llama_cpp 0.0.7 → 0.1.0

data/ext/llama_cpp/src/ggml.c

@@ -135,14 +135,6 @@ inline static void* ggml_aligned_malloc(size_t size) {
 #define UNUSED(x) (void)(x)
 #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
 
-#define GGML_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            abort(); \
-        } \
-    } while (0)
-
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
 #elif defined(GGML_USE_OPENBLAS)
@@ -188,9 +180,13 @@ typedef double ggml_float;
 #undef bool
 #define bool _Bool
 #else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
 #include <immintrin.h>
 #endif
 #endif
+#endif
 
 #ifdef __F16C__
 
@@ -330,7 +326,7 @@ static ggml_fp16_t table_exp_f16[1 << 16];
 // precomputed f32 table for f16 (256 KB)
 static float table_f32_f16[1 << 16];
 
-#if defined(__ARM_NEON)
+#if defined(__ARM_NEON) || defined(__wasm_simd128__)
 #define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
 #define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
 #define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
@@ -370,6 +366,32 @@ ggml_fp16_t ggml_fp32_to_fp16(float x) {
     return GGML_FP32_TO_FP16(x);
 }
 
+void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n) {
+    for (size_t i = 0; i < n; i++) {
+        y[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+}
+
+void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n) {
+    size_t i = 0;
+#if defined(__F16C__)
+    for (; i + 7 < n; i += 8) {
+        __m256 x_vec = _mm256_loadu_ps(x + i);
+        __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storeu_si128((__m128i *)(y + i), y_vec);
+    }
+    for(; i + 3 < n; i += 4) {
+        __m128 x_vec = _mm_loadu_ps(x + i);
+        __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storel_epi64((__m128i *)(y + i), y_vec);
+    }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_FP32_TO_FP16(x[i]);
+    }
+}
+
+
 //
 // timing
 //
@@ -653,19 +675,102 @@ float vmaxvq_f32(float32x4_t v) {
 }
 
 int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) {
-    return vget_low_s8(vcombine_s8(a, b));
+    int8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
 }
 
 int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) {
-    return vget_high_s8(vcombine_s8(a, b));
+    int8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
 }
 
 uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    return vget_low_u8(vcombine_u8(a, b));
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
 }
 
 uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    return vget_high_u8(vcombine_u8(a, b));
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) {
+    int8x16_t res;
+
+    res[0]  = a[0]; res[1]  = b[0]; res[2]  = a[1]; res[3]  = b[1];
+    res[4]  = a[2]; res[5]  = b[2]; res[6]  = a[3]; res[7]  = b[3];
+    res[8]  = a[4]; res[9]  = b[4]; res[10] = a[5]; res[11] = b[5];
+    res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7];
+
+    return res;
+}
+
+int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) {
+    int8x16_t res;
+
+    res[0]  = a[8];  res[1]  = b[8];  res[2]  = a[9];  res[3]  = b[9];
+    res[4]  = a[10]; res[5]  = b[10]; res[6]  = a[11]; res[7]  = b[11];
+    res[8]  = a[12]; res[9]  = b[12]; res[10] = a[13]; res[11] = b[13];
+    res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
+
+    return res;
+}
+
+uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[0]  = a[0];  res[1]  = b[0];  res[2]  = a[1];  res[3]  = b[1];
+    res[4]  = a[2];  res[5]  = b[2];  res[6]  = a[3];  res[7]  = b[3];
+    res[8]  = a[4];  res[9]  = b[4];  res[10] = a[5];  res[11] = b[5];
+    res[12] = a[6];  res[13] = b[6];  res[14] = a[7];  res[15] = b[7];
+
+    return res;
+}
+
+uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[0]  = a[8];  res[1]  = b[8];  res[2]  = a[9];  res[3]  = b[9];
+    res[4]  = a[10]; res[5]  = b[10]; res[6]  = a[11]; res[7]  = b[11];
+    res[8]  = a[12]; res[9]  = b[12]; res[10] = a[13]; res[11] = b[13];
+    res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
+
+    return res;
+}
+
+int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
 }
 
 #endif
@@ -694,14 +799,6 @@ typedef struct {
 } block_q4_2;
 static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
 
-#define QK4_3 16
-typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // 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");
-
 #define QK5_0 32
 typedef struct {
     ggml_fp16_t d;         // delta
@@ -789,6 +886,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         float max = 0.0f;
         float min = 0.0f;
 
+        vector float asrcv [8];
         vector float srcv [8];
         vector float maxv[8];
         vector float minv[8];
@@ -1068,7 +1166,7 @@ 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));
+            const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15));
 
