whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -30,22 +30,32 @@
 #include <regex>
 
 #include <sycl/sycl.hpp>
+#if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
+#    include <sycl/ext/oneapi/experimental/async_alloc/async_alloc.hpp>
+#endif
 #include <sycl/half_type.hpp>
 
+#include "ggml.h"
 #include "ggml-sycl.h"
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
+#include "ggml-sycl/add-id.hpp"
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
-#include "ggml-sycl/presets.hpp"
+#include "ggml-sycl/gated_delta_net.hpp"
 #include "ggml-sycl/gemm.hpp"
-#include "ggml-sycl/set_rows.hpp"
-#include "ggml-sycl/sycl_hw.hpp"
 #include "ggml-sycl/getrows.hpp"
+#include "ggml-sycl/norm.hpp"
+#include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/quantize.hpp"
-#include "ggml.h"
+#include "ggml-sycl/repeat_back.hpp"
+#include "ggml-sycl/set_rows.hpp"
+#include "ggml-sycl/set.hpp"
+#include "ggml-sycl/ssm_conv.hpp"
+#include "ggml-sycl/sycl_hw.hpp"
+
 
 static bool g_sycl_loaded = false;
 int g_ggml_sycl_debug = 0;
@@ -53,6 +63,9 @@ 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;
+int g_ggml_sycl_use_async_mem_op = 0;
+int g_ggml_sycl_enable_flash_attention = 1;
+
 
 static ggml_sycl_device_info ggml_sycl_init() {
     ggml_sycl_device_info info = {};
@@ -85,8 +98,14 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
         info.devices[i].cc =
             100 * prop.get_major_version() + 10 * prop.get_minor_version();
+        info.devices[i].nsm = prop.get_max_compute_units() / 16; //16: Number of Xe Cores
         info.devices[i].opt_feature.reorder = device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
+        info.devices[i].smpbo = prop.get_local_mem_size();
+        info.devices[i].warp_size = WARP_SIZE;
+
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
+        info.devices[i].max_wg_per_cu = info.max_work_group_sizes[i] / prop.get_max_compute_units();
+
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -199,7 +218,37 @@ static void ggml_check_sycl() try {
         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);
+
+#ifdef SYCL_FLASH_ATTN
+        g_ggml_sycl_enable_flash_attention = get_sycl_env("GGML_SYCL_ENABLE_FLASH_ATTN", 1);
+#else
+        g_ggml_sycl_enable_flash_attention = 0;
+#endif
+
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
+
+        GGML_LOG_INFO("Build with Macros:\n");
+#if defined(GGML_SYCL_FORCE_MMQ)
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: no\n");
+#endif
+#if defined(GGML_SYCL_F16)
+        GGML_LOG_INFO("  GGML_SYCL_F16: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_F16: no\n");
+#endif
+#if defined(GGML_SYCL_GRAPH)
+        GGML_LOG_INFO("  GGML_SYCL_GRAPH: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_GRAPH: no\n");
+#endif
+#if defined(GGML_SYCL_DNNL)
+        GGML_LOG_INFO("  GGML_SYCL_DNNL: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_DNNL: no\n");
+#endif
+
         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);
@@ -214,16 +263,12 @@ static void ggml_check_sycl() try {
         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");
-#else
-        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: no\n");
-#endif
-#if defined(GGML_SYCL_F16)
-        GGML_LOG_INFO("  GGML_SYCL_F16: yes\n");
+
+#ifdef SYCL_FLASH_ATTN
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
 #else
-        GGML_LOG_INFO("  GGML_SYCL_F16: no\n");
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d disabled by compile flag\n",
+            g_ggml_sycl_enable_flash_attention);
 #endif
 
 /* NOT REMOVE, keep it for next optimize for XMX.
@@ -233,7 +278,20 @@ static void ggml_check_sycl() try {
         fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
 #endif
 */
-
+        // Currently, we only use async malloc / free when graphs are enabled as it is required for the calls to be
+        // properly recorded. As this SYCL extension matures it may be beneficial to enable as the default path and in
+        // other places.
+#if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
+        g_ggml_sycl_use_async_mem_op = !g_ggml_sycl_disable_graph;
+        if (g_ggml_sycl_use_async_mem_op) {
+            for (unsigned int i = 0; i < dpct::dev_mgr::instance().device_count(); ++i) {
+                if (!dpct::dev_mgr::instance().get_device(i).has(sycl::aspect::ext_oneapi_async_memory_alloc)) {
+                    g_ggml_sycl_use_async_mem_op = 0;
+                    break;
+                }
+            }
+        }
+#endif
         if (CHECK_TRY_ERROR(g_all_sycl_device_count =
                             dpct::dev_mgr::instance().device_count()) != 0) {
             initialized = true;
@@ -1132,13 +1190,28 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+inline void * aligned_malloc_host(size_t alignment, size_t size) {
+#ifdef _WIN32
+    return _aligned_malloc(size, alignment);
+#else
+    return aligned_alloc(alignment, size);
+#endif
+}
+
+inline void free_aligned_mem_host(void * memblock) {
+#ifdef _WIN32
+    _aligned_free(memblock);
+#else
+    free(memblock);
+#endif
+}
+
 static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_sycl_host_free(buffer->context);
+    free_aligned_mem_host((void *)buffer->context);
 }
 
