llama_cpp 0.1.1 → 0.1.2

data/ext/llama_cpp/src/ggml.c

@@ -512,7 +512,7 @@ static inline int hsum_i32_4(const __m128i a) {
     return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
 }
 
-#if __AVX2__ || __AVX512F__
+#if defined(__AVX2__) || defined(__AVX512F__)
 // spread 32 bits to 32 bytes { 0x00, 0xFF }
 static inline __m256i bytes_from_bits_32(const uint8_t * x) {
     uint32_t x32;
@@ -688,7 +688,7 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 #endif // __AVX__ || __AVX2__ || __AVX512F__
 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 
-#if __ARM_NEON
+#if defined(__ARM_NEON)
 
 #if !defined(__aarch64__)
 
@@ -740,19 +740,19 @@ inline static float vaddvq_f32(float32x4_t v) {
     return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
 }
 
-float vminvq_f32(float32x4_t v) {
+inline static float vminvq_f32(float32x4_t v) {
     return
         MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
             MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
 }
 
-float vmaxvq_f32(float32x4_t v) {
+inline static float vmaxvq_f32(float32x4_t v) {
     return
         MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
             MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
 }
 
-int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
     int32x4_t res;
 
     res[0] = roundf(vgetq_lane_f32(v, 0));
@@ -766,21 +766,20 @@ int32x4_t vcvtnq_s32_f32(float32x4_t v) {
 #endif
 #endif
 
-
 #define QK4_0 32
 typedef struct {
-    float   d;          // delta
+    ggml_fp16_t d;          // delta
     uint8_t qs[QK4_0 / 2];  // nibbles / quants
 } block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
 
 #define QK4_1 32
 typedef struct {
-    float   d;          // delta
-    float   m;          // min
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding");
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
 
 #define QK5_0 32
 typedef struct {
@@ -801,16 +800,16 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
 
 #define QK8_0 32
 typedef struct {
-    float   d;          // delta
-    int8_t  qs[QK8_0];  // quants
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
 #define QK8_1 32
 typedef struct {
-    float   d;         // delta
-    float   s;         // d * sum(qs[i])
-    int8_t  qs[QK8_1]; // quants
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
 
@@ -837,7 +836,7 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r
         const float d  = max / -8;
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         for (int j = 0; j < qk/2; ++j) {
             const float x0 = x[i*qk + 0    + j]*id;
@@ -877,8 +876,8 @@ static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * r
         const float d  = (max - min) / ((1 << 4) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
-        y[i].m = min;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].m = GGML_FP32_TO_FP16(min);
 
         for (int j = 0; j < qk/2; ++j) {
             const float x0 = (x[i*qk + 0    + j] - min)*id;
@@ -1009,7 +1008,7 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         for (int j = 0; j < QK8_0; ++j) {
             const float x0 = x[i*QK8_0 + j]*id;
@@ -1044,7 +1043,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const float32x4_t v  = vmulq_n_f32(srcv[j], id);
@@ -1056,6 +1055,39 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
             y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
         }
     }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float 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)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+        }
+    }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
         // Load elements into 4 AVX vectors
@@ -1079,7 +1111,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
 
         // Quantize these floats
         const float d = maxScalar / 127.f;
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
         const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -1178,7 +1210,7 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r
             sum += y[i].qs[QK8_1/2 + j];
         }
 
-        y[i].s = d * sum;
+        y[i].s = sum*d;
     }
 }
 
@@ -1224,6 +1256,48 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int
 
         y[i].s = d * vaddvq_s32(accv);
     }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float 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)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+
+        v128_t accv = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+
+            accv = wasm_i32x4_add(accv, vi);
+        }
+
+        y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
+                      wasm_i32x4_extract_lane(accv, 1) +
+                      wasm_i32x4_extract_lane(accv, 2) +
+                      wasm_i32x4_extract_lane(accv, 3));
+    }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
         // Load elements into 4 AVX vectors
@@ -1330,7 +1404,7 @@ static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict
     const int nb = k / qk;
 
     for (int i = 0; i < nb; i++) {
-        const float d = x[i].d;
+        const float d = GGML_FP16_TO_FP32(x[i].d);
 