             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);
@@ -1291,49 +1389,6 @@ static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int
     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) {
-    assert(k % QK4_3 == 0);
-    const int nb = k / QK4_3;
-
-    for (int i = 0; i < nb; i++) {
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        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;
-        }
-
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-        y[i].m = GGML_FP32_TO_FP16(min);
-
-        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;
-
-            const uint8_t vi0 = (int) (v0 + 0.5f);
-            const uint8_t vi1 = (int) (v1 + 0.5f);
-
-            assert(vi0 < 16);
-            assert(vi1 < 16);
-
-            y[i].qs[l/2] = vi0 | (vi1 << 4);
-        }
-    }
-}
-
-static void quantize_row_q4_3(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK4_3 == 0);
-
-    block_q4_3 * restrict y = vy;
-
-    quantize_row_q4_3_reference(x, y, k);
-}
-
 static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) {
     assert(k % QK5_0 == 0);
     const int nb = k / QK5_0;
@@ -1458,15 +1513,135 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
 }
 
 static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) {
+    assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
 
     block_q8_0 * restrict y = vy;
 
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[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]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+
+        for (int l = 0; l < 8; l++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[l], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __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 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        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 );
+
+        // Quantize these floats
+        const float d = maxScalar / 127.f;
+        y[i].d = d;
+        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#else
+    // scalar
     quantize_row_q8_0_reference(x, y, k);
+#endif
 }
 
 // reference implementation for deterministic creation of model files
 static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) {
+    assert(QK8_1 == 32);
     assert(k % QK8_1 == 0);
     const int nb = k / QK8_1;
 
@@ -1917,36 +2092,6 @@ static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, in
     }
 }
 
-static void dequantize_row_q4_3(const void * restrict vx, float * restrict y, int k) {
-    assert(k % QK4_3 == 0);
-    const int nb = k / QK4_3;
-
-    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 uint8_t * restrict pp = x[i].qs;
-
-        for (int l = 0; l < QK4_3; l += 2) {
-            const uint8_t vi = pp[l/2];
-
-            const int8_t vi0 = vi & 0x0F;
-            const int8_t vi1 = vi >> 4;
-
-            const float v0 = vi0*d + m;
-            const float v1 = vi1*d + m;
-
-            y[i*QK4_3 + l + 0] = v0;
-            y[i*QK4_3 + l + 1] = v1;
-
-            assert(!isnan(y[i*QK4_3 + l + 0]));
-            assert(!isnan(y[i*QK4_3 + l + 1]));
-        }
-    }
-}
-
 static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, int k) {
     assert(k % QK5_0 == 0);
     const int nb = k / QK5_0;
@@ -1965,8 +2110,8 @@ 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;
@@ -2002,8 +2147,8 @@ 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;
@@ -2040,7 +2185,6 @@ static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, in
 static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-static void ggml_vec_dot_q4_3_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
@@ -2070,14 +2214,6 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
         .vec_dot_q                = ggml_vec_dot_q4_2_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
     },
-    [GGML_TYPE_Q4_3] = {
-        .dequantize_row_q         = dequantize_row_q4_3,
-        .quantize_row_q           = quantize_row_q4_3,
-        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_3_reference,
-        .quantize_row_q_dot       = quantize_row_q8_1,
-        .vec_dot_q                = ggml_vec_dot_q4_3_q8_1,
-        .vec_dot_type             = GGML_TYPE_Q8_1,
-    },
     [GGML_TYPE_Q5_0] = {
         .dequantize_row_q         = dequantize_row_q5_0,
         .quantize_row_q           = quantize_row_q5_0,
@@ -2748,35 +2884,35 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
         const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
 
+        // interleave
+        const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs);
+        const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs);
+        const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs);
+        const int8x16_t v0_1hz = vzip2q_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 int8x16_t v1_1l = vld1q_s8(y1->qs);
         const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        // interleave
-        const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
-        const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
-        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h);
 
         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_0ls), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), 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_1ls), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), 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));
@@ -3171,136 +3307,6 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void *
 #endif
 }
 
