.
     struct llama_kv_cache_view_cell {
@@ -639,13 +657,19 @@ extern "C" {
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
     // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "use llama_kv_self_n_tokens instead");
 
     // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx),
+            "use llama_kv_self_used_cells instead");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_cache_clear(
+    LLAMA_API void llama_kv_self_clear(
             struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
@@ -653,7 +677,7 @@ extern "C" {
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_cache_seq_rm(
+    LLAMA_API bool llama_kv_self_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -663,7 +687,7 @@ extern "C" {
     // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_cp(
+    LLAMA_API void llama_kv_self_seq_cp(
             struct llama_context * ctx,
                     llama_seq_id   seq_id_src,
                     llama_seq_id   seq_id_dst,
@@ -671,17 +695,17 @@ extern "C" {
                        llama_pos   p1);
 
     // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_cache_seq_keep(
+    LLAMA_API void llama_kv_self_seq_keep(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_add(
+    LLAMA_API void llama_kv_self_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -691,10 +715,10 @@ extern "C" {
     // Integer division of the positions by factor of `d > 1`
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_div(
+    LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -702,24 +726,76 @@ extern "C" {
                              int   d);
 
     // Returns the largest position present in the KV cache for the specified sequence
-    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache
-    //       how to avoid this?
+                     llama_seq_id   seq_id);
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
+    //   - explicitly with llama_kv_self_update()
+    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
 
     // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+
+    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx),
+            "use llama_kv_self_clear instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_rm instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_cp instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_keep instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_add(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta),
+            "use llama_kv_self_seq_add instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d),
+            "use llama_kv_self_seq_div instead");
+
+    DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_pos_max instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx),
+            "use llama_kv_self_defrag instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx),
+            "use llama_kv_self_can_shift instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx),
+            "use llama_kv_self_update instead");
+
 
     //
     // State / sessions
@@ -883,6 +959,10 @@ extern "C" {
     // If set to true, the model will only attend to the past tokens
     LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
 
+    // Set whether the model is in warmup mode or not
+    // If true, all model tensors are activated during llama_decode() to load and cache their weights.
+    LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup);
+
     // Set abort callback
     LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
 
@@ -1040,7 +1120,7 @@ extern "C" {
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
     /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat
@@ -1114,11 +1194,12 @@ extern "C" {
     };
 
     struct llama_sampler {
-        struct llama_sampler_i  * iface;
-        llama_sampler_context_t   ctx;
+        const struct llama_sampler_i * iface;
+        llama_sampler_context_t        ctx;
     };
 
     // mirror of llama_sampler_i:
+    LLAMA_API struct llama_sampler * llama_sampler_init  (const struct llama_sampler_i * iface, llama_sampler_context_t ctx);
     LLAMA_API const char *           llama_sampler_name  (const struct llama_sampler * smpl);
     LLAMA_API void                   llama_sampler_accept(      struct llama_sampler * smpl, llama_token token);
     LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
@@ -1148,7 +1229,7 @@ extern "C" {
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
     DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
-        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
+        "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1156,7 +1237,7 @@ extern "C" {
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_p      (float   p, size_t min_keep);
 
-    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    /// @details Minimum P sampling as described in https://github.com/ggml-org/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
@@ -1171,6 +1252,9 @@ extern "C" {
     /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
     LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
 
+    /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641
+    LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float   n);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1194,22 +1278,38 @@ extern "C" {
                                float   tau,
                                float   eta);
 
+    /// @details Intializes a GBNF grammar, see grammars/README.md for details.
+    /// @param vocab The vocabulary that this grammar will be used with.
+    /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
+    /// @param grammar_root The name of the start symbol for the grammar.
     LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
             const struct llama_vocab * vocab,
                           const char * grammar_str,
                           const char * grammar_root);
 
-    /// @details Lazy grammar sampler, introduced in https://github.com/ggerganov/llama.cpp/pull/9639
-    /// @param trigger_words A list of words that will trigger the grammar sampler. This may be updated to a loose regex syntax (w/ ^) in a near future.
-    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler.
-    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
             const struct llama_vocab * vocab,
                           const char * grammar_str,
                           const char * grammar_root,
                          const char ** trigger_words,
                                 size_t num_trigger_words,
                    const llama_token * trigger_tokens,
-                                size_t num_trigger_tokens);
+                                size_t num_trigger_tokens),
+        "use llama_sampler_init_grammar_lazy_patterns instead");
+
+
+    /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639
+    /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group.
+    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included.
+    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
 
     /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(

examples/talk-llama/unicode.cpp

@@ -618,7 +618,14 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     result.reserve(utf8.size());
     size_t offset = 0;
     while (offset < utf8.size()) {
-        result.push_back(unicode_cpt_from_utf8(utf8, offset));
+        try {
+            result.push_back(unicode_cpt_from_utf8(utf8, offset));
+        }
+        catch (const std::invalid_argument & /*ex*/) {
+            // Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize
+            ++offset;
+            result.emplace_back(0xFFFD); // replacement character
+        }
     }
     return result;
 }
@@ -701,7 +708,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
     const auto cpts = unicode_cpts_from_utf8(text);
 
     // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
-    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
     std::string text_collapsed;
     if (need_collapse) {
         // collapse all unicode categories

examples/whisper.wasm/CMakeLists.txt

@@ -36,7 +36,7 @@ set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     -s MAXIMUM_MEMORY=2000MB \
     -s ALLOW_MEMORY_GROWTH=1 \
     -s FORCE_FILESYSTEM=1 \
-    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap', 'HEAPU8']\" \
     ${EXTRA_FLAGS} \
     ")
 

ggml/CMakeLists.txt

@@ -107,6 +107,7 @@ message(DEBUG "INS_ENB             : ${INS_ENB}")
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
 option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
+option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
 option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
@@ -170,7 +171,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 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_UMA                         "ggml: use HIP unified memory architecture"       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)
@@ -360,3 +360,18 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4005  # Macro redefinition
+        /wd4244  # Conversion from one type to another type, possible loss of data
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+    )
+    function(disable_msvc_warnings target_name)
+        target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+endif()

ggml/include/ggml-rpc.h

@@ -7,6 +7,9 @@
 extern "C" {
 #endif
 
+#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API

ggml/include/ggml.h

@@ -481,6 +481,7 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
@@ -507,17 +508,12 @@ extern "C" {
 
         GGML_OP_UNARY,
 
-        GGML_OP_MAP_UNARY,
-        GGML_OP_MAP_BINARY,
-
-        GGML_OP_MAP_CUSTOM1_F32,
-        GGML_OP_MAP_CUSTOM2_F32,
-        GGML_OP_MAP_CUSTOM3_F32,
-
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
 
+        GGML_OP_CUSTOM,
+
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
@@ -682,6 +678,9 @@ extern "C" {
     GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
     GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // 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);
+
     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);
 
@@ -1665,7 +1664,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // depthwise
+    // depthwise (via im2col and mul_mat)
     GGML_API struct ggml_tensor * ggml_conv_2d_dw(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,  // convolution kernel
@@ -1677,6 +1676,22 @@ extern "C" {
             int                  d0,  // dilation dimension 0
             int                  d1); // dilation dimension 1
 
+    // Depthwise 2D convolution
+    // may be faster than ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1722,24 +1737,29 @@ extern "C" {
             float                 p0,
             float                 p1);
 
-    // nearest interpolate
+    enum ggml_scale_mode {
+        GGML_SCALE_MODE_NEAREST  = 0,
+        GGML_SCALE_MODE_BILINEAR = 1,
+    };
+
+    // interpolate
     // multiplies ne0 and ne1 by scale factor
-    // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   scale_factor);
+            int                   scale_factor,
+            enum ggml_scale_mode  mode);
 
-    // nearest interpolate
-    // nearest interpolate to specified dimensions
-    // used in tortoise.cpp
+    // interpolate
+    // interpolate scale to specified dimensions
     GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
-            int                   ne3);
+            int                   ne3,
+            enum ggml_scale_mode  mode);
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1916,83 +1936,6 @@ extern "C" {
 
     // custom operators
 
-    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
-            struct ggml_context        * ctx,
-            struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun),
-        "use ggml_map_custom1 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
-            struct ggml_context        * ctx,
-            struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun),
-        "use ggml_map_custom1_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun),
-        "use ggml_map_custom2 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun),
-        "use ggml_map_custom2_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun),
-        "use ggml_map_custom1 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun),
-        "use ggml_map_custom1_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun),
-        "use ggml_map_custom2 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun),
-        "use ggml_map_custom2_inplace instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-            struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun),
-        "use ggml_map_custom3 instead");
-
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
-            struct ggml_context          * ctx,
-            struct ggml_tensor           * a,
-            struct ggml_tensor           * b,
-            struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun),
-        "use ggml_map_custom3_inplace instead");
-
-    // custom operators v2
-
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
     typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
@@ -2048,6 +1991,30 @@ extern "C" {
             int                     n_tasks,
             void                  * userdata);
 
