llama_cpp 0.3.5 → 0.3.6

data/ext/llama_cpp/src/ggml-cuda.h

@@ -27,6 +27,7 @@ void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
 void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);

data/ext/llama_cpp/src/ggml-metal.m

@@ -718,7 +718,8 @@ void ggml_metal_graph_compute(
                             // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
 
                             GGML_ASSERT(ne00 == ne10);
-                            GGML_ASSERT(ne02 == ne12);
+                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            GGML_ASSERT(ne03 == ne13);
 
                             if (ggml_is_contiguous(src0) &&
                                 ggml_is_contiguous(src1) &&
@@ -746,11 +747,11 @@ void ggml_metal_graph_compute(
                                     initWithDevice:ctx->device transposeLeft:false transposeRight:true
                                         resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
 
-                                // we need to do ne02 multiplications
+                                // we need to do ne12 multiplications
                                 // TODO: is there a way to do this in parallel - currently very slow ..
                                 // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
-                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
-                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                for (int64_t i02 = 0; i02 < ne12; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
                                     size_t offs_src1_cur = offs_src1 + i02*nb12;
                                     size_t offs_dst_cur  = offs_dst  + i02*nb2;
 
@@ -772,8 +773,6 @@ void ggml_metal_graph_compute(
                                 switch (src0t) {
                                     case GGML_TYPE_F16:
                                         {
-                                            GGML_ASSERT(ne02 == ne12);
-
                                             nth0 = 64;
                                             nth1 = 1;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
@@ -853,16 +852,18 @@ void ggml_metal_graph_compute(
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                                 [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
                                 [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
-                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
-                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {

data/ext/llama_cpp/src/ggml-metal.metal

@@ -509,11 +509,13 @@ kernel void kernel_mul_mat_f16_f32(
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
+        constant   int64_t & ne02,
         constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant   int64_t & ne10,
         constant   int64_t & ne11,
+        constant   int64_t & ne12,
         constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
@@ -529,7 +531,7 @@ kernel void kernel_mul_mat_f16_f32(
     const int64_t r1 = tgpig.y;
     const int64_t im = tgpig.z;
 
-    device const half  * x = (device const half  *) (src0 + r0*nb01 + im*nb02);
+    device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
     sum[tpitg.x] = 0.0f;
@@ -552,6 +554,7 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
+
 kernel void kernel_alibi_f32(
         device const float * src0,
         device       float * dst,

data/ext/llama_cpp/src/ggml.c

@@ -4557,10 +4557,12 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
 
 static struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_context * ctx,
-        enum   ggml_type type,
-        int    n_dims,
-        const int64_t* ne,
-        void*  data) {
+        enum   ggml_type      type,
+        int                   n_dims,
+        const int64_t       * ne,
+        void                * data) {
+
+    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
     size_t data_size = 0;
 
@@ -4648,22 +4650,22 @@ static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int3
 
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
-        enum   ggml_type type,
-        int    n_dims,
-        const int64_t * ne) {
+        enum   ggml_type      type,
+        int                   n_dims,
+        const int64_t       * ne) {
     return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0) {
     return ggml_new_tensor(ctx, type, 1, &ne0);
 }
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0,
         int64_t ne1) {
     const int64_t ne[2] = { ne0, ne1 };
@@ -4672,7 +4674,7 @@ struct ggml_tensor * ggml_new_tensor_2d(
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0,
         int64_t ne1,
         int64_t ne2) {
@@ -6238,6 +6240,27 @@ struct ggml_tensor * ggml_reshape_4d(
 