-static void ggml_vec_dot_q4_3_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int nb = n / QK8_1;
-
-    assert(n % QK8_1 == 0);
-    assert(nb % 2 == 0);
-    assert(QK8_1 == 2*QK4_3);
-
-    const block_q4_3 * restrict x = vx;
-    const block_q8_1 * 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;
-
-    for (int i = 0; i < nb; ++i) {
-        const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0];
-        const block_q4_3 * restrict x0_1 = &x[2*(i + 0) + 1];
-
-        const block_q8_1 * 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));
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, vdupq_n_u8(0x0F)));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-
-        // interleave
-        const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h);
-        const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-
-        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d);
-        const float x0_1d = GGML_FP16_TO_FP32(x0_1->d);
-
-#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);
-#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));
-        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 int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
-        const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(pl0), x0_0d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(ph0), x0_1d*y0->d);
-#endif
-    }
-
-    *s = vaddvq_f32(vaddq_f32(sumv0, sumv1)) + summs0 + summs1;
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    float summs = 0.0f;
-
-    // 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);
-
-        summs += GGML_FP16_TO_FP32(x[2*i + 0].m) * y[i].s0
-               + GGML_FP16_TO_FP32(x[2*i + 1].m) * y[i].s1;
-
-        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 q = mul_sum_i8_pairs_float(bx, by);
-
-        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
-    }
-
-    *s = hsum_float_8(acc) + summs;
-#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;
-
-        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;
-
-        for (int j = 0; j < QK8_1/4; j++) {
-            const uint8_t v0 = x0[j];
-            const uint8_t v1 = x1[j];
-
-            const int x0_0 = v0 & 0x0F;
-            const int x1_0 = v0 >> 4;
-
-            const int x0_1 = v1 & 0x0F;
-            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_1 = y0[2*(j + QK8_1/4) + 0];
-            const int y1_1 = y0[2*(j + QK8_1/4) + 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;
-#endif
-}
-
 static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     const int nb = n / QK8_0;
 
@@ -3373,6 +3379,72 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = vaddvq_f32(sumv);
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint64_t tmp[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const block_q5_0 * restrict x0 = &x[i];
+        const block_q8_0 * restrict y0 = &y[i];
+
+        const v128_t m4b  = wasm_i8x16_splat(0x0F);
+        const v128_t s16b = wasm_i8x16_splat(0x10);
+
+        // extract the 5th bit
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_u[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_u[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_u[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // interleave
+        const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h,  0, 16,  1, 17,  2, 18,  3, 19,  4, 20,  5, 21,  6, 22,  7, 23);
+        const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h,  8, 24,  9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
+
+        // add high bit and sub 16
+        const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b);
+        const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        const float x0d = GGML_FP16_TO_FP32(x0->d);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
+    }
+
+    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -3413,8 +3485,8 @@ 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;
@@ -3504,6 +3576,77 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
     }
 
     *s = vaddvq_f32(sumv) + summs;
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint64_t tmp[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const block_q5_1 * restrict x0 = &x[i];
+        const block_q8_1 * restrict y0 = &y[i];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_u[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_u[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_u[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        static bool x = true;
+
+        // interleave
+        const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h,  0, 16,  1, 17,  2, 18,  3, 19,  4, 20,  5, 21,  6, 22,  7, 23);
+        const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h,  8, 24,  9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0lz, qhl);
+        const v128_t v0hf = wasm_v128_or(v0hz, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        const float x0d = GGML_FP16_TO_FP32(x0->d);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
+    }
+
+    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -3547,8 +3690,8 @@ 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;
@@ -3925,7 +4068,6 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = QK4_0,
     [GGML_TYPE_Q4_1] = QK4_1,
     [GGML_TYPE_Q4_2] = QK4_2,
-    [GGML_TYPE_Q4_3] = QK4_3,
     [GGML_TYPE_Q5_0] = QK5_0,
     [GGML_TYPE_Q5_1] = QK5_1,
     [GGML_TYPE_Q8_0] = QK8_0,
@@ -3942,7 +4084,6 @@ static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = sizeof(block_q4_0),
     [GGML_TYPE_Q4_1] = sizeof(block_q4_1),
     [GGML_TYPE_Q4_2] = sizeof(block_q4_2),
-    [GGML_TYPE_Q4_3] = sizeof(block_q4_3),
     [GGML_TYPE_Q5_0] = sizeof(block_q5_0),
     [GGML_TYPE_Q5_1] = sizeof(block_q5_1),
     [GGML_TYPE_Q8_0] = sizeof(block_q8_0),
@@ -3960,7 +4101,6 @@ static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = "q4_0",
     [GGML_TYPE_Q4_1] = "q4_1",
     [GGML_TYPE_Q4_2] = "q4_2",
-    [GGML_TYPE_Q4_3] = "q4_3",
     [GGML_TYPE_Q5_0] = "q5_0",
     [GGML_TYPE_Q5_1] = "q5_1",
     [GGML_TYPE_Q8_0] = "q8_0",
@@ -3977,7 +4117,6 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = true,
     [GGML_TYPE_Q4_1] = true,
     [GGML_TYPE_Q4_2] = true,
-    [GGML_TYPE_Q4_3] = true,
     [GGML_TYPE_Q5_0] = true,
     [GGML_TYPE_Q5_1] = true,
     [GGML_TYPE_Q8_0] = true,
@@ -4024,6 +4163,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "DIAG_MASK_INF",
     "SOFT_MAX",
     "ROPE",
+    "ALIBI",
     "CONV_1D_1S",
     "CONV_1D_2S",
 
@@ -4034,7 +4174,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "MAP_BINARY",
 };
 