+    typedef void (*ggml_custom_op_t)(struct ggml_tensor * dst , int ith, int nth, void * userdata);
+
+    GGML_API struct ggml_tensor * ggml_custom_4d(
+            struct ggml_context * ctx,
+            enum ggml_type        type,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            struct ggml_tensor ** args,
+            int                   n_args,
+            ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
+    GGML_API struct ggml_tensor * ggml_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor ** args,
+            int                   n_args,
+            ggml_custom_op_t      fun,
+            int                   n_tasks,
+            void                * userdata);
+
     // loss function
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss(

ggml/src/CMakeLists.txt

@@ -267,6 +267,7 @@ function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
     foreach (feat NATIVE
+                  SSE42
                   AVX AVX2 BMI2 AVX_VNNI FMA F16C
                   AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
                   AMX_TILE AMX_INT8 AMX_BF16)
@@ -286,14 +287,16 @@ if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
     endif()
-    ggml_add_cpu_backend_variant(sandybridge    AVX)
-    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+    ggml_add_cpu_backend_variant(x64)
+    ggml_add_cpu_backend_variant(sse42        SSE42)
+    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
     if (NOT MSVC)
         # MSVC doesn't support AMX
-        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-cann/acl_tensor.cpp

@@ -41,6 +41,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT4;
         case GGML_TYPE_Q8_0:
             return ACL_INT8;
+        case GGML_TYPE_I64:
+            return ACL_INT64;
         default:
             return ACL_DT_UNDEFINED;
     }
@@ -54,9 +56,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     // added.
     int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
 
-    int64_t acl_storage_len = 0;
     if (ne == nullptr) {
-        acl_storage_len = ggml_nbytes(tensor);
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
             acl_ne[i] = tensor->ne[i];
             // The step size of acl is in elements.
@@ -65,14 +65,18 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
     } else {
         // With bcast
         for (int i = 0; i < dims; i++) {
-            acl_storage_len += (ne[i] - 1) * nb[i];
             acl_ne[i] = ne[i];
             acl_stride[i] = nb[i] / ggml_element_size(tensor);
         }
     }
 
-    // Reverse ne and stride.
     int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
+    int64_t acl_storage_len = 1;
+    for (int i = 0; i < final_dims; i++) {
+        acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
+    }
+
+    // Reverse ne and stride.
     std::reverse(acl_ne, acl_ne + final_dims);
     std::reverse(acl_stride, acl_stride + final_dims);
 

ggml/src/ggml-cann/acl_tensor.h

@@ -101,14 +101,14 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
         tmp_stride[i] = nb[i] / type_size;
     }
 
+    int64_t acl_storage_len = 1;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i];
+    }
+
     std::reverse(tmp_ne, tmp_ne + dims);
     std::reverse(tmp_stride, tmp_stride + dims);
 
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
     aclTensor* acl_tensor =
         aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
                         format, &acl_storage_len, 1, data_ptr);

ggml/src/ggml-cann/aclnn_ops.h

@@ -1,15 +1,4 @@
-#ifndef CANN_ACLNN_OPS
-#define CANN_ACLNN_OPS
-
 /**
- * @file    acl_tensor
- * @brief   This file contains related functions of ggml_tensor and acl_tensor.
- *          Contains conversion from ggml_tensor to acl_tensor, broadcast and other
- *          functions.
- * @author  hipudding <huafengchun@gmail.com>
- * @author  wangshuai09 <391746016@qq.com>
- * @date    July 15, 2024
- *
  * Copyright (c) 2023-2024 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
@@ -31,20 +20,31 @@
  * IN THE SOFTWARE.
  */
 
-#include <aclnnop/aclnn_add.h>
+#ifndef CANN_ACLNN_OPS
+#define CANN_ACLNN_OPS
+
+#include <functional>
+#include <aclnnop/aclnn_abs.h>
+#include <aclnnop/aclnn_neg.h>
+#include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_arange.h>
 #include <aclnnop/aclnn_argsort.h>
 #include <aclnnop/aclnn_cat.h>
 #include <aclnnop/aclnn_clamp.h>
-#include <aclnnop/aclnn_div.h>
 #include <aclnnop/aclnn_gelu.h>
+#include <aclnnop/aclnn_gelu_v2.h>
+#include <aclnnop/aclnn_sigmoid.h>
 #include <aclnnop/aclnn_hardsigmoid.h>
 #include <aclnnop/aclnn_hardswish.h>
 #include <aclnnop/aclnn_leaky_relu.h>
-#include <aclnnop/aclnn_mul.h>
 #include <aclnnop/aclnn_relu.h>
 #include <aclnnop/aclnn_silu.h>
 #include <aclnnop/aclnn_tanh.h>
+#include <aclnnop/aclnn_sqrt.h>
+#include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_log.h>
+#include <aclnnop/aclnn_sign.h>
 #include "acl_tensor.h"
 #include "common.h"
 