         for (int j = 0; j < qk/2; ++j) {
             const int x0 = (x[i].qs[j] & 0x0F) - 8;
@@ -1350,8 +1424,8 @@ static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict
     const int nb = k / qk;
 
     for (int i = 0; i < nb; i++) {
-        const float d = x[i].d;
-        const float m = x[i].m;
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float m = GGML_FP16_TO_FP32(x[i].m);
 
         for (int j = 0; j < qk/2; ++j) {
             const int x0 = (x[i].qs[j] & 0x0F);
@@ -1426,7 +1500,7 @@ static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, in
     const block_q8_0 * restrict x = vx;
 
     for (int i = 0; i < nb; i++) {
-        const float d = x[i].d;
+        const float d = GGML_FP16_TO_FP32(x[i].d);
 
         for (int j = 0; j < qk; ++j) {
             y[i*qk + j] = x[i].qs[j]*d;
@@ -1690,8 +1764,9 @@ quantize_fns_t ggml_internal_get_quantize_fn(size_t i) {
 static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
-    for (int i = 0; i < 8; i++)
+    for (int i = 0; i < 8; i++) {
         tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
 
     return _mm256_loadu_ps(tmp);
 }
@@ -2111,8 +2186,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
-        const uint8x16_t m4b   = vdupq_n_u8(0x0F);
-        const int8x16_t  s8b   = vdupq_n_s8(0x8);
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+        const int8x16_t  s8b = vdupq_n_s8(0x8);
 
         const uint8x16_t v0_0 = vld1q_u8(x0->qs);
         const uint8x16_t v0_1 = vld1q_u8(x1->qs);
@@ -2140,8 +2215,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
         const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, 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);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
@@ -2158,8 +2233,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
         const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 #endif
     }
 
@@ -2171,7 +2246,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     // Main loop
     for (int i = 0; i < nb; ++i) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
 
         __m256i bx = bytes_from_nibbles_32(x[i].qs);
 
@@ -2195,7 +2270,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     // Main loop
     for (int i = 0; i < nb; ++i) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
 
         const __m128i lowMask = _mm_set1_epi8(0xF);
         const __m128i off = _mm_set1_epi8(8);
@@ -2237,7 +2312,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) );
 
         const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
 
@@ -2255,7 +2330,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) );
 
         const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);
 
@@ -2288,7 +2363,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
 
         const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);
 
@@ -2306,7 +2381,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) );
 
         const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);
 
@@ -2354,7 +2429,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
             sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
         }
 
-        sumf += (x[i].d*y[i].d)*sumi;
+        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
     }
 
     *s = sumf;
@@ -2384,7 +2459,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
         const block_q8_1 * restrict y0 = &y[i + 0];
         const block_q8_1 * restrict y1 = &y[i + 1];
 
-        summs += x0->m * y0->s + x1->m * y1->s;
+        summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -2408,8 +2483,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
         const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
         const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, 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);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l));
@@ -2426,8 +2501,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
         const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
         const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
 #endif
     }
 
@@ -2440,13 +2515,13 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
 
     // Main loop
     for (int i = 0; i < nb; ++i) {
-        const float * d0 = &x[i].d;
-        const float * d1 = &y[i].d;
+        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+        const float d1 = y[i].d;
 
-        summs += x[i].m * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
 
-        const __m256 d0v = _mm256_broadcast_ss( d0 );
-        const __m256 d1v = _mm256_broadcast_ss( d1 );
+        const __m256 d0v = _mm256_set1_ps( d0 );
+        const __m256 d1v = _mm256_set1_ps( d1 );
 
         // Compute combined scales
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
@@ -2480,7 +2555,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
             sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
         }
 
-        sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
     }
 
     *s = sumf;
@@ -2556,16 +2631,13 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v1_1l = vld1q_s8(y1->qs);
         const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        const float x0d = GGML_FP16_TO_FP32(x0->d);
-        const float x1d = GGML_FP16_TO_FP32(x1->d);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d);
+                        vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d);
+                        vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
@@ -2582,8 +2654,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
         const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 #endif
     }
 
