ggml : Q4_3c using 2x "Full range" approach

Untested

.github/workflows/build.yml

@@ -12,7 +12,7 @@ on:
       - master
     paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
   pull_request:
-    types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
+    types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
 
 env:
@@ -20,8 +20,6 @@ env:
 
 jobs:
   ubuntu-latest-make:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     steps:
@@ -41,8 +39,6 @@ jobs:
           make
 
   ubuntu-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     steps:
@@ -71,8 +67,6 @@ jobs:
           ctest --verbose
 
   ubuntu-latest-cmake-sanitizer:
-    if: github.event.pull_request.draft == false
-
     runs-on: ubuntu-latest
 
     continue-on-error: true
@@ -108,8 +102,6 @@ jobs:
           ctest --verbose
 
   macOS-latest-make:
-    if: github.event.pull_request.draft == false
-
     runs-on: macos-latest
 
     steps:
@@ -128,8 +120,6 @@ jobs:
           make
 
   macOS-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: macOS-latest
 
     steps:
@@ -157,8 +147,6 @@ jobs:
           ctest --verbose
 
   windows-latest-cmake:
-    if: github.event.pull_request.draft == false
-
     runs-on: windows-latest
 
     strategy:

CMakeLists.txt

@@ -312,6 +312,7 @@ add_library(ggml OBJECT
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
@@ -324,6 +325,7 @@ add_library(llama
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
 target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)

examples/main/main.cpp

@@ -25,6 +25,7 @@
 #endif
 
 static console_state con_st;
+static llama_context ** g_ctx;
 
 static bool is_interacting = false;
 
@@ -36,6 +37,7 @@ void sigint_handler(int signo) {
         if (!is_interacting) {
             is_interacting=true;
         } else {
+            llama_print_timings(*g_ctx);
             _exit(130);
         }
     }
@@ -94,6 +96,7 @@ int main(int argc, char ** argv) {
 //bool is_prime(int n) {)";
 
     llama_context * ctx;
+    g_ctx = &ctx;
 
     // load the model
     {

ggml-cuda.cu

@@ -31,8 +31,8 @@ static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2
 
 #define QK4_3 16
 typedef struct {
-    __half  d;              // delta
-    __half  m;              // min
+    __half  d0;             // delta
+    __half  d1;             // delta
     uint8_t qs[QK4_3 / 2];  // nibbles / quants
 } block_q4_3;
 static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
@@ -112,22 +112,32 @@ static __global__ void dequantize_block_q4_3(const void * vx, float * y) {
 
     const int i = blockIdx.x;
 
-    const float d = x[i].d;
-    const float m = x[i].m;
+    const float d0 = x[i].d0;
+    const float d1 = x[i].d1;
 
     const uint8_t * pp = x[i].qs;
 
-    for (int l = 0; l < QK4_3; l += 2) {
-        const uint8_t vi = pp[l/2];
+    for (int l = 0; l < QK4_3/2; l += 2) {
+        const uint8_t vi0 = pp[l/2];
+        const uint8_t vi1 = pp[l/2 + QK4_3/4];
 
-        const int8_t vi0 = vi & 0xf;
-        const int8_t vi1 = vi >> 4;
+        const int8_t vi0_0 = vi0 & 0xf;
+        const int8_t vi0_1 = vi0 >> 4;
 
-        const float v0 = vi0*d + m;
-        const float v1 = vi1*d + m;
+        const int8_t vi1_0 = vi1 & 0xf;
+        const int8_t vi1_1 = vi1 >> 4;
 
-        y[i*QK4_3 + l + 0] = v0;
-        y[i*QK4_3 + l + 1] = v1;
+        const float v0_0 = (vi0_0 - 8)*d0;
+        const float v0_1 = (vi0_1 - 8)*d0;
+
+        const float v1_0 = (vi1_0 - 8)*d1;
+        const float v1_1 = (vi1_1 - 8)*d1;
+
+        y[i*QK4_3 + l + 0] = v0_0;
+        y[i*QK4_3 + l + 1] = v0_1;
+
+        y[i*QK4_3 + l + 0 + QK4_3/2] = v1_0;
+        y[i*QK4_3 + l + 1 + QK4_3/2] = v1_1;
     }
 }
 

ggml.c

@@ -655,8 +655,8 @@ static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2
 
 #define QK4_3 16
 typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // min
+    ggml_fp16_t d0;         // delta
+    ggml_fp16_t d1;         // min
     uint8_t qs[QK4_3 / 2]; // nibbles / quants
 } block_q4_3;
 static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
@@ -680,13 +680,17 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r
 
     for (int i = 0; i < nb; i++) {
         float amax = 0.0f; // absolute max
+        float max = 0.0f;
 
         for (int l = 0; l < QK4_0; l++) {
             const float v = x[i*QK4_0 + l];
-            amax = MAX(amax, fabsf(v));
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+                max = v;
+            }
         }
 
-        const float d = amax / ((1 << 3) - 1);
+        const float d = max / -8;
         const float id = d ? 1.0f/d : 0.0f;
 
         y[i].d = d;
@@ -695,8 +699,8 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r
             const float v0 = x[i*QK4_0 + l + 0]*id;
             const float v1 = x[i*QK4_0 + l + 1]*id;
 
-            const uint8_t vi0 = (int8_t)roundf(v0) + 8;
-            const uint8_t vi1 = (int8_t)roundf(v1) + 8;
+            const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8);
+            const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8);
 
             assert(vi0 < 16);
             assert(vi1 < 16);
@@ -716,28 +720,42 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
 
 #if defined(__POWER9_VECTOR__)
     const vector float v85 = vec_splats(8.5f);
+    const vector signed int v15 = vec_splats(15);
     for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
+        float max = 0.0f;
+        float min = 0.0f;
 
         vector float srcv [8];
-        vector float asrcv[8];
-        vector float amaxv[8];
+        vector float maxv[8];
+        vector float minv[8];
 
         for (int l = 0; l < 8; l++) srcv[l]  = *(vector float *)(x + i*32 + 4*l);
-        for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]);
+        //for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]);
 
-        for (int l = 0; l < 4; l++) amaxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]);
-        //for (int l = 0; l < 2; l++) amaxv[4*l] = vec_max(amaxv[4*l], amaxv[4*l+2]);
-        amaxv[0] = vec_max(amaxv[0], amaxv[2]);
-        amaxv[4] = vec_max(amaxv[4], amaxv[6]);
-        //for (int l = 0; l < 1; l++) amaxv[8*l] = vec_max(amaxv[8*l], amaxv[8*l+4]);
-        amaxv[0] = vec_max(amaxv[0], amaxv[4]);
+        for (int l = 0; l < 4; l++) maxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]);
+        //for (int l = 0; l < 2; l++) maxv[4*l] = vec_max(maxv[4*l], maxv[4*l+2]);
+        maxv[0] = vec_max(maxv[0], maxv[2]);
+        maxv[4] = vec_max(maxv[4], maxv[6]);
+        //for (int l = 0; l < 1; l++) maxv[8*l] = vec_max(maxv[8*l], maxv[8*l+4]);
+        maxv[0] = vec_max(maxv[0], maxv[4]);
 