@@ -63,23 +63,6 @@
  */
 void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
-/**
- * @brief   Adds two ggml tensors using the CANN backend.
- *
- * @details This function performs an element-wise addition of two tensors. In
- *          case the tensors do not have the same shape, one or both tensors
- *          will be broadcasted to match the shape of the other before the
- *          addition is performed.The formula for the operation is given by:
- *          \f[
- *              \text{dst} = \text{acl_src0} + \alpha \cdot \text{acl_src1}
- *          \f]
- *
- * @param ctx The CANN context used for operations.
- * @param dst The ggml tensor representing the destination, result of the
- *            addition is stored at dst->data, and dst->op is `GGML_OP_ADD`
- */
-void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst);
-
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
  *          backend.
@@ -131,19 +114,6 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
-/**
- * @brief   Computes the square of the elements of a ggml tensor using the CANN
- *          backend.
- * @details The function sets the second source tensor of the destination
- *          tensor `dst` to be equal to the first source tensor. This is
- *          effectively squaring the elements since the multiplication becomes
- *          `element * element`.
- * @param ctx The CANN context used for operations.
- * @param dst The destination tensor where the squared values will be stored，
- *            which dst->op is `GGML_OP_SQR`.
- */
-void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst);
-
 /**
  * @brief   Applies a clamp operation to the elements of a ggml tensor using the
  *          CANN backend.
@@ -275,6 +245,20 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Computes the sum of elements in a ggml tensor.
+ *
+ * @details This function performs a reduction sum operation along the last
+ *          dimension of the input tensor `src`. The result of the sum is stored
+ *          in the destination tensor `dst`.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the reduced values will be stored。
+ *
+ */
+
+void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
  *          the CANN backend.
@@ -484,109 +468,616 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
-template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
-                                       aclTensor*, uint64_t*, aclOpExecutor**),
-          aclnnStatus execute(void*, uint64_t, aclOpExecutor*, aclrtStream)>
-void ggml_cann_mul_div(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+/**
+ * @brief   Computes the index of the maximum value along the specified dimension
+ *          of a ggml tensor using the CANN backend.
+ *
+ * @details This function performs an argmax operation on the input tensor.
+ *          It finds the index of the maximum value along the specified axis
+ *          and stores these indices in the destination tensor `dst`. The
+ *          operation is executed using the CANN backend for optimized performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the indices of the maximum values will
+ *            be stored. dst->op is `GGML_OP_ARGMAX`.
+ */
+void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief Adds two tensors element-wise and stores the result in a destination
+ * tensor.
+ *
+ * This function performs the operation:
+ * \f[
+ *    dst = acl\_src0 + alpha \times acl\_src1
+ * \f]
+ * where alpha is a scalar value and defaults to 1.0f.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src0 The first source tensor.
+ * @param acl_src1 The second source tensor.
+ * @param acl_dst The destination tensor where the result will be stored.
+ */
+void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
+    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+
+/**
+ * @brief Sub two tensors element-wise and stores the result in a destination
+ * tensor.
+ *
+ * This function performs the operation:
+ * \f[
+ *    dst = acl\_src0 - alpha \times acl\_src1
+ * \f]
+ * where alpha is a scalar value and defaults to 1.0f.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src0 The first source tensor.
+ * @param acl_src1 The second source tensor.
+ * @param acl_dst The destination tensor where the result will be stored.
+ */
+void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
+    aclTensor* acl_src1, aclTensor* acl_dst = nullptr);
+
+/**
+ * @brief Performs element-wise multiplication of two tensors and stores the
+ * result in a destination tensor.
+ *
+ * This function performs element-wise multiplication of the tensors `acl_src`
+ * and `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=\text {acl_src }_i \times \text {acl_other }_i
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The first tensor for element-wise multiplication.
+ * @param acl_other The second tensor for element-wise multiplication.
+ * @param acl_dst The destination tensor where the result will be stored.
+ */
+void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+
+/**
+ * @brief Matrix division, optionally in-place.
+ *
+ * This function division each element of the source tensor `acl_src` by the
+ * tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * If `inplace` is true, `acl_dst` will not be used and the operation is
+ * performed in-place on `acl_src`. The operation is defined as: \f[
+ *     \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src Numerator tensor..
+ * @param acl_other Denominator tensor.
+ * @param acl_dst The destination tensor where the result will be stored if
+ * `inplace` is false.
+ * @param inplace Flag indicating whether to perform the operation in-place on
+ * `acl_src`.
+ */
+void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_other, aclTensor* acl_dst = nullptr);
+
+/**
+ * @brief Applies element-wise cosine function to the elements of a tensor.
+ *
+ * This function computes the cosine of each element in the source tensor
+ * `acl_src` and stores the result in the destination tensor `acl_dst`. The
+ * operation is defined as: \f[ \text {acl_dst }_i=\cos \left(\text {acl_src
+ * }_i\right) \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor on which the cosine function will be
+ * applied.
+ * @param acl_dst The destination tensor where the cosine results will be
+ * stored.
+ */
+void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_dst);
+
+/**
+ * @brief Applies element-wise sine function to the elements of a tensor.
+ *
+ * This function computes the sine of each element in the source tensor
+ `acl_src`
+ * and stores the result in the destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=\sin \left(\text {acl_src }_i\right)
+ * \f]
+
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor on which the sine function will be applied.
+ * @param acl_dst The destination tensor where the sine results will be stored.
+ */
+void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+    aclTensor* acl_dst);
+
+/**
+ * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one
+ * output tensor.
+ *
+ * This function checks whether broadcasting is needed between `src0` and `src1`.
+ * If broadcasting is required, it calculates the proper shapes and creates
+ * ACL tensors with broadcast parameters. Otherwise, it directly creates ACL tensors
+ * based on the original tensor shapes.
+ *
+ * @param src0     The first input tensor (reference shape).
+ * @param src1     The second input tensor (possibly broadcasted).
+ * @param dst      The destination/output tensor.
+ * @param acl_src0 Output pointer to the created ACL tensor corresponding to src0.
+ * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1.
+ * @param acl_dst  Output pointer to the created ACL tensor corresponding to dst.
+ */
+void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst,
+    aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst);
+
+/**
+ * @brief   Computes the 1D transposed convolution (deconvolution) of a ggml
+ * tensor using the CANN backend.
+ *
+ * @details This function performs a 1D transposed convolution (also known as
+ * deconvolution) operation on the input tensor. The computed result is stored
+ * in the destination tensor `dst`. The operation is optimized using the CANN
+ * backend for improved performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the transposed convolution result
+ * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`.
+ */
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies the ELU (Exponential Linear Unit) activation to a ggml tensor
+ * using the CANN backend.
+ *
+ * @details This function performs an element-wise ELU activation on the input
+ *          tensor.
+ *          The result is written to the destination tensor `dst` in-place.
+ *          The ELU function is defined as:
+ *
+ *          \text{ELU}(x) =
+ *          \begin{cases}
+ *          x, & \text{if } x > 0 \\
+ *          \alpha \left( \exp(x) - 1 \right), & \text{if } x \leq 0
+ *          \end{cases}
+ *
+ *          where α (alpha) is a hyperparameter, typically set to 1.0.
+ *          This operation is optimized using the CANN backend for high-performance
+ *          inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the ELU-activated result will be stored.
+ *            dst->op is expected to be `GGML_OP_ELU`.
+ */
+void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the mean of a ggml tensor element-wise using the CANN backend.
+ *
+ * @details This function calculates the element-wise mean of the input tensor.
+ *          The result is written to the destination tensor `dst`.
+ *          The mean is computed by averaging the values across the entire tensor.
+ *
+ *          This operation is optimized using the CANN backend for high-performance inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the mean result will be stored.
+ *            dst->op is expected to be `GGML_OP_MEAN`.
+ */
+void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies 1D reflect padding to a ggml tensor using the CANN backend.
+ *
+ * @details This function performs 1D reflect padding on the input tensor.
+ *          The amount of padding on each side is specified by parameters stored in `dst->op_params`.
+ *          The operation reflects the values at the borders of the tensor to generate the padded output.
+ *
+ *          This operation is optimized using the CANN backend for high-performance inference or training.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the padded result will be stored.
+ *            dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
+ */
+void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Counts the number of equal elements in two ggml tensors using the CANN backend.
+ *
+ * @details This function performs an element-wise comparison between two input tensors,
+ *          and counts the number of positions where the elements are equal. The result is
+ *          stored in the destination tensor `dst` as a scalar.
+ *
+ *          The operation is optimized using the CANN backend, making it suitable for
+ *          high-performance inference or training scenarios.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
+ */
+void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies the Step activation function to a ggml tensor using the CANN backend.
+ *
+ * @details This function applies a step function element-wise to the input tensor, where
+ *          each element is transformed to 1.0 if it is greater than 0, and 0.0 otherwise.
+ *          The result is stored in the destination tensor `dst`.
+ *
+ *          This operation is accelerated using the CANN backend to improve runtime performance.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored.
+ *            dst->op is expected to be `GGML_OP_STEP`.
+ */
+void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/*
+ * @brief A generic wrapper for ACL resources with custom deleter support.
+ */
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+
+/**
+ * @brief Trait structure used to define how to destroy a given ACL resource type.
+ *
+ * @tparam T ACL resource type.
+ */
+template<typename T>
+struct acl_resource_traits;
+
+/**
+ * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
+ */
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
+ */
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
+ */
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
+ */
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
+};
+
+/**
+ * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ *
+ * @tparam T ACL resource type.
+ * @param ptr Raw pointer to ACL resource.
+ * @return any_acl_resource Smart pointer that handles destruction.
+ */
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
+}
+
+/**
+ * @brief Registers multiple ACL resources into a vector for lifetime management.
+ *
+ * @tparam Args Variadic list of ACL resource types.
+ * @param vec Target vector to hold ACL resources.
+ * @param args Raw pointers to ACL resources.
+ */
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+    (vec.emplace_back(make_acl_resource(args)), ...);
+}
+
+/**
+ * @brief Task class that wraps the execution of an aclnn function call.
+ */
+class aclnn_task : public cann_task {
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
+};
+
+/**
+ * @brief Task class that releases ACL resources after usage.
+ */
+class release_resource_task : public cann_task {
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
+
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
+    std::vector<any_acl_resource> resource_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory copy operations.
+ */
+class async_memcpy_task : public cann_task {
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
+    aclrtMemcpyKind kind_;
+    aclrtStream stream_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory set operations.
+ */
+class async_memset_task : public cann_task {
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
+};
+
+/**
+ * @brief Launches an asynchronous task using the memory allocator.
+ *
+ * This macro submit an asynchronous task on the specified stream.
+ * The task uses memory allocated by the allocator. It is guaranteed
+ * that the memory will not be accessed by other tasks until this task
+ * completes, due to the sequential execution order within the same stream.
+ *
+ * @param OP_NAME aclnn operator name.
+ * @param args Additional arguments required by the task.
+ *
+ * @note
+ * Memory from the allocator will be "freed" immediately and can be
+ * reallocated to other pointers. However, it won't be accessed by any
+ * other task before this asynchronous task ends, because all tasks in the
+ * same stream are executed in queue order.
+ */
+
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
+    } while (0)
+
+/**
+ * @brief Registers and releases multiple ACL resources, optionally deferring the release
+ *        using a task.
+ *
+ * @tparam Args Types of the ACL resources.
+ * @param ctx Backend context which manages task submission and async mode.
+ * @param args Pointers to ACL resources to be released.
+ */
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+    std::vector<any_acl_resource> resources;
+    register_acl_resources(resources, std::forward<Args>(args)...);
+    if(ctx.async_mode) {
+        auto task = std::make_unique<release_resource_task>(std::move(resources));
+        ctx.task_queue.submit_task(std::move(task));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param dst Destination memory address.
+ * @param src Source memory address.
+ * @param len Size of memory to copy (in bytes).
+ * @param kind Type of memory copy (host-to-device, device-to-host, etc).
+ */
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
+    }
+}
+
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx->async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
+        ctx->task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param buffer Memory buffer to be set.
+ * @param size Size of the memory buffer (in bytes).
+ * @param value Value to set in the buffer.
+ */
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
+    }
+}
+
+/**
+ * @brief Applies a element-wise operation to two input tensors using the CANN
+ * backend.
+ *
+ * This templated function takes a binary operator and applies it to two source
+ * tensors
+ * associated with the destination tensor. The function handles broadcasting as
+ * needed.
+ *
+ * @tparam binary_op A callable object (e.g., lambda or function pointer) representing
+ *         the binary operation to be performed. It must take three arguments:
+ *         (ggml_backend_cann_context&, aclTensor*, aclTensor*, aclTensor*).
+ *
+ * @param ctx The CANN backend context used to manage execution and resources.
+ * @param dst The destination tensor.
+ */
+template <auto binary_op>
+void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];
     ggml_tensor* src1 = dst->src[1];
-    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     aclTensor* acl_src0;
     aclTensor* acl_src1;
     aclTensor* acl_dst;
 