-static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38");
+static_assert(GGML_OP_COUNT == 39, "GGML_OP_COUNT != 39");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -4072,6 +4212,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "diag_mask_inf(x)",
     "soft_max(x)",
     "rope(x)",
+    "alibi(x)",
     "conv_1d_1s(x)",
     "conv_1d_2s(x)",
 
@@ -4082,7 +4223,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "f(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38");
+static_assert(GGML_OP_COUNT == 39, "GGML_OP_COUNT != 39");
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
@@ -4252,6 +4393,27 @@ bool ggml_is_quantized(enum ggml_type type) {
     return GGML_IS_QUANTIZED[type];
 }
 
+enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
+    enum ggml_type wtype = GGML_TYPE_COUNT;
+
+    switch (ftype) {
+        case GGML_FTYPE_ALL_F32:              wtype = GGML_TYPE_F32;   break;
+        case GGML_FTYPE_MOSTLY_F16:           wtype = GGML_TYPE_F16;   break;
+        case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
+        case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
+        case GGML_FTYPE_MOSTLY_Q4_2:          wtype = GGML_TYPE_Q4_2;  break;
+        case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
+        case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
+        case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
+        case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
+    }
+
+    GGML_ASSERT(wtype != GGML_TYPE_COUNT);
+
+    return wtype;
+}
+
 static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
@@ -4362,12 +4524,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
-        // initialize cuBLAS
-        #if defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_CUBLAS)
         ggml_init_cublas();
-        #elif defined(GGML_USE_CLBLAST)
+#elif defined(GGML_USE_CLBLAST)
         ggml_cl_init();
-        #endif
+#endif
 
         is_first_call = false;
     }
@@ -4448,7 +4609,7 @@ void ggml_free(struct ggml_context * ctx) {
 }
 
 size_t ggml_used_mem(const struct ggml_context * ctx) {
-    return ctx->objects_end->offs + ctx->objects_end->size;
+    return ctx->objects_end == NULL ? 0 : ctx->objects_end->offs + ctx->objects_end->size;
 }
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) {
@@ -4561,6 +4722,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
         /*.data         =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data,
+        /*.name         =*/ { 0 },
         /*.pad          =*/ { 0 },
     };
 
@@ -4915,6 +5077,15 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) {
     return (float *)(tensor->data);
 }
 
+const char * ggml_get_name(const struct ggml_tensor * tensor) {
+    return tensor->name;
+}
+
+void ggml_set_name(struct ggml_tensor * tensor, const char * name) {
+    strncpy(tensor->name, name, sizeof(tensor->name));
+    tensor->name[sizeof(tensor->name) - 1] = '\0';
+}
+
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
         const struct ggml_tensor * src) {
@@ -6014,6 +6185,7 @@ struct ggml_tensor * ggml_diag_mask_inf(
     //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
     struct ggml_tensor * b = ggml_new_i32(ctx, n_past);
+    ggml_set_name(b, "n_past");
 
     result->op   = GGML_OP_DIAG_MASK_INF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6071,6 +6243,7 @@ struct ggml_tensor * ggml_rope(
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_dims;
     ((int32_t *) b->data)[2] = mode;
+    ggml_set_name(b, "n_past, n_dims, mode");
 
     result->op   = GGML_OP_ROPE;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6080,6 +6253,37 @@ struct ggml_tensor * ggml_rope(
     return result;
 }
 
+// ggml_alibi
+
+struct ggml_tensor * ggml_alibi(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past,
+        int                   n_head) {
+    GGML_ASSERT(n_past >= 0);
+    bool is_node = false;
+
+    if (a->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    // TODO: when implement backward, fix this:
+    //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
+    ((int32_t *) b->data)[0] = n_past;
+    ((int32_t *) b->data)[1] = n_head;
+
+    result->op   = GGML_OP_ALIBI;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src0 = a;
+    result->src1 = b;
+
+    return result;
+}
+
 // ggml_conv_1d_1s
 
 struct ggml_tensor * ggml_conv_1d_1s(
@@ -7199,7 +7403,6 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
-        case GGML_TYPE_Q4_3:
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
@@ -8108,7 +8311,7 @@ static void ggml_compute_forward_rms_norm(
 
 // ggml_compute_forward_mul_mat
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
 // helper function to determine if it is better to use BLAS or not
 // for large matrices, BLAS is faster
 static bool ggml_compute_forward_mul_mat_use_blas(
@@ -8125,7 +8328,8 @@ static bool ggml_compute_forward_mul_mat_use_blas(
 
     // TODO: find the optimal values for these
     if (ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) && ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32))) {
+        ggml_is_contiguous(src1) &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
 
