@fugood/llama.node 0.3.17 → 0.4.0

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

@@ -49,6 +49,8 @@ static bool g_sycl_loaded = false;
 int g_ggml_sycl_debug = 0;
 int g_ggml_sycl_disable_optimize = 0;
 int g_ggml_sycl_disable_graph = 0;
+int g_ggml_sycl_disable_dnn = 0;
+int g_ggml_sycl_prioritize_dmmv = 0;
 
 static ggml_sycl_device_info ggml_sycl_init() {
     ggml_sycl_device_info info = {};
@@ -193,13 +195,25 @@ static void ggml_check_sycl() try {
 
     if (!initialized) {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
+        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
+        g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
+        g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         GGML_LOG_INFO("Running with Environment Variables:\n");
         GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
+#ifdef GGML_SYCL_GRAPH
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: graph disabled by compile flag\n");
+#endif
+#if GGML_SYCL_DNNL
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: %d\n", g_ggml_sycl_disable_dnn);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n");
+#endif
+        GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
         GGML_LOG_INFO("Build with Macros:\n");
 #if defined(GGML_SYCL_FORCE_MMQ)
         GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
@@ -338,7 +352,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         assert(tensor->view_src->buffer->buft == buffer->buft);
         return GGML_STATUS_SUCCESS;
     }
-    if (tensor->type == GGML_TYPE_Q4_0) {
+    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K) && !g_ggml_sycl_disable_optimize) {
         ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
         tensor->extra                 = extra;
         ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
@@ -1982,19 +1996,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne10 = src1->ne[0];
-
+    GGML_ASSERT(ne00 == ne10);
 
     const int64_t row_diff = row_high - row_low;
 
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
-#if !GGML_SYCL_DNNL
-    const int64_t ne0 = dst->ne[0];
+
+    const int64_t ne0 = dst->ne[0]; // used by MKL only
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = id == ctx.device ? ne0 : row_diff;
-#endif
+    int ldc = id == ctx.device ? ne0 : row_diff; // used by MKL only
 
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -2030,25 +2043,29 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
         ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
-#if !GGML_SYCL_DNNL
-        const sycl::half alpha_f16 = 1.0f;
-        const sycl::half beta_f16  = 0.0f;
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
-            *stream, oneapi::math::transpose::trans,
-            oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
-            &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
-            src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
-            dst_f16.get(), dpct::library_data_t::real_half, ldc,
-            dpct::library_data_t::real_half)));
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-#else
-        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ptr,
-                                  DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-                                  dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+#if GGML_SYCL_DNNL
+        if (!g_ggml_sycl_disable_dnn) {
+            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
+                                      DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                                      dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
+            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+        }
+        else
 #endif
+        {
+            const sycl::half alpha_f16 = 1.0f;
+            const sycl::half beta_f16  = 0.0f;
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
+                *stream, oneapi::math::transpose::trans,
+                oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
+                &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
+                src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
+                dst_f16.get(), dpct::library_data_t::real_half, ldc,
+                dpct::library_data_t::real_half)));
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
+            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
+        }
     }
     else {
         // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp32 path\n");
@@ -2069,18 +2086,22 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
-#if !GGML_SYCL_DNNL
-        const float alpha = 1.0f;
-        const float beta  = 0.0f;
-        SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
-            get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
-            src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
-            dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
-#else
-        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i,
-                                  DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
-                                  dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
+#if GGML_SYCL_DNNL
+        if (!g_ggml_sycl_disable_dnn) {
+            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ddf1_i,
+                                      DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
+        }
+        else
 #endif
+        {
+            const float alpha = 1.0f;
+            const float beta  = 0.0f;
+            SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
+                get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
+                src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
+                dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
+        }
     }
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddq_i);
@@ -2694,139 +2715,180 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-static void k_compute_batched_ptrs(const sycl::half *src0_as_f16,
-                                   const sycl::half *src1_as_f16, char *dst,
-                                   const void **ptrs_src, void **ptrs_dst,
-                                   int64_t ne12, int64_t ne13, int64_t ne23,
-                                   size_t nb02, size_t nb03, size_t nb12,
-                                   size_t nb13, size_t nbd2, size_t nbd3,
-                                   int64_t r2, int64_t r3,
-                                   const sycl::nd_item<3> &item_ct1) {
-    int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) +
-                  item_ct1.get_local_id(2);
-    int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) +
-                  item_ct1.get_local_id(1);
+static void k_compute_batched_ptrs(const sycl::half * src0_as_f16, const sycl::half * src1_as_f16, void * dst,
+                                   const void ** ptrs_src, void ** ptrs_dst, int64_t ne12, int64_t ne13, int64_t ne23,
+                                   size_t nb02, size_t nb03, size_t nb12, size_t nb13, size_t nbd2, size_t nbd3,
+                                   int64_t r2, int64_t r3, const sycl::nd_item<3> & item_ct1) {
+    const int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + item_ct1.get_local_id(2);
+    const int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1);
 