     // Need bcast
-    if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) {
-        BCAST_SHAPE(src0, src1)
-        acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0));
-        acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1));
-        acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0));
-    } else {
-        acl_src0 = ggml_cann_create_tensor(src0);
-        acl_src1 = ggml_cann_create_tensor(src1);
-        acl_dst = ggml_cann_create_tensor(dst);
-    }
+    bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
+    binary_op(ctx, acl_src0, acl_src1, acl_dst);
 
-    uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
-    void* workspaceAddr = nullptr;
-
-    ACL_CHECK(getWorkspaceSize(acl_src0, acl_src1, acl_dst, &workspaceSize,
-                               &executor));
-    if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
-        workspaceAddr = workspace_allocator.get();
-    }
-
-    aclrtStream main_stream = ctx.stream();
-    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
-
-    ACL_CHECK(aclDestroyTensor(acl_src0));
-    ACL_CHECK(aclDestroyTensor(acl_src1));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
-// Activation functions template.
-template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
-                                       aclOpExecutor**),
-          aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
-                              const aclrtStream)>
-void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src = dst->src[0];
 
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+/**
+ * @brief Applies a unary operation to an input tensor using the CANN backend.
+ *
+ * This templated function applies a unary operator to the source tensor of `dst`
+ * and stores the result in the destination tensor.
+ *
+ * @tparam unary_op A callable with the signature:
+ *         void(ggml_backend_cann_context&, aclTensor*, aclTensor*)
+ *         where the first aclTensor is the source and the second is the destination.
+ * @param ctx The CANN backend context for managing resources and execution.
+ * @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) {
+    ggml_tensor* src = dst->src[0];
 
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
-    void* workspaceAddr = nullptr;
-
-    ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
-    if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
-        workspaceAddr = workspace_allocator.get();
-    }
-
-    aclrtStream main_stream = ctx.stream();
-    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
-
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    unary_op(ctx, acl_src, acl_dst);
+    ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
-// Activation functions template for const aclTensors.
-template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
-                                       uint64_t*, aclOpExecutor**),
-          aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
-                              const aclrtStream)>
-void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    ggml_tensor* src = dst->src[0];
-
-    GGML_ASSERT(src->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    aclTensor* acl_src = ggml_cann_create_tensor(src);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
-
-    uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
-    void* workspaceAddr = nullptr;
-
-    ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
-    if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
-        workspaceAddr = workspace_allocator.get();
-    }
-
-    aclrtStream main_stream = ctx.stream();
-    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
-
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
-}
+/**
+ * @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.
+ *
+ * @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]`.
+ */
+void ggml_cann_unary_op(
+    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.
+ *
+ * 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.
+ *
+ * Internally, the lambda will call:
+ * @code
+ * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
+ * @endcode
+ *
+ * @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_CALL_ACLNN_OP
+ */
+#define GGML_CANN_CALL_UNARY_OP(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);                         \
+    }                                                                 \
+    while (0)
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -31,9 +31,16 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <unistd.h>
+#include <functional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
+#include "../ggml-impl.h"
 
 #define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
@@ -205,6 +212,127 @@ struct ggml_cann_pool_alloc {
     ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };
 
+/**
+ * @brief Function pointer type for ACLNN operator calls.
+ */
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+
+/**
+ * @brief Base class for all CANN tasks to be submitted to the task queue.
+ *
+ * Users should override the run_task() method with actual task logic.
+ */
+class cann_task {
+public:
+    virtual void run_task() {}
+};
+
+/**
+ * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
+ */
+class cann_task_queue {
+public:
+    /**
+     * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
+     *
+     * @param capacity Queue capacity. Must be a power of 2.
+     * @param device Target device ID (used for context setting).
+     */
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
+        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
+        mask_ = capacity_ - 1;
+    }
+
+    /**
+     * @brief Attempts to enqueue a task into the queue.
+     *
+     * @param item Unique pointer to the task.
+     * @return true if the task was successfully enqueued, false if the queue was full.
+     */
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
+        size_t next_tail = (tail_ + 1) & mask_;
+
+        if (next_tail == head_) {
+            return false;
+        }
+
+        buffer_[tail_] = std::move(item);
+        std::atomic_thread_fence(std::memory_order_release);
+        tail_ = next_tail;
+
+        return true;
+    }
+
+    /**
+     * @brief Submits a task to the queue, and starts the worker thread if not already running.
+     *
+     * @param task Task to be submitted.
+     */
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
+            std::this_thread::yield();
+            continue;
+        }
+
+        if (!running_) {
+            running_ = true;
+            thread_ = std::thread(&cann_task_queue::execute, this);
+        }
+
+    }
+
+    /**
+     * @brief Waits until the queue is completely empty and no tasks are being processed.
+     */
+    void wait() {
+        while (running_ && head_ != tail_) {
+            std::this_thread::yield();
+            continue;
+        }
+    }
+
+    /**
+     * @brief Stops the task queue and joins the worker thread.
+     */
+    void stop() {
+        running_ = false;
+        if (thread_.joinable()) {
+            thread_.join();
+        }
+    }
+
+private:
+    /**
+     * @brief Worker thread function that continuously dequeues and executes tasks.
+     */
+    void execute() {
+        ggml_cann_set_device(device_);
+
+        while (running_) {
+            if(head_ == tail_) {
+                std::this_thread::yield();
+                continue;
+            }
+
+            std::atomic_thread_fence(std::memory_order_acquire);
+            buffer_[head_]->run_task();
+            buffer_[head_].reset();
+            head_ = (head_ + 1) & mask_;
+        }
+    }
+
+    std::vector<std::unique_ptr<cann_task>> buffer_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -213,6 +341,8 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    cann_task_queue task_queue;
+    bool async_mode;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -221,9 +351,12 @@ struct ggml_backend_cann_context {
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
     }
 
     /**
@@ -231,6 +364,7 @@ struct ggml_backend_cann_context {
      */
     ~ggml_backend_cann_context() {
         ggml_cann_set_device(device);
+        task_queue.stop();
         if (copy_event != nullptr) {
             ACL_CHECK(aclrtDestroyEvent(copy_event));
         }

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

@@ -29,6 +29,8 @@
 #include <cstdio>
 #include <cstring>
 #include <mutex>
+#include <queue>
+#include <chrono>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -119,9 +121,10 @@ static ggml_cann_device_info ggml_cann_init() {
         prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
         prop.location.id = id;
         prop.reserve = 0;
-        ACL_CHECK(aclrtMemGetAllocationGranularity(
+        err = aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
-            &info.devices[id].vmm_granularity));
+            &info.devices[id].vmm_granularity);
+        info.devices[id].vmm = err == ACL_SUCCESS;
 
         size_t free, total;
         ggml_backend_cann_get_device_memory(id, &free, &total);
@@ -148,11 +151,223 @@ const ggml_cann_device_info& ggml_cann_info() {
 
 //#define DEBUG_CANN_MALLOC
 /**
- * @brief A pool of CANN buffers(legacy).
+ * @brief A pool of CANN buffers(priority segment buffer).
  *
  * This class manages a pool of CANN buffers for a specific device.
  */
-struct ggml_cann_pool_leg : public ggml_cann_pool {
+struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer.
+        size_t size = 0;      ///< Size of the buffer.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
+
+        bool operator>(const ggml_cann_buffer& other) const {
+            return size > other.size;
+        }
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    std::unordered_map<void*, size_t> buffer_pool;
+    std::priority_queue<ggml_cann_buffer,
+                        std::vector<ggml_cann_buffer>,
+                        std::greater<>> free_buffers ;
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_buf_prio(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_buf_prio() {
+        ggml_cann_set_device(device);
+        for (auto& [b_ptr, b_size] : buffer_pool) {
+            aclrtFree(b_ptr);
+            pool_size -= b_size;
+        }
+        buffer_pool.clear();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        std::vector<ggml_cann_buffer> free_buffers_rest;
+        free_buffers_rest.reserve(free_buffers.size());
+        while (!free_buffers.empty()) {
+            auto b = free_buffers.top();
+            free_buffers.pop();
+
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    ptr = b.ptr;
+#ifdef DEBUG_CANN_MALLOC
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
+#endif
+                    break;
+                }
+            }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+                buffer_pool.erase(b.ptr);
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                continue;
+            }
+            free_buffers_rest.push_back(b);
+        }
+        for (ggml_cann_buffer &b : free_buffers_rest) {
+            free_buffers.push(std::move(b));
+        }
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+        if (ptr != nullptr) {
+            return ptr;
+        }
+
+        // allocate a new buffer if no buffer can be reused
+        ggml_cann_set_device(device);
+        ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: allocate %p, "
+            "pool_size = %5u MB, "
+            "size = %5u MB\n",
+            device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+            (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
+#endif
+        buffer_pool.emplace(ptr, size);
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
+        auto it = buffer_pool.find(ptr);
+        if (it == buffer_pool.end()) {
+            GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
+        }
+
+        auto now = std::chrono::steady_clock::now();
+        free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: return   %p, "
+            "pool_size = %5u MB\n",
+            device, ptr,
+            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers(segment buffer).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_buf : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
     /**
      * @brief The maximum number of buffers in the pool.
      */
@@ -163,12 +378,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      */
     int device;
 