         /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
         return true;
@@ -8133,7 +8337,6 @@ static bool ggml_compute_forward_mul_mat_use_blas(
 
     return false;
 }
-
 #endif
 
 static void ggml_compute_forward_mul_mat_f32(
@@ -8149,7 +8352,7 @@ static void ggml_compute_forward_mul_mat_f32(
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
     const int64_t ne10 = src1->ne[0];
 #endif
     const int64_t ne11 = src1->ne[1];
@@ -8206,7 +8409,16 @@ static void ggml_compute_forward_mul_mat_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CUBLAS)
+    if (ggml_cuda_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
+#endif
+
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         if (params->ith != 0) {
             return;
@@ -8220,42 +8432,13 @@ static void ggml_compute_forward_mul_mat_f32(
             return;
         }
 
-#if defined(GGML_USE_CUBLAS)
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
-        const int x_ne = ne01 * ne10;
-        const int y_ne = ne11 * ne10;
-        const int d_ne = ne11 * ne01;
-
-        size_t x_size, y_size, d_size;
-        float *d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
-        float *d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
-        float *d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
-#endif
-
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
-
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
-#if defined(GGML_USE_CUBLAS)
-                // copy data to device
-                CUDA_CHECK(cudaMemcpyAsync(d_X, x, sizeof(float) * x_ne, cudaMemcpyHostToDevice, g_cudaStream));
-                CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(float) * y_ne, cudaMemcpyHostToDevice, g_cudaStream));
-
-                // compute
-                CUBLAS_CHECK(
-                    cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            &alpha, d_X, ne00,
-                                    d_Y, ne10,
-                            &beta,  d_D, ne01));
-
-                // copy data to host
-                CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, g_cudaStream));
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
                 // zT = y * xT
                 ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
                         ne11, ne01, ne10,
@@ -8272,12 +8455,6 @@ static void ggml_compute_forward_mul_mat_f32(
 #endif
             }
         }
-#if defined(GGML_USE_CUBLAS)
-        CUDA_CHECK(cudaStreamSynchronize(g_cudaStream));
-        ggml_cuda_pool_free(d_X, x_size);
-        ggml_cuda_pool_free(d_Y, y_size);
-        ggml_cuda_pool_free(d_D, d_size);
-#endif
         //printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
 
         return;
@@ -8407,7 +8584,16 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CUBLAS)
+    if (ggml_cuda_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
+#endif
+
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         GGML_ASSERT(nb10 == sizeof(float));
 
@@ -8423,35 +8609,9 @@ static void ggml_compute_forward_mul_mat_f16_f32(
             return;
         }
 
-#if defined(GGML_USE_CUBLAS)
-        ggml_fp16_t * const wdata = params->wdata;
-
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
-        const int x_ne = ne01 * ne10;
-        const int y_ne = ne11 * ne10;
-        const int d_ne = ne11 * ne01;
-
-        size_t x_size, y_size, d_size;
-        float *d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
-        float *d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
-        float *d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
-#else
-        float * const wdata = params->wdata;
-#endif
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-#if defined(GGML_USE_CUBLAS)
-                // with cuBlAS, instead of converting src0 to fp32, we convert src1 to fp16
-                {
-                    size_t id = 0;
-                    for (int64_t i01 = 0; i01 < ne11; ++i01) {
-                        for (int64_t i00 = 0; i00 < ne10; ++i00) {
-                            wdata[id++] = GGML_FP32_TO_FP16(*(float *) ((char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11 + i00*nb10));
-                        }
-                    }
-                }
-#else
+                float * const wdata = params->wdata;
                 {
                     size_t id = 0;
                     for (int64_t i01 = 0; i01 < ne01; ++i01) {
@@ -8459,32 +8619,11 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                             wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00));
                         }
                     }
-                }
-#endif
-
-#if defined(GGML_USE_CUBLAS)
-                const ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + i02*nb02 + i03*nb03);
-                const ggml_fp16_t * y = (ggml_fp16_t *) wdata;
 
-                float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+                    assert(id*sizeof(float) <= params->wsize);
+                }
 
-                // copy data to device
-                CUDA_CHECK(cudaMemcpyAsync(d_X, x, sizeof(ggml_fp16_t) * x_ne, cudaMemcpyHostToDevice, g_cudaStream));
-                CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(ggml_fp16_t) * y_ne, cudaMemcpyHostToDevice, g_cudaStream));
-
-                // compute
-                CUBLAS_CHECK(
-                    cublasGemmEx(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            &alpha, d_X, CUDA_R_16F, ne00,
-                                    d_Y, CUDA_R_16F, ne10,
-                            &beta,  d_D, CUDA_R_32F, ne01,
-                            CUBLAS_COMPUTE_32F,
-                            CUBLAS_GEMM_DEFAULT));
-
-                // copy data to host
-                CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, g_cudaStream));
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
                 const float * x = wdata;
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
 
@@ -8513,12 +8652,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
             }
         }
 