-        amax = MAX(
-                MAX(vec_extract(amaxv[0], 0), vec_extract(amaxv[0], 1)),
-                MAX(vec_extract(amaxv[0], 2), vec_extract(amaxv[0], 3)));
+        for (int l = 0; l < 4; l++) minv[2*l] = vec_min(asrcv[2*l], asrcv[2*l+1]);
+        //for (int l = 0; l < 2; l++) minv[4*l] = vec_min(minv[4*l], minv[4*l+2]);
+        minv[0] = vec_min(minv[0], minv[2]);
+        minv[4] = vec_min(minv[4], minv[6]);
+        //for (int l = 0; l < 1; l++) minv[8*l] = vec_min(minv[8*l], minv[8*l+4]);
+        minv[0] = vec_min(minv[0], minv[4]);
 
-        const float d = amax / ((1 << 3) - 1);
+
+        max = MAX(
+                MAX(vec_extract(maxv[0], 0), vec_extract(maxv[0], 1)),
+                MAX(vec_extract(maxv[0], 2), vec_extract(maxv[0], 3)));
+        min = MIN(
+                MIN(vec_extract(minv[0], 0), vec_extract(minv[0], 1)),
+                MIN(vec_extract(minv[0], 2), vec_extract(minv[0], 3)));
+
+        const float magnitude = max >= fabsf(min) ? max : min;
+        const float d = magnitude / -8;
         const float id = d ? 1.0/d : 0.0;
 
         y[i].d = d;
@@ -747,27 +765,33 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         for (int l = 0; l < 8; l++) {
             const vector float vf  = vec_madd(srcv[l], vid, v85);
             const vector signed int vi = vec_signed(vf);
+            const vector signed int vc = vec_min(vi, v15);
 
-            pb[2*l + 0] = vec_extract(vi, 0) | (vec_extract(vi, 1) << 4);
-            pb[2*l + 1] = vec_extract(vi, 2) | (vec_extract(vi, 3) << 4);
+            pb[2*l + 0] = vec_extract(vc, 0) | (vec_extract(vc, 1) << 4);
+            pb[2*l + 1] = vec_extract(vc, 2) | (vec_extract(vc, 3) << 4);
         }
     }
 #elif __ARM_NEON
     for (int i = 0; i < nb; i++) {
         float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
+        float32x4_t maxv[8];
+        float32x4_t minv[8];
 
         for (int l = 0; l < 8; l++) srcv[l]  = vld1q_f32(x + i*32 + 4*l);
-        for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]);
 
-        for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]);
-        for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]);
-        for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]);
+        for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l+1]);
+        for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l+2]);
+        for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l+4]);
 
-        const float amax = vmaxvq_f32(amaxv[0]);
+        for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l+1]);
+        for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l+2]);
+        for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l+4]);
 
-        const float d = amax / ((1 << 3) - 1);
+        const float max = vmaxvq_f32(maxv[0]);
+        const float min = vminvq_f32(minv[0]);
+
+        const float magnitude = max >= fabsf(min) ? max : min;
+        const float d = magnitude / -8;
         const float id = d ? 1.0f/d : 0.0f;
 
         y[i].d = d;
@@ -776,9 +800,10 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
             const float32x4_t v  = vmulq_n_f32(srcv[l], id);
             const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f));
             const int32x4_t   vi = vcvtq_s32_f32(vf);
+            const int32x4_t   vc = vminq_s32(vi, vdupq_n_s32(15));
 
-            y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
-            y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
+            y[i].qs[2*l + 0] = vgetq_lane_s32(vc, 0) | (vgetq_lane_s32(vc, 1) << 4);
+            y[i].qs[2*l + 1] = vgetq_lane_s32(vc, 2) | (vgetq_lane_s32(vc, 3) << 4);
         }
     }
 #elif defined(__AVX2__)
@@ -790,22 +815,31 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         __m256 v3 = _mm256_loadu_ps( x + 24 );
         x += 32;
 
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+        // Compute max for the block
+        __m256 max  = _mm256_max_ps( v0, v1 );
+        __m256 maxTmp = _mm256_max_ps( v2, v3 );
+        max = _mm256_max_ps( max, maxTmp );
 
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) );
         max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
         max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
         const float maxScalar = _mm_cvtss_f32( max4 );
 
+        // Compute min for the block
+        __m256 min  = _mm256_min_ps( v0, v1 );
+        __m256 minTmp = _mm256_min_ps( v2, v3 );
+        min = _mm256_min_ps( min, minTmp );
+
+        __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) );
+        min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) );
+        min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) );
+        const float minScalar = _mm_cvtss_f32( min4 );
+
         // Quantize these floats
-        const float d = maxScalar / 7.0f;
+        const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar;
+        const float d = magnitude / -8.0f;
         y[i].d = d;
-        const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f;
+        const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
         // Apply the multiplier
@@ -838,9 +872,11 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
         i0 = _mm256_permutevar8x32_epi32( i0, perm );
 
-        // Apply offset to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ]
+        // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ]
         const __m256i off = _mm256_set1_epi8( 8 );
         i0 = _mm256_add_epi8( i0, off );
+        const __m256i maxNibble = _mm256_set1_epi8( 15 );
+        i0 = _mm256_min_epi8( i0, maxNibble );
 
         // Compress the vector into 4 bit/value, and store
         __m128i res = packNibbles( i0 );
@@ -855,22 +891,31 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         __m256 v3 = _mm256_loadu_ps( x + 24 );
         x += 32;
 
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+        // Compute max for the block
+        __m256 max  = _mm256_max_ps( v0, v1 );
+        __m256 maxTmp = _mm256_max_ps( v2, v3 );
+        max = _mm256_max_ps( max, maxTmp );
 
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) );
         max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
         max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
         const float maxScalar = _mm_cvtss_f32( max4 );
 
+        // Compute min for the block
+        __m256 min  = _mm256_min_ps( v0, v1 );
+        __m256 minTmp = _mm256_min_ps( v2, v3 );
+        min = _mm256_min_ps( min, minTmp );
+
+        __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) );
+        min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) );
+        min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) );
+        const float minScalar = _mm_cvtss_f32( min4 );
+
         // Quantize these floats
-        const float d = maxScalar / 7.0f;
+        const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar;
+        const float d = magnitude / -8.0f;
         y[i].d = d;
-        const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f;
+        const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
         // Apply the multiplier
@@ -911,10 +956,13 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         ni0 = _mm_packs_epi16( ni0, ni2 );
         ni4 = _mm_packs_epi16( ni4, ni6 );
 