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
     /**
      * @brief Structure representing a CANN buffer.
      */
     struct ggml_cann_buffer {
         void* ptr = nullptr;  ///< Pointer to the buffer memory.
         size_t size = 0;      ///< Size of the buffer.
+        bool used = false;    ///< Whether the buffer is currently in use.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
     };
 
     /**
@@ -186,17 +408,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      *
      * @param device The device ID to associate with this buffer pool.
      */
-    explicit ggml_cann_pool_leg(int device) : device(device) {}
+    explicit ggml_cann_pool_buf(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
 
     /**
      * @brief Destructor to free all buffers in the pool.
      */
-    ~ggml_cann_pool_leg() {
+    ~ggml_cann_pool_buf() {
         ggml_cann_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
             if (b.ptr != nullptr) {
-                ACL_CHECK(aclrtFree(b.ptr));
+                aclrtFree(b.ptr);
                 pool_size -= b.size;
             }
         }
@@ -212,63 +436,93 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
-        const size_t alignment = 128;
         size = GGML_PAD(size, alignment);
         if (size == 0) {
             size = alignment;
         }
-#ifdef DEBUG_CANN_MALLOC
-        int nnz = 0;
-        size_t max_size = 0;
-#endif
-        size_t best_diff = 1ull << 36;
-        int ibest = -1;
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        int i = 0;
+        for (; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
+            if (b.ptr == nullptr) {
+                break;
+            }
+            if (b.used) {
+                continue;
+            }
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    b.used = true;
+                    ptr = b.ptr;
 #ifdef DEBUG_CANN_MALLOC
-                ++nnz;
-                if (b.size > max_size) max_size = b.size;
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
 #endif
-                if (b.size >= size) {
-                    size_t diff = b.size - size;
-                    if (diff < best_diff) {
-                        best_diff = diff;
-                        ibest = i;
-                        if (!best_diff) {
-                            void* ptr = b.ptr;
-                            *actual_size = b.size;
-                            b.ptr = nullptr;
-                            b.size = 0;
-                            return ptr;
-                        }
-                    }
+                    break;
                 }
             }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                b.ptr = nullptr;
+            }
         }
-        if (ibest >= 0) {
-            ggml_cann_buffer& b = buffer_pool[ibest];
-            void* ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
+        if (ptr != nullptr) {
             return ptr;
         }
-        void* ptr;
-        ggml_cann_set_device(device);
-        ACL_CHECK(
-            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = size;
-        pool_size += size;
+
+        if (i < MAX_BUFFERS) {
+            // allocate a new buffer if no buffer can be reused
+            ggml_cann_buffer& b = buffer_pool[i];
+            ggml_cann_set_device(device);
+            ACL_CHECK(aclrtMalloc(&b.ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+            pool_size += size;
+            *actual_size = size;
+            b.size = size;
+            b.used = true;
+            if (i >= MAX_BUFFERS - 8) {
+                GGML_LOG_WARN("cann pool[%d]: slots almost full\n", device);
+            }
 #ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO(
-            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
-            "requested %u MB\n",
-            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
-            (uint32_t)(pool_size / 1024 / 1024),
-            (uint32_t)(size / 1024 / 1024));
+            GGML_LOG_INFO(
+                "cann pool[%d]: allocate %p, "
+                "pool_size = %5u MB, "
+                "size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
 #endif
-        return ptr;
+            return b.ptr;
+        }
+
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 
     /**
@@ -278,18 +532,24 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @param size Size of the buffer to free.
      */
     void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr == nullptr) {
-                b.ptr = ptr;
-                b.size = size;
-                return;
+            if (b.ptr != ptr) {
+                continue;
             }
+            b.used = false;
+            b.last_used = std::chrono::steady_clock::now();
+#ifdef DEBUG_CANN_MALLOC
+            GGML_LOG_INFO(
+                "cann pool[%d]: return   %p, "
+                "pool_size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+            return;
         }
-        // memory should always buffered. these memory may still needed by
-        // tasks in stream.
-        // TODO, fix me.
-        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 };
 
@@ -347,8 +607,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_vmm(int device)
-        : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+    : device(device) {
         auto dev = ggml_cann_info().devices[device];
         granularity = dev.vmm_granularity;
         max_size = dev.total_vram;
@@ -471,7 +730,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
+    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
+    if (enable_buf_prio) {
+        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
+    }
+    GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
 
 // cann buffer
@@ -803,7 +1073,7 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
         return GGML_STATUS_SUCCESS;
     }
 
-    // TODO: can backend doesn't support quantized yet. Just leave the code
+    // TODO: cann backend doesn't support quantized yet. Just leave the code
     // here.
     if (ggml_is_quantized(tensor->type)) {
         // Initialize padding to 0 to avoid possible NaN values
@@ -1020,8 +1290,11 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
     ggml_cann_set_device(buft_ctx->device);
 
-    size = std::max(size, (size_t)1);
-
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
     void* dev_ptr;
     aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
     if (err != ACL_SUCCESS) {
@@ -1300,47 +1573,69 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_dup(ctx, dst);
             break;
         case GGML_OP_ADD:
-            ggml_cann_add(ctx, dst);
+        case GGML_OP_ADD1:
+            ggml_cann_binary_op<aclnn_add>(ctx, dst);
+            break;
+        case GGML_OP_SUB:
+            ggml_cann_binary_op<aclnn_sub>(ctx, dst);
             break;
         case GGML_OP_ACC:
             ggml_cann_acc(ctx, dst);
             break;
         case GGML_OP_MUL:
-            ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
+            ggml_cann_binary_op<aclnn_mul>(ctx, dst);
             break;
         case GGML_OP_DIV:
-            ggml_cann_mul_div<aclnnDivGetWorkspaceSize, aclnnDiv>(ctx, dst);
+            ggml_cann_binary_op<aclnn_div>(ctx, dst);
             break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_ABS:
+                    GGML_CANN_CALL_UNARY_OP(Abs);
+                    break;
+                case GGML_UNARY_OP_NEG:
+                    GGML_CANN_CALL_UNARY_OP(Neg);
+                    break;
                 case GGML_UNARY_OP_GELU:
-                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
-                        ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Gelu);
                     break;
                 case GGML_UNARY_OP_SILU:
-                    ggml_cann_activation<aclnnSiluGetWorkspaceSize, aclnnSilu>(
-                        ctx, dst);
-                    break;
-                // TODO: Use faster gelu??
-                case GGML_UNARY_OP_GELU_QUICK:
-                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
-                        ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Silu);
                     break;
+                case GGML_UNARY_OP_GELU_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_unary_op(lambda, ctx, dst);
+                } break;
                 case GGML_UNARY_OP_TANH:
-                    ggml_cann_activation<aclnnTanhGetWorkspaceSize, aclnnTanh>(
-                        ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Tanh);
                     break;
                 case GGML_UNARY_OP_RELU:
-                    ggml_cann_activation<aclnnReluGetWorkspaceSize, aclnnRelu>(
-                        ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Relu);
+                    break;
+                case GGML_UNARY_OP_SIGMOID:
+                    GGML_CANN_CALL_UNARY_OP(Sigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSIGMOID:
-                    ggml_cann_activation<aclnnHardsigmoidGetWorkspaceSize,
-                                         aclnnHardsigmoid>(ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Hardsigmoid);
                     break;
                 case GGML_UNARY_OP_HARDSWISH:
-                    ggml_cann_activation<aclnnHardswishGetWorkspaceSize,
-                                         aclnnHardswish>(ctx, dst);
+                    GGML_CANN_CALL_UNARY_OP(Hardswish);
+                    break;
+                case GGML_UNARY_OP_EXP:
+                    GGML_CANN_CALL_UNARY_OP(Exp);
+                    break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_cann_elu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SGN:
+                    GGML_CANN_CALL_UNARY_OP(Sign);
+                    break;
+                case GGML_UNARY_OP_STEP:
+                    ggml_cann_step(ctx, dst);
                     break;
                 default:
                     return false;
@@ -1382,7 +1677,12 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_scale(ctx, dst);
             break;
         case GGML_OP_SQR:
-            ggml_cann_sqr(ctx, dst);
+            GGML_ASSERT(dst->src[1] == nullptr);
+            dst->src[1] = dst->src[0];
+            ggml_cann_binary_op<aclnn_mul>(ctx, dst);
+            break;
+        case GGML_OP_SQRT:
+            GGML_CANN_CALL_UNARY_OP(Sqrt);
             break;
         case GGML_OP_CLAMP:
             ggml_cann_clamp(ctx, dst);
@@ -1414,12 +1714,39 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
         case GGML_OP_POOL_2D:
             ggml_cann_pool2d(ctx, dst);
             break;
+        case GGML_OP_SUM:
+            ggml_cann_sum(ctx, dst);
+            break;
         case GGML_OP_SUM_ROWS:
             ggml_cann_sum_rows(ctx, dst);
             break;
         case GGML_OP_ARGSORT:
             ggml_cann_argsort(ctx, dst);
             break;
+        case GGML_OP_ARGMAX:
+            ggml_cann_argmax(ctx, dst);
+            break;
+        case GGML_OP_COS:
+            ggml_cann_unary_op<aclnn_cos>(ctx, dst);
+            break;
+        case GGML_OP_SIN:
+            ggml_cann_unary_op<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);
+            break;
+        case GGML_OP_MEAN:
+            ggml_cann_mean(ctx, dst);
+            break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_cann_pad_reflect_1d(ctx, dst);
+            break;
+        case GGML_OP_COUNT_EQUAL:
+            ggml_cann_count_equal(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -1458,21 +1785,15 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     ACL_CHECK(aclrtSynchronizeDevice());
     ACL_CHECK(aclrtResetDevice(cann_ctx->device));
 