-#if defined(GGML_USE_CUBLAS)
-        CUDA_CHECK(cudaStreamSynchronize(g_cudaStream));
-        ggml_cuda_pool_free(d_X, x_size);
-        ggml_cuda_pool_free(d_Y, y_size);
-        ggml_cuda_pool_free(d_D, d_size);
-#endif
         /*printf("CBLAS F16 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);*/
 
         return;
@@ -8671,7 +8804,16 @@ static void ggml_compute_forward_mul_mat_q_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CUBLAS)
+    if (ggml_cuda_can_mul_mat(src0, src1, dst)) {
+        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
+            ggml_cuda_mul_mat(src0, src1, dst, params->wdata, params->wsize);
+        }
+        return;
+    }
+#endif
+
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         if (params->ith != 0) {
             return;
@@ -8685,48 +8827,8 @@ static void ggml_compute_forward_mul_mat_q_f32(
             return;
         }
 
-#if defined(GGML_USE_CUBLAS)
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
-        const int x_ne = ne01 * ne10;
-        const int y_ne = ne11 * ne10;
-        const int d_ne = ne11 * ne01;
-
-        size_t x_size, y_size, d_size, q_size;
-        float *d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
-        float *d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
-        float *d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
-        float *d_Q = ggml_cuda_pool_malloc(GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], &q_size);
-
-        void (*dequantize_row_q_cuda)(const void * x, float * y, int k, cudaStream_t stream)  = NULL;
-        if (type == GGML_TYPE_Q4_0) {
-            dequantize_row_q_cuda = dequantize_row_q4_0_cuda;
-        }
-        else if (type == GGML_TYPE_Q4_1) {
-            dequantize_row_q_cuda = dequantize_row_q4_1_cuda;
-        }
-        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 if (type == GGML_TYPE_Q5_0) {
-            dequantize_row_q_cuda = dequantize_row_q5_0_cuda;
-        }
-        else if (type == GGML_TYPE_Q5_1) {
-            dequantize_row_q_cuda = dequantize_row_q5_1_cuda;
-        }
-        else if (type == GGML_TYPE_Q8_0) {
-            dequantize_row_q_cuda = dequantize_row_q8_0_cuda;
-        }
-        else {
-            GGML_ASSERT(false);
-        }
-#elif !defined(GGML_USE_CLBLAST)
         float * const wdata = params->wdata;
         dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
-#endif
 
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -8734,15 +8836,7 @@ static void ggml_compute_forward_mul_mat_q_f32(
 
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
-#if defined(GGML_USE_CUBLAS)
-                // copy and dequantize on device
-                CUDA_CHECK(
-                    cudaMemcpyAsync(d_Q, (char *) src0->data + i03*nb03 + i02*nb02,
-                        GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], cudaMemcpyHostToDevice, g_cudaStream));
-
-                dequantize_row_q_cuda(d_Q, d_X, ne01 * ne00, g_cudaStream);
-                CUDA_CHECK(cudaGetLastError());
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
                 const void* x = (char *) src0->data + i03*nb03 + i02*nb02;
 #else
                 {
@@ -8751,26 +8845,14 @@ static void ggml_compute_forward_mul_mat_q_f32(
                         dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
                         id += ne00;
                     }
+
+                    assert(id*sizeof(float) <= params->wsize);
                 }
+
                 const float * x = wdata;
 #endif
 
-
-#if defined(GGML_USE_CUBLAS)
-                // copy data to device
-                CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(float) * y_ne, cudaMemcpyHostToDevice, g_cudaStream));
-
-                // compute
-                CUBLAS_CHECK(
-                    cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            &alpha, d_X, ne00,
-                                    d_Y, ne10,
-                            &beta,  d_D, ne01));
-
-                // copy data to host
-                CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, g_cudaStream));
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
                 // zT = y * xT
                 ggml_cl_sgemm_wrapper(GGML_BLAS_ORDER_ROW_MAJOR, GGML_BLAS_OP_N, GGML_BLAS_OP_T,
                         ne11, ne01, ne10,
@@ -8788,13 +8870,6 @@ static void ggml_compute_forward_mul_mat_q_f32(
             }
         }
 
-#if defined(GGML_USE_CUBLAS)
-        CUDA_CHECK(cudaStreamSynchronize(g_cudaStream));
-        ggml_cuda_pool_free(d_X, x_size);
-        ggml_cuda_pool_free(d_Y, y_size);
-        ggml_cuda_pool_free(d_D, d_size);
-        ggml_cuda_pool_free(d_Q, q_size);
-#endif
         //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
 
         return;
@@ -8883,7 +8958,6 @@ static void ggml_compute_forward_mul_mat(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
-        case GGML_TYPE_Q4_3:
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
@@ -9115,7 +9189,6 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
-        case GGML_TYPE_Q4_3:
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
@@ -9300,6 +9373,161 @@ static void ggml_compute_forward_soft_max(
     }
 }
 