-        // Apply offset to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ]
-        const __m128i off = _mm_set1_epi8( 8);
+        // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ]
+        const __m128i off = _mm_set1_epi8( 8 );
         ni0 = _mm_add_epi8( ni0, off );
         ni4 = _mm_add_epi8( ni4, off );
+        const __m128i maxNibble = _mm_set1_epi8( 15 );
+        ni0 = _mm_min_epi8( ni0, maxNibble );
+        ni4 = _mm_min_epi8( ni4, maxNibble );
 
         // Compress the vector into 4 bit/value, and store
         __m128i res = packNibbles( ni0, ni4 );
@@ -922,24 +970,32 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
     }
 #elif defined(__wasm_simd128__)
     for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
+        float max = 0.0f;
+        float min = 0.0f;
 
         v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
+        v128_t maxv[8];
+        v128_t minv[8];
 
         for (int l = 0; l < 8; l++) srcv[l]  = wasm_v128_load(x + i*32 + 4*l);
-        for (int l = 0; l < 8; l++) asrcv[l] = wasm_f32x4_abs(srcv[l]);
 
-        for (int l = 0; l < 4; l++) amaxv[2*l] = wasm_f32x4_max(asrcv[2*l], asrcv[2*l+1]);
-        for (int l = 0; l < 2; l++) amaxv[4*l] = wasm_f32x4_max(amaxv[4*l], amaxv[4*l+2]);
-        for (int l = 0; l < 1; l++) amaxv[8*l] = wasm_f32x4_max(amaxv[8*l], amaxv[8*l+4]);
+        for (int l = 0; l < 4; l++) maxv[2*l] = wasm_f32x4_max(srcv[2*l], srcv[2*l+1]);
+        for (int l = 0; l < 2; l++) maxv[4*l] = wasm_f32x4_max(maxv[4*l], maxv[4*l+2]);
+        for (int l = 0; l < 1; l++) maxv[8*l] = wasm_f32x4_max(maxv[8*l], maxv[8*l+4]);
 
-        amax = MAX(
-                MAX(wasm_f32x4_extract_lane(amaxv[0], 0), wasm_f32x4_extract_lane(amaxv[0], 1)),
-                MAX(wasm_f32x4_extract_lane(amaxv[0], 2), wasm_f32x4_extract_lane(amaxv[0], 3)));
+        for (int l = 0; l < 4; l++) minv[2*l] = wasm_f32x4_min(srcv[2*l], srcv[2*l+1]);
+        for (int l = 0; l < 2; l++) minv[4*l] = wasm_f32x4_min(minv[4*l], minv[4*l+2]);
+        for (int l = 0; l < 1; l++) minv[8*l] = wasm_f32x4_min(minv[8*l], minv[8*l+4]);
 
-        const float d = amax / ((1 << 3) - 1);
+        max = MAX(
+                MAX(wasm_f32x4_extract_lane(maxv[0], 0), wasm_f32x4_extract_lane(maxv[0], 1)),
+                MAX(wasm_f32x4_extract_lane(maxv[0], 2), wasm_f32x4_extract_lane(maxv[0], 3)));
+        min = MIN(
+                MIN(wasm_f32x4_extract_lane(minv[0], 0), wasm_f32x4_extract_lane(minv[0], 1)),
+                MIN(wasm_f32x4_extract_lane(minv[0], 2), wasm_f32x4_extract_lane(minv[0], 3)));
+
+        const float magnitude = max >= fabsf(min) ? max : min;
+        const float d = magnitude / -8;
         const float id = d ? 1.0/d : 0.0;
 
         y[i].d = d;
@@ -948,9 +1004,10 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
             const v128_t v  = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id));
             const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f));
             const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf);
+            const v128_t vc = wasm_i32x4_min_u(vi, wasm_i32x4_splat(15));
 
-            y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vi, 0) | (wasm_i32x4_extract_lane(vi, 1) << 4);
-            y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vi, 2) | (wasm_i32x4_extract_lane(vi, 3) << 4);
+            y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4);
+            y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4);
         }
     }
 #else
@@ -1131,13 +1188,17 @@ static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * r
 
     for (int i = 0; i < nb; i++) {
         float amax = 0.0f; // absolute max
+        float max  = 0.0f;
 
         for (int l = 0; l < QK4_2; l++) {
             const float v = x[i*QK4_2 + l];
-            amax = MAX(amax, fabsf(v));
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+                max = v;
+            }
         }
 
-        const float d = amax / ((1 << 3) - 1);
+        const float d = max / -8;
 
         const float id = d ? 1.0f/d : 0.0f;
 
@@ -1147,8 +1208,8 @@ static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * r
             const float v0 = x[i*QK4_2 + l + 0]*id;
             const float v1 = x[i*QK4_2 + l + 1]*id;
 
-            const uint8_t vi0 = (uint8_t)(v0 + 8.5f);
-            const uint8_t vi1 = (uint8_t)(v1 + 8.5f);
+            const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8);
+            const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8);
 
             assert(vi0 < 16);
             assert(vi1 < 16);
@@ -1158,93 +1219,12 @@ static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * r
     }
 }
 