-    // finalize when last backend freed.
-    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
-        ACL_CHECK(aclFinalize());
-    }
-
     delete cann_ctx;
     delete backend;
 }
 
+
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
+ * This function asynchronously sets tensor data in the CANN backend.
  *
  * @param backend Pointer to the CANN backend structure.
  * @param tensor Pointer to the tensor structure to set data for.
@@ -1487,23 +1808,28 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
                                                size_t size) {
     ggml_backend_cann_context *cann_ctx =
         (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+        "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
+    ggml_cann_async_memcpy(cann_ctx, (char *)tensor->data + offset, data, size,
+        ACL_MEMCPY_HOST_TO_DEVICE);
 }
 
+/**
+ * @brief Gets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously gets tensor data in the CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to get data from.
+ * @param data Pointer to the host data to copy from the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
 static void ggml_backend_cann_get_tensor_async(
     ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
@@ -1514,20 +1840,11 @@ static void ggml_backend_cann_get_tensor_async(
 
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
                 "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
+
+    ggml_cann_async_memcpy(cann_ctx, data, (char *)tensor->data + offset, size,
+        ACL_MEMCPY_DEVICE_TO_HOST);
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
 }
 
 /**
@@ -1587,6 +1904,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
 
+        // wait for task_queue empty to keep task order.
+        cann_ctx_src->task_queue.wait();
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
@@ -1614,9 +1933,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
+    cann_ctx->task_queue.wait();
     ggml_cann_set_device(cann_ctx->device);
-
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
@@ -1675,24 +1993,38 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_STEP:
                     return true;
                 default:
                     return false;
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                case GGML_TYPE_Q8_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
-                case GGML_TYPE_Q4_0:
                     return true;
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
+                    // only support contiguous for quantized types.
+                    return ggml_is_contiguous(op->src[0]) &&
+                            ggml_is_contiguous(op->src[1]);
                 default:
                     return false;
             }
@@ -1738,13 +2070,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
-            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            float ext_factor = 0.0f;
+            memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
             }
             // TODO: ext_factor != 0
-            if (*ext_factor != 0) {
+            if (ext_factor != 0) {
                 return false;
             }
 
@@ -1756,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 return false;
             }
 
+            if(!ggml_is_contiguous(op->src[0])){
+                return false;
+            }
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -1764,8 +2100,28 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
                 return false;
             }
+            if (op->op_params[0] != GGML_SCALE_MODE_NEAREST) {
+                return false;
+            }
             return true;
         }
+        case GGML_OP_POOL_2D: {
+            const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
+            const int       k0   = opts[1];
+            const int       k1   = opts[2];
+            const int       p0   = opts[5];
+            const int       p1   = opts[6];
+            // value of paddingH should be at most half of kernelH
+            // value of paddingW should be at most half of kernelW
+            return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
+        }
+        case GGML_OP_SUM:
         case GGML_OP_DUP:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:
@@ -1777,15 +2133,17 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
         case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_SUB:
         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_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
@@ -1794,6 +2152,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_ARGMAX:
+        case GGML_OP_COS:
+        case GGML_OP_SIN:
+        case GGML_OP_CONV_TRANSPOSE_1D:
+        case GGML_OP_LOG:
+        case GGML_OP_MEAN:
+        case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_COUNT_EQUAL:
             return true;
         default:
             return false;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -222,7 +222,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             elseif (GGML_AVX)
                 list(APPEND ARCH_FLAGS /arch:AVX)
                 list(APPEND ARCH_DEFINITIONS GGML_AVX)
-            else ()
+            elseif (GGML_SSE42)
                 list(APPEND ARCH_FLAGS /arch:SSE4.2)
                 list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
@@ -237,8 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_NATIVE)
                 list(APPEND ARCH_FLAGS -march=native)
             else ()
-                list(APPEND ARCH_FLAGS -msse4.2)
-                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_SSE42)
+                    list(APPEND ARCH_FLAGS -msse4.2)
+                    list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                endif()
                 if (GGML_F16C)
                     list(APPEND ARCH_FLAGS -mf16c)
                     list(APPEND ARCH_DEFINITIONS GGML_F16C)

ggml/src/ggml-cpu/cpu-feats-x86.cpp

@@ -263,7 +263,7 @@ void test_x86_is() {
 static int ggml_backend_cpu_x86_score() {
     // FIXME: this does not check for OS support
 
-    int score = 0;
+    int score = 1;
     cpuid_x86 is;
 
 #ifdef GGML_FMA

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -4,13 +4,13 @@
 
 #include "ggml.h"
 #include "ggml-impl.h"
+
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 //#include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
 #include <math.h>   // fabsf
 
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -69,33 +69,16 @@ struct ggml_compute_params {
 #endif
 
 #if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
 #include <sys/prctl.h>
 #endif
 
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
 #if defined(__ARM_NEON)
 
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
+// ref: https://github.com/ggml-org/llama.cpp/pull/5404
 #ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
 #define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
 #else
-
-typedef __fp16 ggml_fp16_internal_t;
-
 #define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
 #endif // _MSC_VER
 
 #if !defined(__aarch64__)
@@ -340,8 +323,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #else
 #ifdef __POWER9_VECTOR__
 #include <altivec.h>
-#undef bool
-#define bool _Bool
 #else
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1932,6 +1932,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D_DW:
+            {
+                ggml_compute_forward_conv_2d_dw(params, tensor);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 ggml_compute_forward_conv_transpose_2d(params, tensor);
@@ -2027,41 +2031,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
-        case GGML_OP_MAP_UNARY:
-            {
-                ggml_unary_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_unary(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_BINARY:
-            {
-                ggml_binary_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_binary(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM1_F32:
-            {
-                ggml_custom1_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1_f32(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM2_F32:
-            {
-                ggml_custom2_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2_f32(params, tensor, fun);
-            }
-            break;
-        case GGML_OP_MAP_CUSTOM3_F32:
-            {
-                ggml_custom3_op_f32_t fun;
-                memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3_f32(params, tensor, fun);
-            }
-            break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2077,6 +2046,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_map_custom3(params, tensor);
             }
             break;
+        case GGML_OP_CUSTOM:
+            {
+                ggml_compute_forward_custom(params, tensor);
+            }
+            break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 ggml_compute_forward_cross_entropy_loss(params, tensor);
@@ -2298,6 +2272,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
@@ -2328,11 +2303,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
-        case GGML_OP_MAP_UNARY:
-        case GGML_OP_MAP_BINARY:
-        case GGML_OP_MAP_CUSTOM1_F32:
-        case GGML_OP_MAP_CUSTOM2_F32:
-        case GGML_OP_MAP_CUSTOM3_F32:
             {
                 n_tasks = 1;
             } break;
@@ -2366,6 +2336,16 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     n_tasks = MIN(p.n_tasks, n_threads);
                 }
             } break;
+        case GGML_OP_CUSTOM:
+            {
+                struct ggml_custom_op_params p;
+                memcpy(&p, node->op_params, sizeof(p));
+                if (p.n_tasks == GGML_N_TASKS_MAX) {
+                    n_tasks = n_threads;
+                } else {
+                    n_tasks = MIN(p.n_tasks, n_threads);
+                }
+            } break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -425,6 +425,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
         }
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
+        case GGML_OP_GET_ROWS_BACK:
+            return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
         case GGML_OP_OUT_PROD:
             return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
                 src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;