     if (i13 >= ne13 || i12 >= ne12) {
         return;
     }
 
-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
+    const int64_t i03 = i13 / r3;
+    const int64_t i02 = i12 / r2;
+
+    const uint8_t * src0_bytes = reinterpret_cast<const uint8_t *>(src0_as_f16);
+    const uint8_t * src1_bytes = reinterpret_cast<const uint8_t *>(src1_as_f16);
+    uint8_t *       dst_bytes  = static_cast<uint8_t *>(dst);
 
-    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
-    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
+    ptrs_src[0 * ne23 + i12 + i13 * ne12] = src0_bytes + i02 * nb02 + i03 * nb03;
+    ptrs_src[1 * ne23 + i12 + i13 * ne12] = src1_bytes + i12 * nb12 + i13 * nb13;
+    ptrs_dst[0 * ne23 + i12 + i13 * ne12] = dst_bytes + i12 * nbd2 + i13 * nbd3;
 }
 
-static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
-                                             const ggml_tensor *src0,
-                                             const ggml_tensor *src1,
-                                             ggml_tensor *dst) try {
+static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * src0,
+                                           const ggml_tensor * src1, ggml_tensor * dst) try {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
+    // TODO: see https://github.com/ggml-org/llama.cpp/pull/13155
+    // Batched mul_mat requires a rewrite to support both oneDNN and non-contiguous dst
+    GGML_ASSERT(ggml_is_contiguous(dst));
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    queue_ptr main_stream = ctx.stream();;
+    queue_ptr queue = ctx.stream();
 
-    void * src0_ddq = src0->data;
-    sycl::half *src0_as_f16 = (sycl::half *)src0_ddq;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf = (float *) dst->data;
+    dpct::has_capability_or_fail(queue->get_device(), { sycl::aspect::fp16 });
 
-    // convert src1 to fp16
+    const sycl::half * src0_f16 = static_cast<const sycl::half *>(src0->data);
+    float *            dst_ddf  = static_cast<float *>(dst->data);
+
+    const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+    const size_t       type_size_src1 = ggml_type_size(src1->type);
+    GGML_ASSERT(nb10 == type_size_src1);
+
+    // SRC1 strides
+    int64_t                          s11 = nb11 / type_size_src1;
+    int64_t                          s12 = nb12 / type_size_src1;
+    int64_t                          s13 = nb13 / type_size_src1;
     ggml_sycl_pool_alloc<sycl::half> src1_f16_alloc(ctx.pool());
+
+    // convert src1 to fp16
     if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
+        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
+        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
-        GGML_ASSERT(to_fp16_sycl != nullptr);
-        to_fp16_sycl(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+        to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue);
+
+        src1_f16 = src1_f16_alloc.get();
+        s11      = ne10;
+        s12      = ne11 * s11;
+        s13      = ne12 * s12;
     }
-    sycl::half *src1_f16 = src1->type == GGML_TYPE_F16 ? (sycl::half *)src1_ddf
-                                                       : src1_f16_alloc.get();
 
-    char * dst_t;
+    ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool());
 
-    dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float;
-    dpct::library_data_t cu_data_type = dpct::library_data_t::real_float;
+    dpct::library_data_t mkl_compute_type = dpct::library_data_t::real_float;
+    dpct::library_data_t mkl_data_type    = dpct::library_data_t::real_float;
 
     // dst strides
     size_t nbd2 = dst->nb[2];
     size_t nbd3 = dst->nb[3];
 
     const float alpha_f32 = 1.0f;
-    const float beta_f32 = 0.0f;
+    const float beta_f32  = 0.0f;
 
     const void * alpha = &alpha_f32;
     const void * beta  = &beta_f32;
 
-    dst_t = (char *) dst_ddf;
-
     GGML_ASSERT(ne12 % ne02 == 0);
     GGML_ASSERT(ne13 % ne03 == 0);
+    GGML_ASSERT(ne01 == static_cast<int64_t>(nb1/nb0));
+    GGML_ASSERT(ne10 == ne00);
 
     // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
-        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
-            *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-            (const char *) src0_as_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
-            (const char *) src1_f16, dpct::library_data_t::real_half, nb11 / nb10, nb12 / nb10, beta, (char *) dst_t,
-            cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type)));
-    } else {
-        const int ne23 = ne12*ne13;
-
-        ggml_sycl_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
-        ggml_sycl_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
-        ggml_sycl_pool_alloc<matrix_info_t<float>> matrix_info(ctx.host_pool(), 1);
-
-        sycl::range<3> block_dims(1, ne12, ne13);
-        /*
-        DPCT1049:47: The work-group size passed to the SYCL kernel may exceed
-        the limit. To get the device limit, query
-        info::device::max_work_group_size. Adjust the work-group size if needed.
-        */
-        {
-            dpct::has_capability_or_fail(main_stream->get_device(),
-                                         {sycl::aspect::fp16});
-
-            main_stream->submit([&](sycl::handler &cgh) {
-                const void **ptrs_src_get = ptrs_src.get();
-                void **ptrs_dst_get = ptrs_dst.get();
-                size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2;
-                size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2;
-                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims),
-                                 [=](sycl::nd_item<3> item_ct1) {
-                                     k_compute_batched_ptrs(
-                                         src0_as_f16, src1_f16,
-                                         dst_t, ptrs_src_get,
-                                         ptrs_dst_get, ne12, ne13, ne23,
-                                         nb02, nb03, nb12_scaled, nb13_scaled,
-                                         nbd2, nbd3, r2, r3, item_ct1);
-                                 });
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
+#if GGML_SYCL_DNNL
+    if (!g_ggml_sycl_disable_dnn) {
+        auto dnn_gemm = [&ctx, queue, ne11, ne01, ne10, nb00, nb01, nb02, s11, s12]
+            (const sycl::half* src1, const sycl::half* src0, float* dst, const dnnl_dim_t batches_a, const dnnl_dim_t batches_b) {
+
+            DnnlGemmWrapper::gemm(ctx, ne11,ne01, ne10,
+                            src1, DnnlGemmWrapper::to_dt<sycl::half>(), s11, 1, s12,
+                            src0, DnnlGemmWrapper::to_dt<sycl::half>(), 1, nb01/nb00, nb02/nb00,
+                            dst, DnnlGemmWrapper::to_dt<float>(), queue, batches_a, batches_b);
+        };
+
+        if (r2 == 1 && r3 == 1) {
+            if (ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+                dnn_gemm(src1_f16, src0_f16, dst_ddf, ne12*ne13, ne02 * ne03);
+            }
+            else {
+                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/sizeof(sycl::half)); // nb is in bytes
+                    const sycl::half* src1_f16_shifted = src1_f16 + ie03*s13;
+                    float* dst_shifted = dst_ddf + ((ie03*nb3)/sizeof(float));
+                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, ne12, ne02);
+                }
+            }
+        } else {
+            // iterate over batches from smaller set of matrices (matrix 0)
+            for (int64_t ie02 = 0; ie02 < ne02; ++ie02) {
+                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/sizeof(sycl::half));
+                    const sycl::half* src1_f16_shifted = src1_f16 + ie02*s12*r2 + ie03*s13*r3;
+                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/sizeof(float));
+                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, r2*r3, 1);
+                }
+            }
+        }
+    }
+    else
+#endif
+    {
+        if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+            // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
+                                                        oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
+                                                        src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_ddf,
+                                                        mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
+        } else {
+            const int ne23 = ne12 * ne13;
+
+            ggml_sycl_pool_alloc<const void *>         ptrs_src(ctx.pool(), 2 * ne23);
+            ggml_sycl_pool_alloc<void *>               ptrs_dst(ctx.pool(), 1 * ne23);
+            ggml_sycl_pool_alloc<matrix_info_t<float>> matrix_info(ctx.host_pool(), 1);
+
+            sycl::range<3> block_dims(1, ne12, ne13);
+            queue->submit([&](sycl::handler & cgh) {
+                const void ** ptrs_src_get = ptrs_src.get();
+                void **       ptrs_dst_get = ptrs_dst.get();
+                size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
+                size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
+                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
+                                           nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
+                });
             });
+
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
+                *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
+                (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, s11, beta,
+                (void **) (ptrs_dst.get() + 0 * ne23), mkl_data_type, ne0, ne23, mkl_compute_type, matrix_info.get())));
         }
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
-            *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-            (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
-            (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, nb11 / nb10, beta,
-            (void **) (ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type, matrix_info.get())));
     }
+} catch (const sycl::exception & exc) {
+    std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+    std::exit(1);
 }
-catch (sycl::exception const &exc) {
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-            << ", line:" << __LINE__ << std::endl;
-  std::exit(1);
-}
+
+enum class mul_mat_algo {
+    DMMV         = 0,
+    MMVQ         = 1,
+    MUL_MAT_SYCL = 2,
+};
 
 inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     // TODO: accuracy issues in MMQ
@@ -2834,6 +2896,36 @@ inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     return false;
 }
 
+inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return true;
+        case GGML_TYPE_Q4_K:
+            return !g_ggml_sycl_prioritize_dmmv;
+        default:
+            return false;
+    }
+}
+
+inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_K:
+            return true;
+        default:
+            return false;
+    }
+}
+
 static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -2853,16 +2945,16 @@ static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     }
 }
 
-static void reorder_qw(char *data_device, const int ncols, const int nrows,
-                size_t size, size_t offset, dpct::queue_ptr stream) {
-    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
+static void reorder_qw_q4_0(uint8_t * data_device, const int ncols, const int nrows, size_t size, size_t offset,
+                            dpct::queue_ptr stream) {
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
     SYCL_CHECK(
         CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
             .wait()));
     GGML_ASSERT((size % sizeof(block_q4_0) == 0));
     GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
     int offset_blks = offset / sizeof(block_q4_0);
-    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
+    auto qs_ptr      = data_device + offset_blks * QK4_0 / 2;
     auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
 
     stream->parallel_for(
@@ -2876,48 +2968,119 @@ static void reorder_qw(char *data_device, const int ncols, const int nrows,
                 *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
             }
             *(d_ptr + ib) = x[ib].d;
-        });
+        }).wait_and_throw();
+
+    sycl::free(tmp_buf, *stream);
+}
+
+static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
+    GGML_ASSERT(size % sizeof(block_q4_K) == 0);
+    GGML_ASSERT(offset % sizeof(block_q4_K) == 0);
+
+    const int nblocks = size / sizeof(block_q4_K);
+
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+
+    auto * qs_ptr     = data_device;
+    auto * scales_ptr = qs_ptr + QK_K / 2 * nblocks;
+    auto * dm_ptr     = (sycl::half2 *) (scales_ptr + K_SCALE_SIZE * nblocks);
+
+    stream->parallel_for(nblocks, [=](auto i) {
+        const block_q4_K * x  = (const block_q4_K *) tmp_buf;
+        const int          ib = i;
+
+        for (int j = 0; j < QK_K / 2; ++j) {
+            qs_ptr[ib * (QK_K / 2) + j] = x[ib].qs[j];
+        }
+
+        for (int j = 0; j < K_SCALE_SIZE; ++j) {
+            scales_ptr[ib * K_SCALE_SIZE + j] = x[ib].scales[j];
+        }
+
+        dm_ptr[ib] = x[ib].dm;
+    }).wait_and_throw();
 
     sycl::free(tmp_buf, *stream);
 }
 
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
-    char*data_device = (char*)src0->data;
+    uint8_t * data_device = (uint8_t *) src0->data;
     size_t ncols = src0->ne[0];
     size_t nrows = src0->ne[1];
     size_t size = ggml_nbytes(src0);
 
-    reorder_qw(data_device, ncols, nrows, size, 0, stream);
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            reorder_qw_q4_0(data_device, ncols, nrows, size, 0, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            reorder_qw_q4_k(data_device, size, 0, stream);
+            break;
+        default:
+            GGML_ABORT("reorder_qw() called with unsupported type");
+            break;
+    }
 }
 
-/*
-* This function could be called when the OP (mul_mat) function support reorder optimizition.
-*/
-static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1,
-    ggml_tensor * dst) {
-    if (!g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
-        ctx->opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
-        dst->op == GGML_OP_MUL_MAT &&    //limit to some supported cases of Q4_0, to do for more cases.
-        src0->type == GGML_TYPE_Q4_0 &&
-        src1->ne[2]==1 && src1->ne[3]==1) {
+static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_tensor * dst) {
+    return !g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
+            ctx.opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
+            dst->op == GGML_OP_MUL_MAT &&   //limit to some supported cases of Q4_0, to do for more cases.
+            dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
+}
 