-static inline int nearest_int(float fval) {
-    assert(fval <= 4194303.f);
-    float val = fval + 12582912.f;
-    int i; memcpy(&i, &val, sizeof(int));
-    return (i & 0x007fffff) - 0x00400000;
-}
-
-static float kquantize_q4_with_bounds(int n, int nmin, int nmax, const float * restrict X, int nCandidates,
-        const float * restrict candidates, int8_t * restrict L) {
-    assert (nmin >= INT8_MIN);
-    assert (nmax <= INT8_MAX);
-    float amax = 0;
-    for (int i=0; i<n; ++i) amax = MAX(amax, fabsf(X[i]));
-    if (!amax) { // all zero
-        for (int i=0; i<n; ++i) L[i] = 0;
-        return 1.f;
-    }
-    float best = 0, bestScale = 0;
-    for (int si=0; si<nCandidates; ++si) {
-        float iscale = candidates[si]/amax;
-        float sumlxP = 0; int suml2P = 0;
-        float sumlxM = 0; int suml2M = 0;
-        for (int i=0; i<n; ++i) {
-            int l = nearest_int(iscale*X[i]);
-            int lp = MAX(nmin, MIN(nmax, +l));
-            int lm = MAX(nmin, MIN(nmax, -l));
-            sumlxP += X[i]*lp; suml2P += lp*lp;
-            sumlxM += X[i]*lm; suml2M += lm*lm;
-        }
-        float sumlxP2 = sumlxP*sumlxP;
-        float sumlxM2 = sumlxM*sumlxM;
-        if (sumlxP2*suml2M > sumlxM2*suml2P) {
-            if (sumlxP2 > best*suml2P) {
-                best = sumlxP2/suml2P; bestScale = iscale;
-            }
-        } else {
-            if (sumlxM2 > best*suml2M) {
-                best = sumlxM2/suml2M; bestScale = -iscale;
-            }
-        }
-    }
-    float sumlx = 0; int suml2 = 0;
-    for (int i=0; i<n; ++i) {
-        int l = nearest_int(bestScale*X[i]);
-        l = MAX(nmin, MIN(nmax, l));
-        sumlx += X[i]*l; suml2 += l*l;
-        L[i] = l;
-    }
-    float scale = sumlx/suml2;
-    return scale;
-}
-
-static void quantize_row_q4_2_rmse(const float * restrict x, block_q4_2 * restrict y, int k) {
-#define CANDIDATE_COUNT 8
-    static const float candidates[CANDIDATE_COUNT] = { +8.7f, +8.3f, +8.1f, +7.8f, +7.3f, +7.0f, +6.3f, +5.7f };
-    assert(k % QK4_2 == 0);
-
-    int8_t L[QK4_2];
-
-    const int nb = k / QK4_2;
-
-    for (int i = 0; i < nb; i++) {
-        float scale = kquantize_q4_with_bounds(QK4_2, -8, 7, x, CANDIDATE_COUNT, candidates, L);
-        y[i].d = GGML_FP32_TO_FP16(scale);
-
-        for (int l = 0; l < QK4_2; l += 2) {
-            const uint8_t vi0 = (uint8_t)(L[l+0] + 8);
-            const uint8_t vi1 = (uint8_t)(L[l+1] + 8);
-
-            assert(vi0 < 16);
-            assert(vi1 < 16);
-
-            y[i].qs[l/2] = vi0 | (vi1 << 4);
-        }
-
-        x += QK4_2;
-    }
-}
-
 static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK4_2 == 0);
 
     block_q4_2 * restrict y = vy;
 
-    //quantize_row_q4_2_reference(x, y, k);
-    // This produces the exact same format, just better match to the input floats ("better" as measured by RMSE)
-    quantize_row_q4_2_rmse(x, y, k);
+    quantize_row_q4_2_reference(x, y, k);
 }
 
 static void quantize_row_q4_3_reference(const float * restrict x, block_q4_3 * restrict y, int k) {
@@ -1252,32 +1232,50 @@ static void quantize_row_q4_3_reference(const float * restrict x, block_q4_3 * r
     const int nb = k / QK4_3;
 
     for (int i = 0; i < nb; i++) {
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
+        float amax0 = 0.0f;
+        float max0  = 0.0f;
+        float amax1 = 0.0f;
+        float max1  = 0.0f;
 
-        for (int l = 0; l < QK4_3; l++) {
-            const float v = x[i*QK4_3 + l];
-            if (v < min) min = v;
-            if (v > max) max = v;
+        for (int l = 0; l < QK4_3/2; l++) {
+            const float v0 = x[i*QK4_3 + l];
+            const float v1 = x[i*QK4_3 + l + QK4_3/2];
+
+            if (amax0 < fabsf(v0)) {
+                amax0 = fabsf(v0);
+                max0 = v0;
+            }
+
+            if (amax1 < fabsf(v1)) {
+                amax1 = fabsf(v1);
+                max1 = v1;
+            }
         }
 
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
+        const float d0 = max0 / -8;
+        const float d1 = max1 / -8;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
-        y[i].m = GGML_FP32_TO_FP16(min);
+        const float id0 = d0 ? 1.0f/d0 : 0.0f;
+        const float id1 = d1 ? 1.0f/d1 : 0.0f;
 
-        for (int l = 0; l < QK4_3; l += 2) {
-            const float v0 = (x[i*QK4_3 + l + 0] - min)*id;
-            const float v1 = (x[i*QK4_3 + l + 1] - min)*id;
+        y[i].d0 = GGML_FP32_TO_FP16(d0);
+        y[i].d1 = GGML_FP32_TO_FP16(d1);
 
-            const uint8_t vi0 = (int) (v0 + 0.5f);
-            const uint8_t vi1 = (int) (v1 + 0.5f);
+        for (int l = 0; l < QK4_3/2; l += 2) {
+            const float v0_0 = x[i*QK4_3 + l + 0]*id0;
+            const float v0_1 = x[i*QK4_3 + l + 1]*id0;
 
-            assert(vi0 < 16);
-            assert(vi1 < 16);
+            const float v1_0 = x[i*QK4_3 + l + 0 + QK4_3/2]*id1;
+            const float v1_1 = x[i*QK4_3 + l + 1 + QK4_3/2]*id1;
 
-            y[i].qs[l/2] = vi0 | (vi1 << 4);
+            const uint8_t vi0_0 = MIN(15, (int8_t)roundf(v0_0) + 8);
+            const uint8_t vi0_1 = MIN(15, (int8_t)roundf(v0_1) + 8);
+
+            const uint8_t vi1_0 = MIN(15, (int8_t)roundf(v1_0) + 8);
+            const uint8_t vi1_1 = MIN(15, (int8_t)roundf(v1_1) + 8);
+
+            y[i].qs[l/2          ] = vi0_0 | (vi0_1 << 4);
+            y[i].qs[l/2 + QK4_3/4] = vi1_0 | (vi1_1 << 4);
         }
     }
 }
@@ -1749,25 +1747,32 @@ static void dequantize_row_q4_3(const void * restrict vx, float * restrict y, in
     const block_q4_3 * restrict x = vx;
 
     for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float m = GGML_FP16_TO_FP32(x[i].m);
+        const float d0 = GGML_FP16_TO_FP32(x[i].d0);
+        const float d1 = GGML_FP16_TO_FP32(x[i].d1);
 
         const uint8_t * restrict pp = x[i].qs;
 
-        for (int l = 0; l < QK4_3; l += 2) {
-            const uint8_t vi = pp[l/2];
+        for (int l = 0; l < QK4_3/2; l += 2) {
+            const uint8_t vi0 = pp[l/2];
+            const uint8_t vi1 = pp[l/2 + QK4_3/4];
 
-            const int8_t vi0 = vi & 0xf;
-            const int8_t vi1 = vi >> 4;
+            const int8_t vi0_0 = vi0 & 0xf;
+            const int8_t vi0_1 = vi0 >> 4;
 
-            const float v0 = vi0*d + m;
-            const float v1 = vi1*d + m;
+            const int8_t vi1_0 = vi1 & 0xf;
+            const int8_t vi1_1 = vi1 >> 4;
 
-            y[i*QK4_3 + l + 0] = v0;
-            y[i*QK4_3 + l + 1] = v1;
+            const float v0_0 = (vi0_0 - 8)*d0;
+            const float v0_1 = (vi0_1 - 8)*d0;
 