ggml/src/ggml-cpu/ops.cpp

@@ -6064,6 +6064,178 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
+// ggml_compute_forward_conv_2d_dw
+
+struct ggml_conv_2d_dw_params {
+    int64_t channels;
+    int64_t batch;
+    int64_t src_w;
+    int64_t src_h;
+    int64_t dst_w;
+    int64_t dst_h;
+    int64_t knl_w;
+    int64_t knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+};
+
+static void ggml_compute_forward_conv_2d_dw_cwhn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t c = p.channels;
+    const float * knl_data = (const float *)kernel->data;
+
+    const int64_t rows_total = p.dst_h * p.batch;
+    const int64_t rows_per_thread = (rows_total + params->nth - 1) / params->nth;
+    const int64_t row_start = params->ith * rows_per_thread;
+    const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
+
+#ifdef GGML_SIMD
+    const int64_t pkg_size = GGML_F32_EPR;
+    const int64_t pkg_count = c / pkg_size;
+    const int64_t c_pkg_end = pkg_count * pkg_size;
+#else
+    const int64_t c_pkg_end = 0;
+#endif
+
+    for (int64_t row = row_start; row < row_end; ++row) {
+        const int64_t dst_y = row % p.dst_h;
+        const float * src_data = (const float *)src->data + (row / p.dst_h) * p.src_w * p.src_h * c;
+        for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+            float * dst_data = (float *)dst->data + (row * p.dst_w + dst_x) * c;
+            const int64_t src_y_base = dst_y * p.stride_y - p.pad_y;
+            const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;
+
+#ifdef GGML_SIMD
+            // Vectorized loop
+            for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {
+                GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(knl_data + (knl_y * p.knl_w + knl_x) * c + c_i);
+                        GGML_F32_VEC s = GGML_F32_VEC_LOAD(src_data + (src_y * p.src_w + src_x) * c + c_i);
+                        sum = GGML_F32_VEC_FMA(sum, k, s);
+                    }
+                }
+                GGML_F32_VEC_STORE(dst_data + c_i, sum);
+            }
+#endif
+            // Scalar loop
+            for (int64_t c_i = c_pkg_end; c_i < c; ++c_i) {
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[(knl_y * p.knl_w + knl_x) * c + c_i]
+                             * src_data[(src_y * p.src_w + src_x) * c + c_i];
+                    }
+                }
+                dst_data[c_i] = sum;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_2d_dw_whcn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t n = p.channels * p.batch;
+    const int64_t per_thread = (n + params->nth - 1) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end = MIN(start + per_thread, n);
+
+    for (int64_t i = start; i < end; ++i) {
+        const float * knl_data = (const float *)kernel->data + (i % p.channels) * p.knl_w * p.knl_h;
+        const float * src_data = (const float *)src->data + i * p.src_w * p.src_h;
+        float * dst_data = (float *)dst->data + i * p.dst_w * p.dst_h;
+
+        for (int64_t dst_y = 0; dst_y < p.dst_h; ++dst_y) {
+            for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[knl_y * p.knl_w + knl_x]
+                             * src_data[src_y * p.src_w + src_x];
+                    }
+                }
+                dst_data[dst_y * p.dst_w + dst_x] = sum;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_2d_dw(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * src = dst->src[1];
+    ggml_conv_2d_dw_params p;
+    p.channels = src->ne[2];
+    p.batch = src->ne[3];
+    p.src_w = src->ne[0];
+    p.src_h = src->ne[1];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.knl_w = kernel->ne[0];
+    p.knl_h = kernel->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(kernel->ne[3] == p.channels);
+    GGML_ASSERT(dst->ne[3] == p.batch);
+
+    if (ggml_is_contiguous(src)) {
+        ggml_compute_forward_conv_2d_dw_whcn(params, src, kernel, dst, p);
+    } else if (ggml_is_contiguous_channels(src)) {
+        // kernel should also have channels most contiguous in memory
+        GGML_ASSERT(kernel->nb[0] >= kernel->nb[2] && kernel->nb[1] >= kernel->nb[0]);
+        ggml_compute_forward_conv_2d_dw_cwhn(params, src, kernel, dst, p);
+    } else {
+        GGML_ABORT("non-contiguous memory layout not supported");
+    }
+}
+
 // ggml_compute_forward_pool_1d_sk_p0
 
 static void ggml_compute_forward_pool_1d_sk_p0(
@@ -6351,24 +6523,72 @@ static void ggml_compute_forward_upscale_f32(
     const float sf2 = (float)ne2/src0->ne[2];
     const float sf3 = (float)ne3/src0->ne[3];
 
-    // TODO: optimize
+    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
 
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        const int64_t i03 = i3 / sf3;
-        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
-            const int64_t i02 = i2 / sf2;
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                const int64_t i01 = i1 / sf1;
-                for (int64_t i0 = 0; i0 < ne0; i0++) {
-                    const int64_t i00 = i0 / sf0;
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const int64_t i01 = i1 / sf1;
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const int64_t i00 = i0 / sf0;
 
-                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
+                        const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
 
-                    *y = *x;
+                        *y = *x;
+                    }
                 }
             }
         }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        // setting a pixel offset of 0 would replicate the behavior of pytorch interpolate with align_corners=True
+        const float pixel_offset = 0.5f;
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    int64_t y0 = (int64_t)floorf(y);
+                    int64_t y1 = y0 + 1;
+
+                    y0 = std::max(int64_t(0), std::min(y0, ne01 - 1));
+                    y1 = std::max(int64_t(0), std::min(y1, ne01 - 1));
+
+                    float dy = y - (float)y0;
+                    dy = std::max(0.0f, std::min(dy, 1.0f));
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        int64_t x0 = (int64_t)floorf(x);
+                        int64_t x1 = x0 + 1;
+
+                        x0 = std::max(int64_t(0), std::min(x0, ne00 - 1));
+                        x1 = std::max(int64_t(0), std::min(x1, ne00 - 1));
+
+                        float dx = x - (float)x0;
+                        dx = std::max(0.0f, std::min(dx, 1.0f));
+
+                        // fetch the four surrounding pixel values and interpolate
+                        const float a = *(const float *)((const char *)src0->data + x0*nb00 + y0*nb01 + i02*nb02 + i03*nb03);
+                        const float b = *(const float *)((const char *)src0->data + x1*nb00 + y0*nb01 + i02*nb02 + i03*nb03);
+                        const float c = *(const float *)((const char *)src0->data + x0*nb00 + y1*nb01 + i02*nb02 + i03*nb03);
+                        const float d = *(const float *)((const char *)src0->data + x1*nb00 + y1*nb01 + i02*nb02 + i03*nb03);
+
+                        const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
+
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else {
+        GGML_ABORT("unsupported upscale mode");
     }
 }
 
@@ -6721,8 +6941,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     ggml_vec_dot_t    const kq_vec_dot     = ggml_get_type_traits_cpu(k->type)->vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
 
-    GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type");
-    GGML_ASSERT(v_to_float   && "fattn: unsupported V-type");
+    GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
+    GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
@@ -6818,10 +7038,14 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                     vs = expf(s - M);
                 }
 
-                v_to_float(v_data, V32, DV);
-
                 // V += v*expf(s - M)
-                ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                if (v_to_float) {
+                    v_to_float(v_data, V32, DV);
+                    ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                } else {
+                    // V is F32
+                    ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs);
+                }
             }
 
             S = S*ms + vs; // scale and increment sum with partial sum
@@ -8264,152 +8488,6 @@ void ggml_compute_forward_rwkv_wkv7(
     }
 }
 
-// ggml_compute_forward_map_unary
-
-static void ggml_compute_forward_map_unary_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_unary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, dst));
-
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
-
-    for (int i = 0; i < n; i++) {
-        fun(nc,
-                (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])));
-    }
-}
-
-void ggml_compute_forward_map_unary(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_unary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_unary_f32(params, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_map_binary
-
-static void ggml_compute_forward_map_binary_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_binary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(src1));
-    assert(ggml_is_contiguous_1(dst));
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
-
-    const int n  = ggml_nrows(src0);
-    const int nc = src0->ne[0];
-
-    for (int i = 0; i < n; i++) {
-        fun(nc,
-                (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])),
-                (float *) ((char *) src1->data + i*(src1->nb[1])));
-    }
-}
-
-void ggml_compute_forward_map_binary(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_binary_op_f32_t fun) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_binary_f32(params, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
-// ggml_compute_forward_map_custom1
-
-void ggml_compute_forward_map_custom1_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom1_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a);
-}
-
-// ggml_compute_forward_map_custom2
-
-void ggml_compute_forward_map_custom2_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom2_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-    const ggml_tensor * b = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a, b);
-}
-
-// ggml_compute_forward_map_custom3
-
-void ggml_compute_forward_map_custom3_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst,
-        const ggml_custom3_op_f32_t fun) {
-
-    const ggml_tensor * a = dst->src[0];
-    const ggml_tensor * b = dst->src[1];
-    const ggml_tensor * c = dst->src[1];
-
-    if (params->ith != 0) {
-        return;
-    }
-
-    fun(dst, a, b, c);
-}
-
 // ggml_compute_forward_map_custom1
 
 void ggml_compute_forward_map_custom1(
@@ -8455,6 +8533,18 @@ void ggml_compute_forward_map_custom3(
     p.fun(dst, a, b, c, params->ith, params->nth, p.userdata);
 }
 
+// ggml_compute_forward_custom
+
+void ggml_compute_forward_custom(
+    const struct ggml_compute_params * params,
+          struct ggml_tensor * dst) {
+
+    struct ggml_custom_op_params p;
+    memcpy(&p, dst->op_params, sizeof(p));
+
+    p.fun(dst, params->ith, params->nth, p.userdata);
+}
+
 // ggml_compute_forward_cross_entropy_loss
 
 static void ggml_compute_forward_cross_entropy_loss_f32(

ggml/src/ggml-cpu/ops.h

@@ -65,6 +65,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -96,29 +97,10 @@ void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params,
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_map_unary(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_unary_op_f32_t fun);
-void ggml_compute_forward_map_binary(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_binary_op_f32_t fun);
-void ggml_compute_forward_map_custom1_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom1_op_f32_t fun);
-void ggml_compute_forward_map_custom2_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom2_op_f32_t fun);
-void ggml_compute_forward_map_custom3_f32(
-    const struct ggml_compute_params * params,
-    struct ggml_tensor * dst,
-    const ggml_custom3_op_f32_t fun);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/simd-mappings.h