@@ -2600,7 +2672,6 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         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
         memcpy(&qh, x0->qh, sizeof(qh));
@@ -2638,15 +2709,14 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         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)));
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
     }
 
     *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -2658,7 +2728,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     // Main loop
     for (int i = 0; i < nb; i++) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d));
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
 
         __m256i bx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
@@ -2682,7 +2752,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     // Main loop
     for (int i = 0; i < nb; i++) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d));
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
 
         __m256i bx = bytes_from_nibbles_32(x[i].qs);
         const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@@ -2725,7 +2795,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
             sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
     }
 
     *s = sumf;
@@ -2807,16 +2877,13 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         const int8x16_t v1_1l = vld1q_s8(y1->qs);
         const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        const float x0d = GGML_FP16_TO_FP32(x0->d);
-        const float x1d = GGML_FP16_TO_FP32(x1->d);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d);
+                        vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d);
+                        vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
 #else
         const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
@@ -2833,8 +2900,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
         const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
 #endif
     }
 
@@ -2873,8 +2940,6 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         const v128_t v0l = wasm_v128_and (v0, m4b);
         const v128_t v0h = wasm_u8x16_shr(v0, 4);
 
-        static bool x = true;
-
         // add high bit
         const v128_t v0lf = wasm_v128_or(v0l, qhl);
         const v128_t v0hf = wasm_v128_or(v0h, qhh);
@@ -2894,15 +2959,14 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         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(
+        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)));
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
     }
 
     *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -2924,7 +2988,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
         bx = _mm256_or_si256(bx, bxhi);
 
-        const __m256 dy = _mm256_broadcast_ss(&y[i].d);
+        const __m256 dy = _mm256_set1_ps(y[i].d);
         const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx, by);
@@ -2958,7 +3022,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         bxh = _mm_or_si128(bxh, bxhih);
         bx = _mm256_set_m128i(bxh, bxl);
 
-        const __m256 dy = _mm256_broadcast_ss(&y[i].d);
+        const __m256 dy = _mm256_set1_ps(y[i].d);
         const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx, by);
@@ -3028,11 +3092,11 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
 #if defined(__ARM_FEATURE_DOTPROD)
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), x0->d*y0->d);
+                        vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
 
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), x1->d*y1->d);
+                        vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 
 #else
         const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0));
@@ -3050,8 +3114,8 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
         const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
         const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), x0->d*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), x1->d*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
 #endif
     }
 
@@ -3063,7 +3127,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
     // Main loop
     for (int i = 0; i < nb; ++i) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
         __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
         __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
@@ -3089,7 +3153,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
             sumi += x[i].qs[j]*y[i].qs[j];
         }
 
-        sumf += (x[i].d*y[i].d)*sumi;
+        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
     }
 
     *s = sumf;
@@ -3478,6 +3542,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "ROPE",
     "ROPE_BACK",
     "ALIBI",
+    "CLAMP",
     "CONV_1D_1S",
     "CONV_1D_2S",
 
@@ -3488,7 +3553,8 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "MAP_BINARY",
 };
 
-static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50");
+static_assert(GGML_OP_COUNT == 51, "GGML_OP_COUNT != 51");
+
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3538,6 +3604,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "rope(x)",
     "rope_back(x)",
     "alibi(x)",
+    "clamp(x)",
     "conv_1d_1s(x)",
     "conv_1d_2s(x)",
 
@@ -3548,7 +3615,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "f(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50");
+static_assert(GGML_OP_COUNT == 51, "GGML_OP_COUNT != 51");
 
 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");
@@ -3782,6 +3849,12 @@ static inline bool ggml_can_repeat(const struct ggml_tensor * t0, const struct g
         (t1->ne[3]%t0->ne[3] == 0);
 }
 