+// ggml_compute_forward_alibi
+
+static void ggml_compute_forward_alibi_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    assert(params->ith == 0);
+    assert(src1->type == GGML_TYPE_I32);
+    assert(ggml_nelements(src1) == 2);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int n_past = ((int32_t *) src1->data)[0];
+    const int n_head = ((int32_t *) src1->data)[1];
+
+    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
+    const int ne1 = src0->ne[1]; // seq_len_without_past
+    //const int ne2 = src0->ne[2]; // n_head -> this is k
+    //const int ne3 = src0->ne[3]; // 1 -> bsz
+
+    const int n  = ggml_nrows(src0);
+    const int ne2_ne3 = n/ne1; // ne2*ne3
+
+    const int nb0 = src0->nb[0];
+    const int nb1 = src0->nb[1];
+    const int nb2 = src0->nb[2];
+    //const int nb3 = src0->nb[3];
+
+    assert(nb0 == sizeof(float));
+    assert(ne1 + n_past == ne0); (void) n_past;
+
+    // add alibi to src0 (KQ_scaled)
+    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -8.0f / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -4.0f / n_heads_log2_floor);
+
+    for (int i = 0; i < ne0; i++) {
+        for (int j = 0; j < ne1; j++) {
+            for (int k = 0; k < ne2_ne3; k++) {
+                float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
+                float *      pdst = (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
+
+                // TODO: k*nb2 or k*nb3
+
+                float m_k;
+
+                if (k < n_heads_log2_floor) {
+                    m_k = powf(m0, k + 1);
+                } else {
+                    m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
+                }
+
+                pdst[0] = (j+1) * m_k + src[0];
+            }
+        }
+    }
+}
+
+
+static void ggml_compute_forward_alibi_f16(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    assert(params->ith == 0);
+    assert(src1->type == GGML_TYPE_I32);
+    assert(ggml_nelements(src1) == 2);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int n_past = ((int32_t *) src1->data)[0];
+    const int n_head = ((int32_t *) src1->data)[1];
+
+    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
+    const int ne1 = src0->ne[1]; // seq_len_without_past
+    //const int ne2 = src0->ne[2]; // n_head -> this is k
+    //const int ne3 = src0->ne[3]; // 1 -> bsz
+
+    const int n  = ggml_nrows(src0);
+    const int ne2_ne3 = n/ne1; // ne2*ne3
+
+    const int nb0 = src0->nb[0];
+    const int nb1 = src0->nb[1];
+    const int nb2 = src0->nb[2];
+    //const int nb3 = src0->nb[3];
+
+    assert(nb0 == sizeof(ggml_fp16_t));
+    assert(ne1 + n_past == ne0); (void) n_past;
+
+    // add alibi to src0 (KQ_scaled)
+    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -8.0f / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -4.0f / n_heads_log2_floor);
+
+    for (int i = 0; i < ne0; i++) {
+        for (int j = 0; j < ne1; j++) {
+            for (int k = 0; k < ne2_ne3; k++) {
+                ggml_fp16_t * const src  = (ggml_fp16_t *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
+                      float *      pdst  =       (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
+
+                // TODO: k*nb2 or k*nb3
+
+                float m_k;
+
+                if (k < n_heads_log2_floor) {
+                    m_k = powf(m0, k + 1);
+                } else {
+                    m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
+                }
+
+                // we return F32
+                pdst[0] = (j+1) * m_k + GGML_FP16_TO_FP32(src[0]);
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_alibi(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+        struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_alibi_f16(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_alibi_f32(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q4_2:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q8_1:
+        case GGML_TYPE_I8:
+        case GGML_TYPE_I16:
+        case GGML_TYPE_I32:
+        case GGML_TYPE_COUNT:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_rope
 
 static void ggml_compute_forward_rope_f32(
@@ -10938,6 +11166,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rope(params, tensor->src0, tensor->src1, tensor);
             } break;
+        case GGML_OP_ALIBI:
+            {
+                ggml_compute_forward_alibi(params, tensor->src0, tensor->src1, tensor);
+            } break;
         case GGML_OP_CONV_1D_1S:
             {
                 ggml_compute_forward_conv_1d_1s(params, tensor->src0, tensor->src1, tensor);
@@ -11140,6 +11372,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_ALIBI:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_SILU:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -11617,15 +11853,21 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 
                         size_t cur = 0;
 
