@fugood/llama.node 0.3.1 → 0.3.3

package/src/llama.cpp/ggml/src/ggml-sycl/im2col.cpp

@@ -0,0 +1,125 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "im2col.hpp"
+
+template <typename T>
+static void im2col_kernel(
+        const float *x, T *dst, int64_t batch_offset, int64_t offset_delta,
+        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
+        int64_t pelements, int64_t CHW, int s0, int s1, int p0, int p1, int d0, int d1,
+        const sycl::nd_item<3> &item_ct1) {
+    const int64_t work_group_size = item_ct1.get_local_range(2);
+    const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
+
+    // make each work-item deal with more elements since sycl global range can not exceed max int
+    for (int64_t i = global_id; i < pelements; i += work_group_size * item_ct1.get_group_range(2)) {
+
+        const int64_t ksize = OW * (KH > 1 ? KW : 1);
+        const int64_t kx = i / ksize;
+        const int64_t kd = kx * ksize;
+        const int64_t ky = (i - kd) / OW;
+        const int64_t ix = i % OW;
+
+        const int64_t  oh = item_ct1.get_group(1);
+        const int64_t  batch = item_ct1.get_group(0) / IC;
+        const int64_t  ic = item_ct1.get_group(0) % IC;
+
+        const int64_t iiw = ix * s0 + kx * d0 - p0;
+        const int64_t iih = oh * s1 + ky * d1 - p1;
+
+        const int64_t offset_dst =
+            ((batch * OH + oh) * OW + ix) * CHW +
+            (ic * (KW * KH) + ky * KW + kx);
+
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+            dst[offset_dst] =
+                sycl::vec<float, 1>(0.0f)
+                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        } else {
+            const int64_t offset_src = ic * offset_delta + batch * batch_offset;
+            dst[offset_dst] =
+                sycl::vec<float, 1>(x[offset_src + iih * IW + iiw])
+                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        }
+    }
+}
+
+template <typename T>
+static void im2col_sycl(
+        const float *x, T *dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+        int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
+        int s0, int s1, int p0, int p1, int d0, int d1,
+        queue_ptr stream) {
+    const int64_t parallel_elements = OW * KW * KH;
+    const int64_t num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+
+    // decrease global range when it exceeds the max int
+    int64_t local_size = downsample_sycl_global_range(batch * IC * OH * num_blocks, SYCL_IM2COL_BLOCK_SIZE);
+    sycl::range<3> block_nums(batch * IC, OH, num_blocks);
+    sycl::range<3> local_range(1, 1, local_size);
+
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * local_range, local_range),
+            [=](sycl::nd_item<3> item_ct1) {
+                im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH,
+                               parallel_elements, (IC * KH * KW), s0, s1, p0,
+                               p1, d0, d1, item_ct1);
+            });
+    }
+}
+
+void ggml_sycl_op_im2col(
+        ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+        ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
+        const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const int64_t IC = src1->ne[is_2D ? 2 : 1];
+    const int64_t IH = is_2D ? src1->ne[1] : 1;
+    const int64_t IW =         src1->ne[0];
+
+    const int64_t KH = is_2D ? src0->ne[1] : 1;
+    const int64_t KW =         src0->ne[0];
+
+    const int64_t OH = is_2D ? dst->ne[2] : 1;
+    const int64_t OW =         dst->ne[1];
+
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch = src1->ne[3];
+    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    if (dst->type == GGML_TYPE_F16) {
+        im2col_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
+    } else {
+        im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
+    }
+
+    (void) src0;
+    (void) src0_dd;
+}

package/src/llama.cpp/ggml/src/ggml-sycl/im2col.hpp

@@ -0,0 +1,23 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_IM2COL_HPP
+#define GGML_SYCL_IM2COL_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_im2col(
+        ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+        ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
+        const queue_ptr &main_stream);
+
+#endif // GGML_SYCL_IM2COL_HPP

package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp

@@ -1,6 +1,6 @@
 #include "mmvq.hpp"
 #include "vecdotq.hpp"
-
+#include <cassert>
 
 template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_sycl_t vec_dot_q_sycl>
 static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
@@ -13,7 +13,8 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 
 // partial sum for each thread
     float tmp = 0.0f;
@@ -37,7 +38,7 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -61,7 +62,8 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 
 // partial sum for each thread
     float tmp = 0.0f;
@@ -85,7 +87,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -109,8 +111,8 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -133,7 +135,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -157,8 +159,8 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -181,7 +183,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -205,8 +207,8 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -229,7 +231,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -253,8 +255,8 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -277,7 +279,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -301,8 +303,8 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -325,7 +327,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -349,8 +351,8 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -373,7 +375,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -397,8 +399,8 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -421,7 +423,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -446,8 +448,8 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
     }
 
     const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
+    const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
+    assert(blocks_per_warp>0);
 // partial sum for each thread
     float tmp = 0.0f;
 
@@ -470,7 +472,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -487,7 +489,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK4_0 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -495,7 +497,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK4_0, QI4_0, block_q4_0,
                                       VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -511,7 +513,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK4_1 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -519,7 +521,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
                                       VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -535,7 +537,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK5_0 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -543,7 +545,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
                                       VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -559,7 +561,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK5_1 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -567,7 +569,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
                                       VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -583,7 +585,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK8_0 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -591,7 +593,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
                                       VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -607,7 +609,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -615,7 +617,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
                                       VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -631,7 +633,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -639,7 +641,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
                                       VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -655,7 +657,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -663,7 +665,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
                                       VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -679,7 +681,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -687,7 +689,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
                                       VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -703,7 +705,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -711,7 +713,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
                                       VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
@@ -728,13 +730,13 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
         stream->submit([&](sycl::handler &cgh) {
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -749,7 +751,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -759,7 +761,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -774,7 +776,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -784,7 +786,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -799,7 +801,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -809,7 +811,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -824,7 +826,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -833,7 +835,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -848,7 +850,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
@@ -858,7 +860,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -873,13 +875,13 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
         stream->submit([&](sycl::handler &cgh) {
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
@@ -894,15 +896,15 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK4_NL == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 1>(
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });
         });
@@ -916,14 +918,14 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
     GGML_ASSERT(ncols % QK_K == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
 
         stream->submit([&](sycl::handler &cgh) {
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
-                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
                         mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
                             vx, vy, dst, ncols, nrows, item_ct1);
                     });