-            assert(!isnan(y[i*QK4_3 + l + 0]));
-            assert(!isnan(y[i*QK4_3 + l + 1]));
+            const float v1_0 = (vi1_0 - 8)*d1;
+            const float v1_1 = (vi1_1 - 8)*d1;
+
+            y[i*QK4_3 + l + 0] = v0_0;
+            y[i*QK4_3 + l + 1] = v0_1;
+
+            y[i*QK4_3 + l + 0 + QK4_3/2] = v1_0;
+            y[i*QK4_3 + l + 1 + QK4_3/2] = v1_1;
         }
     }
 }
@@ -1795,7 +1800,7 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_2] = {
         .dequantize_row_q         = dequantize_row_q4_2,
         .quantize_row_q           = quantize_row_q4_2,
-        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_rmse, //quantize_row_q4_2_reference,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_reference,
         .quantize_row_q_dot       = quantize_row_q8_0,
         .vec_dot_q                = ggml_vec_dot_q4_2_q8_0,
     },
@@ -2635,15 +2640,15 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
-        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs));
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
 
         const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
         const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@@ -2876,17 +2881,16 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
 
     assert(n % QK8_0 == 0);
     assert(nb % 2 == 0);
-    assert(QK8_0 == 2*QK4_2);
+    assert(QK8_0 == 2*QK4_3);
 
     const block_q4_3 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
 
 #if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs0 = 0.0f;
-    float summs1 = 0.0f;
+    float32x2_t sumv0 = vdup_n_f32(0.0f);
+    float32x2_t sumv1 = vdup_n_f32(0.0f);
+    float32x2_t sumv2 = vdup_n_f32(0.0f);
+    float32x2_t sumv3 = vdup_n_f32(0.0f);
 
     for (int i = 0; i < nb; ++i) {
         const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0];
@@ -2894,29 +2898,46 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
 
         const block_q8_0 * restrict y0 = &y[i + 0];
 
-        summs0 += GGML_FP16_TO_FP32(x0_0->m) * y0->s0;
-        summs1 += GGML_FP16_TO_FP32(x0_1->m) * y0->s1;
-
-        const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs));
+        const uint8x8_t v0_0 = vld1_u8(x0_0->qs);
+        const uint8x8_t v0_1 = vld1_u8(x0_1->qs);
 
         // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, vdupq_n_u8(0xf)));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x8_t v0_0l = vreinterpret_s8_u8(vand_u8  (v0_0, vdup_n_u8(0xf)));
+        const int8x8_t v0_0h = vreinterpret_s8_u8(vshr_n_u8(v0_0, 4));
+        const int8x8_t v0_1l = vreinterpret_s8_u8(vand_u8  (v0_1, vdup_n_u8(0xf)));
+        const int8x8_t v0_1h = vreinterpret_s8_u8(vshr_n_u8(v0_1, 4));
+
+        // sub 8
+        const int8x8_t v0_0ls = vsub_s8(v0_0l, vdup_n_s8(8));
+        const int8x8_t v0_0hs = vsub_s8(v0_0h, vdup_n_s8(8));
+        const int8x8_t v0_1ls = vsub_s8(v0_1l, vdup_n_s8(8));
+        const int8x8_t v0_1hs = vsub_s8(v0_1h, vdup_n_s8(8));
 
         // interleave
-        const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h);
-        const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h);
+        const int8x8_t v0_0lz = vzip1_s8(v0_0ls, v0_0hs);
+        const int8x8_t v0_0hz = vzip2_s8(v0_0ls, v0_0hs);
+        const int8x8_t v0_1lz = vzip1_s8(v0_1ls, v0_1hs);
+        const int8x8_t v0_1hz = vzip2_s8(v0_1ls, v0_1hs);
 
         // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x8_t v1_0l = vld1_s8(y0->qs);
+        const int8x8_t v1_0h = vld1_s8(y0->qs + 8);
+        const int8x8_t v1_1l = vld1_s8(y0->qs + 16);
+        const int8x8_t v1_1h = vld1_s8(y0->qs + 24);
 
-        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d);
-        const float x0_1d = GGML_FP16_TO_FP32(x0_1->d);
+        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d0);
+        const float x0_1d = GGML_FP16_TO_FP32(x0_0->d1);
+        const float x1_0d = GGML_FP16_TO_FP32(x0_1->d0);
+        const float x1_1d = GGML_FP16_TO_FP32(x0_1->d1);
 
 #if defined(__ARM_FEATURE_DOTPROD)
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
+        //sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d);
+        //sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
+
+        sumv0 = vmla_n_f32(sumv0, vcvt_f32_s32(vdot_s32(vdup_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d);
+        sumv1 = vmla_n_f32(sumv1, vcvt_f32_s32(vdot_s32(vdup_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
+        sumv2 = vmla_n_f32(sumv2, vcvt_f32_s32(vdot_s32(vdup_n_s32(0), v0_1lz, v1_1l)), x1_0d*y0->d);
+        sumv3 = vmla_n_f32(sumv3, vcvt_f32_s32(vdot_s32(vdup_n_s32(0), v0_1hz, v1_1h)), x1_1d*y0->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
@@ -2931,77 +2952,79 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void *
 #endif
     }
 
-    *s = vaddvq_f32(vaddq_f32(sumv0, sumv1)) + summs0 + summs1;
+    *s = vaddv_f32(vadd_f32(vadd_f32(sumv0, sumv1), vadd_f32(sumv2, sumv3)));
 #elif defined(__AVX2__)
+    GGML_ASSERT(false); // TODO
     // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
+    //__m256 acc = _mm256_setzero_ps();
 
-    // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d));
-        const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d));
-        const __m256 dx = _mm256_set_m128(d1, d0);
+    //// Main loop
+    //for (int i = 0; i < nb; i++) {
+    //    const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d));
+    //    const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d));
+    //    const __m256 dx = _mm256_set_m128(d1, d0);
 
-        const __m128 m0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].m));
-        const __m128 m1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].m));
-        const __m256 mx = _mm256_set_m128(m1, m0);
+    //    const __m128 m0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].m));
+    //    const __m128 m1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].m));
+    //    const __m256 mx = _mm256_set_m128(m1, m0);
 
-        const __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs);
-        const __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs);
-        const __m256i bx = _mm256_set_m128i(bx1, bx0);
+    //    const __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs);
+    //    const __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs);
+    //    const __m256i bx = _mm256_set_m128i(bx1, bx0);
 