+static inline bool ggml_can_repeat_rows(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return (t0->ne[0] == t1->ne[0]) && ggml_can_repeat(t0, t1);
+}
+
 static inline int ggml_up32(int n) {
     return (n + 31) & ~31;
 }
@@ -4664,11 +4737,15 @@ struct ggml_tensor * ggml_mul_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    GGML_ASSERT(ggml_are_same_shape(a, b));
+    // TODO: support less-strict constraint
+    //       GGML_ASSERT(ggml_can_repeat(b, a));
+    GGML_ASSERT(ggml_can_repeat_rows(b, a));
 
     bool is_node = false;
 
     if (!inplace && (a->grad || b->grad)) {
+        // TODO: support backward pass for broadcasting
+        GGML_ASSERT(ggml_are_same_shape(a, b));
         is_node = true;
     }
 
@@ -6210,7 +6287,8 @@ struct ggml_tensor * ggml_alibi(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         int                   n_past,
-        int                   n_head) {
+        int                   n_head,
+        float                 bias_max) {
     GGML_ASSERT(n_past >= 0);
     bool is_node = false;
 
@@ -6229,6 +6307,8 @@ struct ggml_tensor * ggml_alibi(
 
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_head;
+    GGML_ASSERT(sizeof(float) == sizeof(int32_t));
+    (((float *) b->data)[2]) = bias_max;
 
     ggml_scratch_load(ctx);
 
@@ -6240,6 +6320,40 @@ struct ggml_tensor * ggml_alibi(
     return result;
 }
 
+// ggml_clamp
+
+struct ggml_tensor * ggml_clamp(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 min,
+        float                 max) {
+    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 = ggml_view_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3);
+
+    ((float *) b->data)[0] = min;
+    ((float *) b->data)[1] = max;
+
+    ggml_scratch_load(ctx);
+
+    result->op   = GGML_OP_CLAMP;
+    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(
@@ -7966,7 +8080,7 @@ static void ggml_compute_forward_mul_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat_rows(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -7974,10 +8088,25 @@ static void ggml_compute_forward_mul_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int nr  = ggml_nrows(src0);
-    const int64_t ne0 = src0->ne[0];
-    const int64_t ne1 = src0->ne[1];
-    const int64_t ne2 = src0->ne[2];
+#ifdef GGML_USE_CUBLAS
+    if (src1->backend == GGML_BACKEND_CUDA) {
+        if (ith == 0) {
+            ggml_cuda_mul(src0, src1, dst);
+        }
+        return;
+    }
+#endif
+
+    const int64_t nr = ggml_nrows(src0);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const size_t nb00 = src0->nb[0];
     const size_t nb01 = src0->nb[1];
@@ -7996,44 +8125,51 @@ static void ggml_compute_forward_mul_f32(
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(ne00 == ne10);
 
     if (nb10 == sizeof(float)) {
-        for (int ir = ith; ir < nr; ir += nth) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
 
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
 #ifdef GGML_USE_ACCELERATE
             UNUSED(ggml_vec_mul_f32);
 
-            vDSP_vmul(
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ), 1,
-                    ne0);
+            vDSP_vmul( src0_ptr, 1, src1_ptr, 1, dst_ptr,  1, ne00);
 #else
-            ggml_vec_mul_f32(ne0,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+            ggml_vec_mul_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
 #endif
                 // }
             // }
         }
     } else {
         // src1 is not contiguous
-        for (int ir = ith; ir < nr; ir += nth) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-            float * dst_ptr  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
-            float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+            for (int64_t i0 = 0; i0 < ne00; i0++) {
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr);
             }
@@ -10527,6 +10663,7 @@ static void ggml_compute_forward_diag_mask_f32(
 
     const int  n_past  =       ((int32_t *) src1->data)[0];
     const bool inplace = (bool)((int32_t *) src1->data)[1];
+
     assert(n_past >= 0);
 
     if (!inplace && (params->type == GGML_TASK_INIT)) {
@@ -10697,14 +10834,15 @@ static void ggml_compute_forward_alibi_f32(
         struct ggml_tensor * dst) {
     assert(params->ith == 0);
     assert(src1->type == GGML_TYPE_I32);
-    assert(ggml_nelements(src1) == 2);
+    assert(ggml_nelements(src1) == 3);
 