 // ggml_view_1d
 
+static struct ggml_tensor * ggml_view_tensor_offset(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_dims,
+        const int64_t       * ne,
+        size_t                offset) {
+    // don't calculate an offset from an unallocated tensor
+    void * data = NULL;
+    if (a->data != NULL) {
+        data = (char *) a->data + offset;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data);
+
+    ggml_format_name(result, "%s (view)", a->name);
+
+    ggml_set_op_params(result, &offset, sizeof(offset));
+
+    return result;
+}
+
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6250,10 +6273,7 @@ struct ggml_tensor * ggml_view_1d(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset);
 
     result->op   = GGML_OP_VIEW;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6280,10 +6300,7 @@ struct ggml_tensor * ggml_view_2d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = result->nb[1]*ne1;
@@ -6316,10 +6333,7 @@ struct ggml_tensor * ggml_view_3d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
@@ -6354,10 +6368,7 @@ struct ggml_tensor * ggml_view_4d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
@@ -6741,6 +6752,18 @@ struct ggml_tensor * ggml_rope_inplace(
     return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, true);
 }
 
+struct ggml_tensor * ggml_rope_custom(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past,
+        int                   n_dims,
+        int                   mode,
+        int                   n_ctx,
+        float                 freq_base,
+        float                 freq_scale) {
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, false);
+}
+
 struct ggml_tensor * ggml_rope_custom_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,

data/ext/llama_cpp/src/ggml.h

@@ -1170,7 +1170,18 @@ extern "C" {
             int                   mode,
             int                   n_ctx);
 
-    // custom RoPE, in-place, returns view(a)
+    // custom RoPE
+    GGML_API struct ggml_tensor * ggml_rope_custom(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
+    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

data/ext/llama_cpp/src/k_quants.c

@@ -39,6 +39,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -1353,7 +1355,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -1421,7 +1423,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
 
         // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(_mm256_set_m128i(summs_1, summs_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
 
         const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
         const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
@@ -1493,7 +1495,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -1644,8 +1646,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * smin;
 
         const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m256i q2_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 2), q2bits), m3);
-        const __m256i q2_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
+        const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
+        const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
 
         const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
         const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
@@ -1709,10 +1711,10 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
         const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
 
-        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
-        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
-        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
-        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+        const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
 
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
@@ -1917,7 +1919,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         // high bit
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
@@ -2128,7 +2130,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // multiply with block scale and accumulate
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2303,13 +2305,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         aux16[0] = a & 0x0f0f;
         aux16[1] = (a >> 4) & 0x0f0f;
 
-        const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
-        const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
+        const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
+        const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
 
         memcpy(&aux64, x[i].hmask, 8);
 
         const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m256i q3h_0 = _mm256_set_m128i(_mm_srli_epi16(haux, 2), haux);
+        __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
         __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
         q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
         q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
@@ -2318,7 +2320,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
 
         // prepare low and high bits
-        const __m256i q3aux  = _mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits);
+        const __m256i q3aux  = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
         const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
         const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
 
@@ -2429,7 +2431,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
         p16_0 = _mm_add_epi32(p16_0, p16_2);
         p16_1 = _mm_add_epi32(p16_1, p16_3);
-        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+        __m256i p16 = MM256_SET_M128I(p16_1, p16_0);
 
         // multiply with block scale and accumulate
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
@@ -2620,7 +2622,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -2727,7 +2729,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2968,11 +2970,11 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
         const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
 
         const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
         const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
 
     }
 
@@ -3160,7 +3162,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * _mm_extract_epi32(hsum, 0);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
         __m256i hmask = mone;
@@ -3299,7 +3301,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -3462,13 +3464,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
 
-        const __m256i scale_l = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
-        const __m256i scale_h = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
+        const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
+        const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
 
         int64_t aux64;
         memcpy(&aux64, x[i].qh, 8);
         const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m256i haux256 = _mm256_set_m128i(_mm_srli_epi16(haux128, 2), haux128);
+        const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
 
         const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
         const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
@@ -3543,7 +3545,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
         const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
 
     }
 
@@ -3925,7 +3927,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
         }
 
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -4083,8 +4085,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
         const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
 
-        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
-        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
+        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
+        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
 
         const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
         const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
@@ -4177,7 +4179,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
         sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
     }
 
     *s = hsum_float_8(acc);