-        const __m256 dy = _mm256_broadcast_ss(&y[i].d);
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+    //    const __m256 dy = _mm256_broadcast_ss(&y[i].d);
+    //    const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256i syi = _mm256_maddubs_epi16(_mm256_set1_epi8(1), by);
-        const __m256 syf = sum_i16_pairs_float(syi);
+    //    const __m256i syi = _mm256_maddubs_epi16(_mm256_set1_epi8(1), by);
+    //    const __m256 syf = sum_i16_pairs_float(syi);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+    //    const __m256 q = mul_sum_i8_pairs_float(bx, by);
 
-        const __m256 sxy = _mm256_fmadd_ps(q, dx, _mm256_mul_ps(mx, syf));
-        acc = _mm256_fmadd_ps(sxy, dy, acc);
-    }
+    //    const __m256 sxy = _mm256_fmadd_ps(q, dx, _mm256_mul_ps(mx, syf));
+    //    acc = _mm256_fmadd_ps(sxy, dy, acc);
+    //}
 
-    *s = hsum_float_8(acc);
+    //*s = hsum_float_8(acc);
 #else
-    // scalar
-    float sumf = 0.0;
-    for (int i = 0; i < nb; i++) {
-        const uint8_t * restrict x0 = x[2*i + 0].qs;
-        const uint8_t * restrict x1 = x[2*i + 1].qs;
-        const  int8_t * restrict y0 = y[i].qs;
+    GGML_ASSERT(false); // TODO
+    //// scalar
+    //float sumf = 0.0;
+    //for (int i = 0; i < nb; i++) {
+    //    const uint8_t * restrict x0 = x[2*i + 0].qs;
+    //    const uint8_t * restrict x1 = x[2*i + 1].qs;
+    //    const  int8_t * restrict y0 = y[i].qs;
 
-        const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d);
-        const float m0 = GGML_FP16_TO_FP32(x[2*i + 0].m);
-        const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d);
-        const float m1 = GGML_FP16_TO_FP32(x[2*i + 1].m);
+    //    const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d);
+    //    const float m0 = GGML_FP16_TO_FP32(x[2*i + 0].m);
+    //    const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d);
+    //    const float m1 = GGML_FP16_TO_FP32(x[2*i + 1].m);
 
-        int sxy_0 = 0;
-        int sxy_1 = 0;
+    //    int sxy_0 = 0;
+    //    int sxy_1 = 0;
 
-        for (int j = 0; j < QK8_0/4; j++) {
-            const uint8_t v0 = x0[j];
-            const uint8_t v1 = x1[j];
+    //    for (int j = 0; j < QK8_0/4; j++) {
+    //        const uint8_t v0 = x0[j];
+    //        const uint8_t v1 = x1[j];
 
-            const int x0_0 = v0 & 0xf;
-            const int x1_0 = v0 >> 4;
+    //        const int x0_0 = v0 & 0xf;
+    //        const int x1_0 = v0 >> 4;
 
-            const int x0_1 = v1 & 0xf;
-            const int x1_1 = v1 >> 4;
+    //        const int x0_1 = v1 & 0xf;
+    //        const int x1_1 = v1 >> 4;
 
-            const int y0_0 = y0[2*j + 0];
-            const int y1_0 = y0[2*j + 1];
+    //        const int y0_0 = y0[2*j + 0];
+    //        const int y1_0 = y0[2*j + 1];
 
-            const int y0_1 = y0[2*(j + QK8_0/4) + 0];
-            const int y1_1 = y0[2*(j + QK8_0/4) + 1];
+    //        const int y0_1 = y0[2*(j + QK8_0/4) + 0];
+    //        const int y1_1 = y0[2*(j + QK8_0/4) + 1];
 
-            sxy_0 += x0_0*y0_0 + x1_0*y1_0;
-            sxy_1 += x0_1*y0_1 + x1_1*y1_1;
-        }
+    //        sxy_0 += x0_0*y0_0 + x1_0*y1_0;
+    //        sxy_1 += x0_1*y0_1 + x1_1*y1_1;
+    //    }
 
-        sumf += (d0*sxy_0 + d1*sxy_1)*y[i].d + m0*y[i].s0 + m1*y[i].s1;
-    }
-    *s = sumf;
+    //    sumf += (d0*sxy_0 + d1*sxy_1)*y[i].d + m0*y[i].s0 + m1*y[i].s1;
+    //}
+    //*s = sumf;
 #endif
 }
 
@@ -7992,6 +8015,9 @@ static void ggml_compute_forward_mul_mat_q_f32(
         else if (type == GGML_TYPE_Q4_2) {
             dequantize_row_q_cuda = dequantize_row_q4_2_cuda;
         }
+        else if (type == GGML_TYPE_Q4_3) {
+            dequantize_row_q_cuda = dequantize_row_q4_3_cuda;
+        }
         else {
             GGML_ASSERT(false);
         }
@@ -12124,8 +12150,7 @@ size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t *
     for (int j = 0; j < n; j += k) {
         block_q4_2 * restrict y = (block_q4_2 *)dst + j/QK4_2;
 
-        //quantize_row_q4_2_reference(src + j, y, k);
-        quantize_row_q4_2_rmse(src + j, y, k);
+        quantize_row_q4_2_reference(src + j, y, k);
 
         for (int i = 0; i < nb; i++) {
             for (int l = 0; l < QK4_2; l += 2) {

llama.cpp

@@ -2256,7 +2256,6 @@ std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_te
 
 // Returns the size of the state
 size_t llama_get_state_size(struct llama_context * ctx) {
-    const size_t s_bool = sizeof(int32_t);
     // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
     // for reference, std::mt19937(1337) serializes to 6701 bytes.
     const size_t s_rng_size = sizeof(size_t);

tests/CMakeLists.txt

@@ -6,5 +6,6 @@ function(llama_add_test source)
 endfunction()
 
 # llama_add_test(test-double-float.c) # SLOW
-llama_add_test(test-quantize.c)
+llama_add_test(test-quantize-fns.cpp)
+llama_add_test(test-quantize-perf.cpp)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)