@@ -71,7 +71,7 @@
     #define GGML_F16x8              float16x8_t
     #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
     #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
-    #define GGML_F16x8_LOAD(x)      vld1q_f16((const ggml_fp16_internal_t *)(x))
+    #define GGML_F16x8_LOAD(x)      vld1q_f16((const __fp16 *)(x))
     #define GGML_F16x8_STORE        vst1q_f16
     #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
     #define GGML_F16x8_ADD          vaddq_f16
@@ -99,7 +99,7 @@
     #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
     #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
     #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i])
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
     #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
     #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
     #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
@@ -114,7 +114,7 @@
     #define GGML_F32Cx4              float32x4_t
     #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
     #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
-    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const ggml_fp16_internal_t *)(x)))
+    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
     #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
     #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
     #define GGML_F32Cx4_ADD          vaddq_f32
@@ -125,7 +125,7 @@
     #define GGML_F16_VEC_ZERO           GGML_F32Cx4_ZERO
     #define GGML_F16_VEC_SET1           GGML_F32Cx4_SET1
     #define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx4_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((ggml_fp16_internal_t *)(p), r[i])
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
     #define GGML_F16_VEC_FMA            GGML_F32Cx4_FMA
     #define GGML_F16_VEC_ADD            GGML_F32Cx4_ADD
     #define GGML_F16_VEC_MUL            GGML_F32Cx4_MUL
@@ -392,7 +392,11 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ?                   \
   vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \
   vec_extract_fp32_from_shortl(vec_xl(0, p))
-#define GGML_ENDIAN_BYTE(i) ((unsigned char *)&(uint16_t){1})[i]
+static inline unsigned char ggml_endian_byte(int i) {
+       uint16_t tmp_val = 1;
+       return ((unsigned char *)&tmp_val)[i];
+}
+#define GGML_ENDIAN_BYTE(i) ggml_endian_byte(i)
 #define GGML_F16_VEC_STORE(p, r, i)                             \
   if (i & 0x1)                                                  \
     vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)],  \
@@ -851,13 +855,17 @@ static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
         tmp[i] = GGML_FP16_TO_FP32(x[i]);
     }
 
-    return vec_xl(0, tmp);
+    // note: keep type-cast here to prevent compiler bugs
+    // see: https://github.com/ggml-org/llama.cpp/issues/12846
+    return vec_xl(0, (const float *)(tmp));
 }
 
 static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
     float arr[4];
 
-    vec_xst(y, 0, arr);
+    // note: keep type-cast here to prevent compiler bugs
+    // see: https://github.com/ggml-org/llama.cpp/issues/12846
+    vec_xst(y, 0, (float *)(arr));
 
     for (int i = 0; i < 4; i++) {
         x[i] = GGML_FP32_TO_FP16(arr[i]);

ggml/src/ggml-cuda/common.cuh

@@ -729,7 +729,13 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    std::vector<char **> updated_kernel_arg;
+    bool use_cpy_indirection = false;
+    std::vector<char *> cpy_dest_ptrs;
+    char ** dest_ptrs_d;
+    int dest_ptrs_size = 0;
+    // Index to allow each cpy kernel to be aware of it's position within the graph
+    // relative to other cpy nodes.
+    int graph_cpynode_index = -1;
 #endif
 };
 

ggml/src/ggml-cuda/convert.cu

@@ -579,7 +579,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
 
     const src_t * x = (const src_t *) vx;
 
-    y[i] = x[i];
+    y[i] = float(x[i]);
 }
 
 template <typename src_t, typename dst_t>
@@ -588,6 +588,17 @@ static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict_
     convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
+to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half>;
+        default:
+            return nullptr;
+    }
+}
+
 to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -633,6 +644,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cuda<nv_bfloat16>;
         default:
             return nullptr;
     }

ggml/src/ggml-cuda/convert.cuh

@@ -7,7 +7,10 @@ using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, in
 
 typedef to_t_cuda_t<float> to_fp32_cuda_t;
 typedef to_t_cuda_t<half> to_fp16_cuda_t;
+typedef to_t_cuda_t<nv_bfloat16> to_bf16_cuda_t;
 
 to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
 
+to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
+
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);

ggml/src/ggml-cuda/cpy.cu

@@ -10,6 +10,13 @@ static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
     *dsti = *xi;
 }
 
+static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
+
+    *dsti = *xi;
+}
+
 static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
     const float * xi = (const float *) cxi;
     half * dsti = (half *) cdsti;
@@ -32,16 +39,18 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_1>
-static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
                                    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                   const int nb12, const int nb13) {
+                                   const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
     // then combine those indices with the corresponding byte offsets to get the total offsets
     const int64_t i03 = i/(ne00 * ne01 * ne02);
@@ -288,16 +297,18 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -314,16 +325,18 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
 }
 
 template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
+static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
                                  const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                 const int nb12, const int nb13) {
+                                 const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
 
     if (i >= ne) {
         return;
     }
 
+    char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct;
+
     const int i03 = i/(ne00 * ne01 * ne02);
     const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
     const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
@@ -339,66 +352,97 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
+// Copy destination pointers to GPU to be available when pointer indirection is in use
+
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) {
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+        if (cuda_graph->dest_ptrs_d != nullptr) {
+            CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d));
+        }
+        CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *)));
+        cuda_graph->dest_ptrs_size = host_dest_ptrs_size;
+    }
+    // copy destination pointers to GPU
+    CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
+    cuda_graph->graph_cpynode_index = 0; // reset index
+#else
+    GGML_UNUSED(cuda_graph); GGML_UNUSED(host_dest_ptrs);
+    GGML_UNUSED(host_dest_ptrs_size); GGML_UNUSED(stream);
+#endif
+}
+
 static void ggml_cpy_f16_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
+}
+
+static void ggml_cpy_f32_bf16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q8_0_f32_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_0_f32_cuda(
@@ -407,22 +451,22 @@ static void ggml_cpy_q4_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q4_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q4_1_f32_cuda(
@@ -431,22 +475,22 @@ static void ggml_cpy_q4_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+         ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_0_f32_cuda(
@@ -455,22 +499,22 @@ static void ggml_cpy_q5_0_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_q5_1_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_q5_1_f32_cuda(
@@ -479,35 +523,35 @@ static void ggml_cpy_q5_1_f32_cuda(
     const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12,
     const int nb10, const int nb11, const int nb12, const int nb13,
-    cudaStream_t stream) {
+    cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
     const int num_blocks = ne;
     cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
         cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f32_iq4_nl_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
 static void ggml_cpy_f16_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -541,51 +585,70 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
 
+    char ** dest_ptrs_d = nullptr;
+    int graph_cpynode_index = -1;
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
+        dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
+        graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
+    }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
+#endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
         CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
-        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
-            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+            nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
-        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
-        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
+#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
+        ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
+    }
+#endif
+
 }
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
 }
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
@@ -593,6 +656,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         return (void*) cpy_f32_f16<cpy_1_f32_f16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {

ggml/src/ggml-cuda/cpy.cuh

@@ -2,8 +2,10 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
+
+void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream);

ggml/src/ggml-cuda/fattn-common.cuh

@@ -62,7 +62,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
     T sum = 0.0f;
 
 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
         const int k_KQ = k_KQ_0 + threadIdx.x;
 
         const int ib    = k_KQ /  QI8_1;
@@ -102,7 +102,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1(
     T sum = 0.0f;
 
 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
         const int k_KQ = k_KQ_0 + threadIdx.x;
 
         const int ib    = k_KQ /  QI8_1;
@@ -146,7 +146,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0(
     T sum = 0.0f;
 
 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
         const int k_KQ = k_KQ_0 + threadIdx.x;
 
         const int ib    = k_KQ /  QI8_1;
@@ -193,7 +193,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
     T sum = 0.0f;
 
 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
         const int k_KQ = k_KQ_0 + threadIdx.x;
 
         const int ib    = k_KQ /  QI8_1;
@@ -244,7 +244,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
     T sum = 0.0f;
 
 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += warp_size) {
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
         const int k_KQ = k_KQ_0 + threadIdx.x;
 
         const int ib  = k_KQ / QI8_0;

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -52,6 +52,18 @@ static __global__ void flash_attn_tile_ext_f32(
     return;
 #endif // FP16_MMA_AVAILABLE
     if (use_logit_softcap && !(D == 128 || D == 256)) {
+        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
+        GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
+        GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
+        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
+        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
+        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
+        GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
+        GGML_UNUSED(ne2); GGML_UNUSED(ne3);
         NO_DEVICE_CODE;
         return;
     }