release : v1.4.1

* whisper : debug

* whisper : fix bug during quantization

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.4.0)
+project(whisper.cpp VERSION 1.4.1)
 
 if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
     add_compile_options(/utf-8)

README.md

@@ -6,7 +6,7 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Beta: [v1.4.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.0) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Beta: [v1.4.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.1) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.4.0",
+  "version": "1.4.1",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

examples/common-ggml.cpp

@@ -90,7 +90,7 @@ bool ggml_common_quantize_0(
         }
 
         int32_t nelements = 1;
-        int32_t ne[2] = { 1, 1 };
+        int32_t ne[4] = { 1, 1, 1, 1 };
         for (int i = 0; i < n_dims; ++i) {
             finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
             nelements *= ne[i];
@@ -99,7 +99,7 @@ bool ggml_common_quantize_0(
         std::string name(length, 0);
         finp.read (&name[0], length);
 
-        printf("%64s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ggml_type_name((ggml_type) ttype));
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
 
         bool quantize = false;
 
@@ -204,11 +204,11 @@ bool ggml_common_quantize_0(
             total_size_new += cur_size;
 
             printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
-            for (int i = 0; i < hist_cur.size(); ++i) {
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
                 hist_all[i] += hist_cur[i];
             }
 
-            for (int i = 0; i < hist_cur.size(); ++i) {
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
                 printf("%5.3f ", hist_cur[i] / (float)nelements);
             }
             printf("\n");
@@ -226,12 +226,12 @@ bool ggml_common_quantize_0(
 
     {
         int64_t sum_all = 0;
-        for (int i = 0; i < hist_all.size(); ++i) {
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
             sum_all += hist_all[i];
         }
 
         printf("%s: hist: ", __func__);
-        for (int i = 0; i < hist_all.size(); ++i) {
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
             printf("%5.3f ", hist_all[i] / (float)sum_all);
         }
         printf("\n");

ggml.c

@@ -1911,8 +1911,8 @@ static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, in
             const uint8_t vi = pp[l/2];
 
             // extract the 5-th bit from qh
-            const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-            const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+            const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+            const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
             const int8_t vi0 = (vi & 0x0F) | vh0;
             const int8_t vi1 = (vi >>   4) | vh1;
@@ -1948,8 +1948,8 @@ static void dequantize_row_q5_1(const void * restrict vx, float * restrict y, in
             const uint8_t vi = pp[l/2];
 
             // extract the 5-th bit from qh
-            const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-            const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+            const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+            const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
             const uint8_t vi0 = (vi & 0x0F) | vh0;
             const uint8_t vi1 = (vi >>   4) | vh1;
@@ -3286,8 +3286,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         for (int j = 0; j < QK8_0/2; j++) {
             const uint8_t v0 = x0[j];
 
-            const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
-            const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
+            const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
+            const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
 
             const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16;
             const int x1_0 = ((v0 >>   4) | x1_0h) - 16;
@@ -3491,8 +3491,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         for (int j = 0; j < QK8_1/2; j++) {
             const uint8_t v0 = x0[j];
 
-            const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
-            const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
+            const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
+            const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
 
             const int x0_0 = (v0 & 0x0F) | x0_0h;
             const int x1_0 = (v0 >>   4) | x1_0h;
@@ -13057,8 +13057,8 @@ size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int l = 0; l < QK5_0; l += 2) {
-                const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-                const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+                const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+                const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;
@@ -13087,8 +13087,8 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int l = 0; l < QK5_1; l += 2) {
-                const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-                const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+                const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+                const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;

whisper.cpp

@@ -1333,7 +1333,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             }
 
             int32_t nelements = 1;
-            int32_t ne[3] = { 1, 1, 1 };
+            int32_t ne[4] = { 1, 1, 1, 1 };
             for (int i = 0; i < n_dims; ++i) {
                 read_safe(loader, ne[i]);
                 nelements *= ne[i];
@@ -1352,6 +1352,8 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             auto tensor = model.tensors[name.data()];
             if (ggml_nelements(tensor) != nelements) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                fprintf(stderr, "%s: shape: [%d, %d, %d], expected: [%d, %d, %d]\n",
+                        __func__, ne[0], ne[1], ne[2], (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2]);
                 return false;
             }