tests/test-quantize-fns.cpp

@@ -0,0 +1,154 @@
+// Unit tests for quantization specific functions - quantize, dequantize and dot product
+
+#include "ggml.h"
+
+#undef NDEBUG
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string>
+#include <vector>
+
+
+const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
+const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
+const float MAX_DOT_PRODUCT_ERROR = 0.02;
+
+const char* RESULT_STR[] = {"ok", "FAILED"};
+
+
+// Generate synthetic data
+void generate_data(float offset, size_t n, float * dst) {
+    for (size_t i = 0; i < n; i++) {
+        dst[i] = 0.1 + 2*cosf(i + offset);
+    }
+}
+
+// Calculate RMSE between two float arrays
+float array_rmse(const float * a1, const float * a2, size_t n) {
+    double sum = 0;
+    for (size_t i = 0; i < n; i++) {
+        double diff = a1[i] - a2[i];
+        sum += diff * diff;
+    }
+    return sqrtf(sum) / n;
+}
+
+// Total quantization error on test data
+float total_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+    std::vector<uint8_t> tmp_q(test_size);
+    std::vector<float> tmp_out(test_size);
+
+    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+    return array_rmse(test_data, tmp_out.data(), test_size);
+}
+
+// Total quantization error on test data
+float reference_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
+    std::vector<uint8_t> tmp_q(test_size);
+    std::vector<float> tmp_out(test_size);
+    std::vector<float> tmp_out_ref(test_size);
+
+    qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
+
+    qfns.quantize_row_q_reference(test_data, tmp_q.data(), test_size);
+    qfns.dequantize_row_q(tmp_q.data(), tmp_out_ref.data(), test_size);
+
+    return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
+}
+
+float dot_product(const float * a1, const float * a2, size_t test_size) {
+    double sum = 0;
+    for (size_t i = 0; i < test_size; i++) {
+        sum += a1[i] * a2[i];
+    }
+    return sum;
+}
+
+// Total dot product error
+float dot_product_error(quantize_fns_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
+    std::vector<uint8_t> tmp_q1(test_size);
+    std::vector<uint8_t> tmp_q2(test_size*2);
+
+    qfns.quantize_row_q(test_data1, tmp_q1.data(), test_size);
+    qfns.quantize_row_q_dot(test_data2, tmp_q2.data(), test_size);
+
+    float result = INFINITY;
+    qfns.vec_dot_q(test_size, &result, tmp_q1.data(), tmp_q2.data());
+
+    const float dot_ref = dot_product(test_data1, test_data2, test_size);
+
+    return fabsf(result - dot_ref) / test_size;
+}
+
+int main(int argc, char * argv[]) {
+    bool verbose = false;
+    const size_t test_size = 32 * 128;
+
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "-v") {
+            verbose = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            return 1;
+        }
+    }
+
+    std::vector<float> test_data(test_size);
+    std::vector<float> test_data2(test_size);
+
+    generate_data(0.0, test_data.size(), test_data.data());
+    generate_data(1.0, test_data2.size(), test_data2.data());
+
+    // Initialize GGML, ensures float conversion tables are initialized
+    struct ggml_init_params ggml_params = {
+        /* .mem_size   = */ 1*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true,
+    };
+    struct ggml_context * ctx = ggml_init(ggml_params);
+
+    int num_failed = 0;
+    bool failed = false;
+
+    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        ggml_type type = (ggml_type) i;
+        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+
+        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            const float total_error = total_quantization_error(qfns, test_size, test_data.data());
+            failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s absolute quantization error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
+            }
+
+            const float reference_error = reference_quantization_error(qfns, test_size, test_data.data());
+            failed = !(reference_error < MAX_QUANTIZATION_REFERENCE_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
+            }
+
+            const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
+            failed = !(vec_dot_error < MAX_DOT_PRODUCT_ERROR);
+            num_failed += failed;
+            if (failed || verbose) {
+                printf("%5s dot product error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error);
+            }
+        }
+    }
+
+    if (num_failed || verbose) {
+        printf("%d tests failed\n", num_failed);
+    }
+
+    ggml_free(ctx);
+
+    return num_failed > 0;
+}