     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   n_past   = ((int32_t *) src1->data)[0];
+    const int   n_head   = ((int32_t *) src1->data)[1];
+    const float max_bias = ((float *)   src1->data)[2];
 
     assert(n_past >= 0);
 
@@ -10727,8 +10865,8 @@ static void ggml_compute_forward_alibi_f32(
     // 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);
+    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
     for (int i = 0; i < ne0; i++) {
         for (int j = 0; j < ne1; j++) {
@@ -10746,13 +10884,13 @@ static void ggml_compute_forward_alibi_f32(
                     m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
                 }
 
-                pdst[0] = i * m_k + src[0];
+                pdst[0] = (i-ne0+1) * m_k + src[0];
+
             }
         }
     }
 }
 
-
 static void ggml_compute_forward_alibi_f16(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -10760,14 +10898,15 @@ static void ggml_compute_forward_alibi_f16(
         struct ggml_tensor * dst) {
     assert(params->ith == 0);
     assert(src1->type == GGML_TYPE_I32);
-    assert(ggml_nelements(src1) == 2);
+    assert(ggml_nelements(src1) == 3);
 
     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   n_past   = ((int32_t *) src1->data)[0];
+    const int   n_head   = ((int32_t *) src1->data)[1];
+    const float max_bias = ((float *)   src1->data)[2];
 
     assert(n_past >= 0);
 
@@ -10790,8 +10929,8 @@ static void ggml_compute_forward_alibi_f16(
     // 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);
+    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
     for (int i = 0; i < ne0; i++) {
         for (int j = 0; j < ne1; j++) {
@@ -10810,7 +10949,7 @@ static void ggml_compute_forward_alibi_f16(
                 }
 
                 // we return F32
-                pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]);
+                pdst[0] = (i-ne0+1) * m_k + GGML_FP16_TO_FP32(src[0]);
             }
         }
     }
@@ -10846,6 +10985,77 @@ static void ggml_compute_forward_alibi(
     }
 }
 
+
+// ggml_compute_forward_clamp
+
+static void ggml_compute_forward_clamp_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 min = ((float *) src1->data)[0];
+    const int max = ((float *) src1->data)[1];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    const size_t nb00 = src0->nb[0];
+    const size_t nb01 = src0->nb[1];
+
+    const size_t nb0 = dst->nb[0];
+    const size_t nb1 = dst->nb[1];
+
+    GGML_ASSERT( nb0 == sizeof(float));
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    for (int j = ith; j < n; j += nth) {
+        float * dst_ptr  = (float *) ((char *)  dst->data + j*nb1);
+        float * src0_ptr = (float *) ((char *) src0->data + j*nb01);
+
+        for (int i = 0; i < nc; i++) {
+            dst_ptr[i] = MAX(MIN(src0_ptr[i], max), min);
+        }
+    }
+}
+
+static void ggml_compute_forward_clamp(
+        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_F32:
+            {
+                ggml_compute_forward_clamp_f32(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F16:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        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(
@@ -12827,6 +13037,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_alibi(params, tensor->src0, tensor->src1, tensor);
             } break;
+        case GGML_OP_CLAMP:
+            {
+                ggml_compute_forward_clamp(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);
@@ -13134,6 +13348,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_CLAMP:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_SILU:
             {
                 // necessary for llama
@@ -14013,6 +14231,10 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                     {
                         node->n_tasks = 1; //TODO
                     } break;
+                case GGML_OP_CLAMP:
+                    {
+                        node->n_tasks = 1; //TODO
+                    } break;
                 case GGML_OP_CONV_1D_1S:
                 case GGML_OP_CONV_1D_2S:
                     {
@@ -14409,9 +14631,12 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             fprintf(fp, "%s |", node->name);
         }
 
-        fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s",
-                i, node->ne[0], node->ne[1],
-                GGML_OP_SYMBOL[node->op]);
+        if (node->n_dims == 2) {
+            fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], GGML_OP_SYMBOL[node->op]);
+        } else {
+            fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], GGML_OP_SYMBOL[node->op]);
+        }
+
 
         if (node->grad) {
             fprintf(fp, " | <g>%s\"; ]\n", GGML_OP_SYMBOL[node->grad->op]);