+#if defined(GGML_USE_CUBLAS)
+                        if (ggml_cuda_can_mul_mat(node->src0, node->src1, node)) {
+                            node->n_tasks = 1; // TODO: this actually is doing nothing
+                                                //       the threads are still spinning
+                            cur = ggml_cuda_mul_mat_get_wsize(node->src0, node->src1, node);
+                        }
+                        else
+#endif
                         if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1; // TODO: this actually is doing nothing
                                                    //       the threads are still spinning
+                                // here we need memory just for single 2D matrix from src0
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
-                                //printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
-                                //printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
-                                //printf("cur = %zu\n", cur);
                             } else {
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
                             }
@@ -11634,8 +11876,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 #endif
                         } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+                            if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
+                                node->n_tasks = 1;
+                            }
+#endif
                         } else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
@@ -11673,6 +11920,10 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                     {
                         node->n_tasks = n_threads;
                     } break;
+                case GGML_OP_ALIBI:
+                    {
+                        node->n_tasks = 1; //TODO
+                    } break;
                 case GGML_OP_CONV_1D_1S:
                 case GGML_OP_CONV_1D_2S:
                     {
@@ -12060,10 +12311,16 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             snprintf(color, sizeof(color), "white");
         }
 
-        fprintf(fp, "  \"%p\" [ \
-style = filled; fillcolor = %s; shape = record; \
-label=\"%d [%" PRId64 ", %" PRId64 "] | <x>%s",
-                (void *) node, color,
+        fprintf(fp, "  \"%p\" [ "
+                    "style = filled; fillcolor = %s; shape = record; "
+                    "label=\"",
+                (void *) node, color);
+
+        if (strlen(node->name) > 0) {
+            fprintf(fp, "%s |", node->name);
+        }
+
+        fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s",
                 i, node->ne[0], node->ne[1],
                 GGML_OP_SYMBOL[node->op]);
 
@@ -12079,18 +12336,26 @@ label=\"%d [%" PRId64 ", %" PRId64 "] | <x>%s",
 
         snprintf(color, sizeof(color), "pink");
 
+        fprintf(fp, "  \"%p\" [ "
+                    "style = filled; fillcolor = %s; shape = record; "
+                    "label=\"<x>",
+                (void *) node, color);
+
+        if (strlen(node->name) > 0) {
+                fprintf(fp, "%s | ", node->name);
+        }
         if (ggml_nelements(node) == 1) {
-            fprintf(fp, "  \"%p\" [ \
-style = filled; fillcolor = %s; shape = record; \
-label=\"<x>%.1e\"; ]\n",
-                    (void *) node, color, (double)ggml_get_f32_1d(node, 0));
-        } else {
-            fprintf(fp, "  \"%p\" [ \
-style = filled; fillcolor = %s; shape = record; \
-label=\"<x>CONST %d [%" PRId64 ", %" PRId64 "]\"; ]\n",
-                    (void *) node, color,
-                    i, node->ne[0], node->ne[1]);
+            if (node->type == GGML_TYPE_I8 || node->type == GGML_TYPE_I16 || node->type == GGML_TYPE_I32) {
+                fprintf(fp, "%d", ggml_get_i32_1d(node, 0));
+            }
+            else {
+                fprintf(fp, "%.1e", (double)ggml_get_f32_1d(node, 0));
+            }
         }
+        else {
+            fprintf(fp, "CONST %d [%" PRId64 ", %" PRId64 "]", i, node->ne[0], node->ne[1]);
+        }
+        fprintf(fp, "\"; ]\n");
     }
 
     for (int i = 0; i < gb->n_nodes; i++) {
@@ -12889,29 +13154,6 @@ size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t *
     return (n/QK4_2*sizeof(block_q4_2));
 }
 
-size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK4_3 == 0);
-    const int nb = k / QK4_3;
-
-    for (int j = 0; j < n; j += k) {
-        block_q4_3 * restrict y = (block_q4_3 *)dst + j/QK4_3;
-
-        quantize_row_q4_3_reference(src + j, y, k);
-
-        for (int i = 0; i < nb; i++) {
-            for (int l = 0; l < QK4_3; l += 2) {
-                const uint8_t vi0 = y[i].qs[l/2] & 0x0F;
-                const uint8_t vi1 = y[i].qs[l/2] >> 4;
-
-                hist[vi0]++;
-                hist[vi1]++;
-            }
-        }
-    }
-
-    return (n/QK4_3*sizeof(block_q4_3));
-}
-
 size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK5_0 == 0);
     const int nb = k / QK5_0;
@@ -12926,8 +13168,8 @@ 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;
@@ -12956,8 +13198,8 @@ 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;
@@ -13014,12 +13256,6 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
                 block_q4_2 * block = (block_q4_2*)dst + start / QK4_2;
                 result = ggml_quantize_q4_2(src + start, block, n, n, hist);
             } break;
-        case GGML_TYPE_Q4_3:
-            {
-                GGML_ASSERT(start % QK4_3 == 0);
-                block_q4_3 * block = (block_q4_3*)dst + start / QK4_3;
-                result = ggml_quantize_q4_3(src + start, block, n, n, hist);
-            } break;
         case GGML_TYPE_Q5_0:
             {
                 GGML_ASSERT(start % QK5_0 == 0);