-        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
-        if (!extra) return; //only happen in CI/UT permute case.
+static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * /* src1 */,
+                            ggml_tensor * dst, mul_mat_algo mm_algorithm) {
+    if (!should_reorder_tensor(*ctx, dst)) {
+        return;
+    }
 
-        if (extra->optimized_feature.reorder) return; //skip the tensor which is handled for reorder.
+    ggml_tensor_extra_gpu * extra = static_cast<ggml_tensor_extra_gpu *>(src0->extra);
+    if (!extra || extra->optimized_feature.reorder) {
+        return;  // Skip permutations and already reordered tensors
+    }
 
-        reorder_qw(src0, ctx->stream());
-        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
+    switch (mm_algorithm) {
+        case mul_mat_algo::DMMV:
+            if (!ggml_sycl_supports_reorder_dmmv(src0->type)) {
+                return;
+            }
+            break;
+        case mul_mat_algo::MMVQ:
+            if (!ggml_sycl_supports_reorder_mmvq(src0->type)) {
+                return;
+            }
+            break;
+        case mul_mat_algo::MUL_MAT_SYCL:
+            if (!ggml_sycl_supports_reorder_mul_mat_sycl(src0->type)) {
+                return;
+            }
+            break;
     }
+
+    reorder_qw(src0, ctx->stream());
+    extra->optimized_feature.reorder = true;  // Used to decode/dequan in next steps and avoid re-reordering
 }
 
-static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 
+static bool can_use_dequantize_mul_mat_vec(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    return ggml_sycl_supports_dmmv(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
+           src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
+}
+
+static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    return ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
+           src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+}
+
+static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
     int64_t min_compute_capability = INT_MAX;
 
     if (split) {
-        ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
+        ggml_backend_sycl_split_buffer_type_context * buft_ctx =
+            (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
         auto & tensor_split = buft_ctx->tensor_split;
         for (int id = 0; id < ggml_sycl_info().device_count; ++id) {
             // skip devices that are not going to do any work:
@@ -2930,17 +3093,13 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             }
         }
     } else {
-        min_compute_capability    = ggml_sycl_info().devices[ctx.device].cc;
+        min_compute_capability = ggml_sycl_info().devices[ctx.device].cc;
     }
 
     // check data types and tensor shapes for custom matrix multiplication kernels:
-    bool use_dequantize_mul_mat_vec = ggml_sycl_supports_dmmv(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
+    bool use_dequantize_mul_mat_vec = can_use_dequantize_mul_mat_vec(src0, src1, dst);
 
-    bool use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+    bool use_mul_mat_vec_q = can_use_mul_mat_vec_q(src0, src1, dst);
 
     bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
@@ -2952,9 +3111,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // SYCL_USE_XMX
 
+
     // mmvq path is faster in the CUDA backend.
-    if (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda)
+    if (!g_ggml_sycl_prioritize_dmmv && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda
+        // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization
+        // is enabled takes precedence over DMMV, the current if-else implementation
+        // requires disabling DMMV if both conditions are met
+        || (should_reorder_tensor(ctx, dst) && ggml_sycl_supports_reorder_mmvq(src0->type)))) {
         use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+    }
 
     if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         // TODO: Refactor and cleanup of mul mat dispatching.
@@ -2966,24 +3131,30 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             // The kernel from the if path is faster for that specific case, but does not support all mul mats.
             ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
         }
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
-        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
-        // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream());
+        constexpr bool convert_src1_to_q8_1 = false;
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::DMMV);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, convert_src1_to_q8_1);
     } else if (use_mul_mat_vec_q) {
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
+        constexpr bool convert_src1_to_q8_1 = true;
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MMVQ);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, convert_src1_to_q8_1);
     } else if (use_mul_mat_q) {
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
+        constexpr bool convert_src1_to_q8_1 = true;
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, convert_src1_to_q8_1);
     } else {
-        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
+        constexpr bool convert_src1_to_q8_1 = false;
+        // MUL_MAT_SYCL supports reorder
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MUL_MAT_SYCL);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1);
     }
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
 
 
@@ -3651,7 +3822,8 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
             return GGML_STATUS_SUCCESS;
         }
 
-        sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream()));
+        sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream()), {sycl_ex::property::graph::assume_buffer_outlives_graph{}});
+
         model_sycl_graph.begin_recording(*(sycl_ctx->stream()));
         ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
         model_sycl_graph.end_recording();