tests/test-quantize-perf.cpp

@@ -0,0 +1,310 @@
+// Benchmark quantization specific functions on synthetic data
+
+#include "ggml.h"
+
+#undef NDEBUG
+#include <algorithm>
+#include <assert.h>
+#include <functional>
+#include <inttypes.h>
+#include <math.h>
+#include <memory>
+#include <stdio.h>
+#include <string>
+#include <vector>
+
+#define MAX_ALIGNMENT 64
+#define QK 32
+#define WARMUP 5
+#define ITERATIONS 10
+
+#define L1_SIZE      32*128
+#define L2_SIZE     32*2048
+#define L3_SIZE    32*20480
+#define MEM_SIZE 32*2048000
+
+struct quantize_perf_params {
+    std::vector<std::string> include_types;
+    std::vector<size_t> test_sizes;
+    size_t alignment_offset = 0;
+    bool op_quantize_row_q_reference = false;
+    bool op_quantize_row_q = false;
+    bool op_dequantize_row_q = false;
+    bool op_quantize_row_q_dot = false;
+    bool op_vec_dot_q = false;
+};
+
+
+#if defined(__x86_64__) || defined(__i386__)
+
+#include <x86intrin.h>
+inline int64_t cpu_cycles() {
+// Rough way to detect new-ish CPUs
+#ifdef __POPCNT__
+    unsigned int dummy;
+    return __rdtscp(&dummy);
+#else
+    return __rdtsc();
+#endif
+}
+
+#else
+
+#define cpu_cycles() 0
+
+#endif
+
+
+// Generate synthetic data
+void generate_data(float offset, size_t n, float * dst) {
+    for (size_t i = 0; i < n; i++) {
+        dst[i] = 0.1 + 2*cosf(i + offset);
+    }
+}
+
+float gigabytes_per_second(size_t bytes, int64_t usecs) {
+    return bytes / (float) usecs * 1000000 / (1024*1024*1024);
+}
+
+void * align_with_offset(void * ptr, int offset) {
+    size_t dummy_size = MAX_ALIGNMENT * 4;
+    return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset;
+}
+
+void benchmark_function(size_t size, size_t q_size, std::function<size_t(void)> function) {
+    int64_t min_time_us = INT64_MAX;
+    int64_t total_time_us = 0;
+    int64_t min_time_cycles = INT64_MAX;
+    int64_t total_time_cycles = 0;
+
+    for (int i = 0; i < WARMUP; i++) {
+        function();
+    }
+
+
+    for (int i = 0; i < ITERATIONS; i++) {
+        const int64_t start_time = ggml_time_us();
+        const int64_t start_cycles = cpu_cycles();
+
+        function();
+
+        const int64_t end_cycles = cpu_cycles();
+        const int64_t end_time = ggml_time_us();
+
+        total_time_cycles += end_cycles - start_cycles;
+        min_time_cycles = std::min(min_time_cycles, end_cycles - start_cycles);
+        total_time_us += end_time - start_time;
+        min_time_us = std::min(min_time_us, end_time - start_time);
+    }
+
+    printf("      min cycles/%d vals   : %9.2f\n",  QK, QK * min_time_cycles / (float) size);
+    printf("      avg cycles/%d vals   : %9.2f\n",  QK, QK * total_time_cycles / (float) (size * ITERATIONS));
+    printf("      float32 throughput   : %9.2f GB/s\n",  gigabytes_per_second(4 * size * ITERATIONS, total_time_us));
+    printf("      quantized throughput : %9.2f GB/s\n",  gigabytes_per_second(q_size * ITERATIONS, total_time_us));
+}
+
+int main(int argc, char * argv[]) {
+    quantize_perf_params params {};
+
+    // read command line
+
+    bool invalid_param = false;
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "--size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            size_t size = std::stoi(argv[i]);
+            if (size % 32 != 0) {
+                fprintf(stderr, "error: size %zu not divisible by 32\n", size);
+                invalid_param = true;
+                break;
+            }
+            params.test_sizes.push_back(size);
+        } else if (arg == "-3") {
+            // quick select sizes that probably fit in CPU caches
+            params.test_sizes.push_back(L1_SIZE);
+            params.test_sizes.push_back(L2_SIZE);
+            params.test_sizes.push_back(L3_SIZE);
+        } else if (arg == "-4") {
+            // quick select cache sizes + memory
+            params.test_sizes.push_back(L1_SIZE);
+            params.test_sizes.push_back(L2_SIZE);
+            params.test_sizes.push_back(L3_SIZE);
+            params.test_sizes.push_back(MEM_SIZE);
+        } else if (arg == "--op") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string op {argv[i]};
+            if (op == "quantize_row_q_reference") {
+                params.op_quantize_row_q_reference = true;
+            } else if (op == "quantize_row_q") {
+                params.op_quantize_row_q = true;
+            } else if (op == "dequantize_row_q") {
+                params.op_dequantize_row_q = true;
+            } else if (op == "quantize_row_q_dot") {
+                params.op_quantize_row_q_dot = true;
+            } else if (op == "vec_dot_q") {
+                params.op_vec_dot_q = true;
+            } else {
+                invalid_param = true;
+                break;
+            }
+        } else if (arg == "--type") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.include_types.push_back(argv[i]);
+        } else if (arg == "--alignment-offset") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            int alignment = std::stoi(argv[i]);
+            if (alignment < 0 || alignment > MAX_ALIGNMENT) {
+            fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
+                invalid_param = true;
+                break;
+            }
+            params.alignment_offset = alignment;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            return 1;
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        return 1;
+    }
+
+    if (params.test_sizes.empty()) {
+        params.test_sizes.push_back(L1_SIZE);
+    }
+    if (!(params.op_quantize_row_q_reference || params.op_quantize_row_q || params.op_dequantize_row_q || params.op_quantize_row_q_dot || params.op_vec_dot_q)) {
+        params.op_quantize_row_q_reference = params.op_quantize_row_q = params.op_dequantize_row_q = params.op_quantize_row_q_dot = params.op_vec_dot_q = true;
+    }
+
+    std::sort(params.test_sizes.begin(), params.test_sizes.end());
+    size_t largest = params.test_sizes.back();
+
+    std::vector<uint8_t> test_data1_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_data2_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q1_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q2_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_out_v(largest*4 + MAX_ALIGNMENT*2);
+
+    float * test_data1 = (float *) align_with_offset(test_data1_v.data(), params.alignment_offset);
+    float * test_data2 = (float *) align_with_offset(test_data2_v.data(), params.alignment_offset);
+    float * test_q1 = (float *) align_with_offset(test_q1_v.data(), params.alignment_offset);
+    float * test_q2 = (float *) align_with_offset(test_q2_v.data(), params.alignment_offset);
+    float * test_out = (float *) align_with_offset(test_out_v.data(), params.alignment_offset);
+
+    generate_data(0, largest, test_data1);
+    generate_data(1, largest, test_data2);
+
+
+    // Initialize GGML, ensures float conversion tables are initialized
+    struct ggml_init_params ggml_params = {
+        /* .mem_size   = */ 1*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true,
+    };
+    struct ggml_context * ctx = ggml_init(ggml_params);
+
+    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        ggml_type type = (ggml_type) i;
+        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+        if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
+            continue;
+        }
+
+        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            printf("%s\n", ggml_type_name(type));
+
+            if (params.op_quantize_row_q_reference) {
+                printf("  quantize_row_q_reference\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q_reference(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_quantize_row_q) {
+                printf("  quantize_row_q\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_dequantize_row_q) {
+                printf("  dequantize_row_q\n");
+                qfns.quantize_row_q(test_data1, test_q1, largest);
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.dequantize_row_q(test_q1, test_out, size);
+                        return test_out[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_quantize_row_q_dot) {
+                printf("  quantize_row_q_dot\n");
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        qfns.quantize_row_q_dot(test_data1, test_q1, size);
+                        return test_q1[0];
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+
+            if (params.op_vec_dot_q) {
+                printf("  vec_dot_q\n");
+                qfns.quantize_row_q(test_data1, test_q1, largest);
+                qfns.quantize_row_q(test_data2, test_q2, largest);
+                for (size_t size : params.test_sizes) {
+                    printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
+                    auto quantize_fn = [&](void ) {
+                        float result;
+                        qfns.vec_dot_q(size, &result, test_q1, test_q2);
+                        return result;
+                    };
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
+                    benchmark_function(size, quantized_size, quantize_fn);
+                }
+                printf("\n");
+            }
+        }
+    }
+
+    ggml_free(ctx);
+
+    return 0;
+}

tests/test-quantize.c

@@ -1,42 +0,0 @@
-#include "ggml.h"
-#undef NDEBUG
-#include <assert.h>
-#include <math.h>
-
-int main(void) {
-    #define QK 32
-    float src[QK];
-    uint8_t dst[24];
-    int64_t hist[16];
-
-    for (int i = 0; i < QK; i++) {
-        src[i] = (float)(i + 1);
-    }
-
-    size_t size = ggml_quantize_q4_0(src, dst, QK, QK, hist);
-    assert(size == 20);
-    float max_result = ((float *)dst)[0];
-    float max_expected = src[31] / ((1 << 3) - 1);
-    assert(max_result == max_expected);
-    for (int i = 0; i < QK; i++) {
-        uint8_t q4_result = (i % 2) ? (dst[sizeof(float) + i/2] >> 4) : (dst[sizeof(float) + i/2] & 0xF);
-        uint8_t q4_expected = roundf(src[i] / max_expected) + 8;
-        assert(q4_result == q4_expected);
-    }
-
-    size = ggml_quantize_q4_1(src, dst, QK, QK, hist);
-    assert(size == 24);
-    float delta_result = ((float *)dst)[0];
-    float delta_expected = (src[31] - src[0]) / ((1 << 4) - 1);
-    assert(delta_result == delta_expected);
-    float min_result = ((float *)dst)[1];
-    float min_expected = src[0];
-    assert(min_result == min_expected);
-    for (int i = 0; i < QK; i++) {
-        uint8_t q4_result = (i % 2) ? (dst[sizeof(float)*2 + i/2] >> 4) : (dst[sizeof(float)*2 + i/2] & 0xF);
-        uint8_t q4_expected = roundf((src[i] - min_expected) / delta_expected);
-        assert(q4_result == q4_expected);
-    }
-
-    return 0;
-}