 static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr = ggml_sycl_host_malloc(size);
-
+    void * ptr = aligned_malloc_host(TENSOR_ALIGNMENT, size);
     if (ptr == nullptr) {
         // fallback to cpu buffer
         return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
@@ -1511,60 +1584,70 @@ static inline void ggml_sycl_swap(T & a, T & b) {
 template <ggml_sort_order order>
 __dpct_inline__ static void
 k_argsort_f32_i32(const float *x, int *dst, const int ncols, int ncols_pad,
-                  const sycl::nd_item<3> &item_ct1, uint8_t *dpct_local) {
+                  const int tasks_per_thread, const sycl::nd_item<3> &item_ct1,
+                  uint8_t *dpct_local) {
     // bitonic sort
-    int col = item_ct1.get_local_id(2);
+    int col_index =  item_ct1.get_local_id(2);
     int row = item_ct1.get_group(1);
 
-    if (col >= ncols_pad) {
-        return;
+    for (int i = 0; i < tasks_per_thread; i++) {
+        int col = col_index * tasks_per_thread + i;
+        if (col >= ncols_pad) {
+            return;
+        }
     }
 
     const float * x_row = x + row * ncols;
     auto dst_row = (int *)dpct_local;
 
     // initialize indices
-    dst_row[col] = col;
+    for (int i=0;i<tasks_per_thread;i++){
+        int col = col_index*tasks_per_thread+i;
+        dst_row[col] = col;
+    }
 
     item_ct1.barrier(sycl::access::fence_space::local_space);
 
     for (int k = 2; k <= ncols_pad; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= ncols ||
-                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (dst_row[ixj] >= ncols ||
-                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+            for (int i = 0; i < tasks_per_thread; i++) {
+                int col = col_index * tasks_per_thread + i;
+                int ixj = col ^ j;
+                if (ixj > col) {
+                    if ((col & k) == 0) {
+                        if (dst_row[col] >= ncols ||
+                            (dst_row[ixj] < ncols &&
+                             (order == GGML_SORT_ORDER_ASC
+                                  ? x_row[dst_row[col]] > x_row[dst_row[ixj]]
+                                  : x_row[dst_row[col]] <
+                                        x_row[dst_row[ixj]]))) {
+                            ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                        }
+                    } else {
+                        if (dst_row[ixj] >= ncols ||
+                            (dst_row[col] < ncols &&
+                             (order == GGML_SORT_ORDER_ASC
+                                  ? x_row[dst_row[col]] < x_row[dst_row[ixj]]
+                                  : x_row[dst_row[col]] >
+                                        x_row[dst_row[ixj]]))) {
+                            ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                        }
                     }
                 }
+                item_ct1.barrier(sycl::access::fence_space::local_space);
             }
-            /*
-            DPCT1118:1: SYCL group functions and algorithms must be encountered
-            in converged control flow. You may need to adjust the code.
-            */
-            item_ct1.barrier(sycl::access::fence_space::local_space);
         }
     }
 
     // copy the result to dst without the padding
-    if (col < ncols) {
-        dst[row * ncols + col] = dst_row[col];
+    for (int i = 0; i < tasks_per_thread; i++) {
+        int col = col_index * tasks_per_thread + i;
+        if (col < ncols) {
+            dst[row * ncols + col] = dst_row[col];
+        }
     }
 }
 
-
 static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past,
                               const sycl::nd_item<3> &item_ct1) {
     const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
@@ -1737,13 +1820,23 @@ static int next_power_of_2(int x) {
 
 static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                                  const int nrows, ggml_sort_order order,
-                                 queue_ptr stream) {
+                                 queue_ptr stream, int device) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
 
-    const sycl::range<3> block_dims(1, 1, ncols_pad);
+    int nth = 1;
+    int max_block_size = ggml_sycl_info().max_work_group_sizes[device];
+    while (nth < ncols_pad && nth < max_block_size)
+        nth *= 2;
+    if (nth > max_block_size)
+        nth = max_block_size;
+
+    const int tasks_per_thread = ncols_pad / nth;
+
+    const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
+    GGML_ASSERT(shared_mem<=ggml_sycl_info().devices[device].smpbo);
 
     if (order == GGML_SORT_ORDER_ASC) {
         stream->submit([&](sycl::handler &cgh) {
@@ -1754,8 +1847,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
-                        x, dst, ncols, ncols_pad, item_ct1,
-                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                        x, dst, ncols, ncols_pad, tasks_per_thread, item_ct1,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
                             .get());
                 });
         });
@@ -1768,8 +1862,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
-                        x, dst, ncols, ncols_pad, item_ct1,
-                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                        x, dst, ncols, ncols_pad, tasks_per_thread, item_ct1,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
                             .get());
                 });
         });
@@ -1778,6 +1873,110 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
     }
 }
 
+static void top_k_f32_sycl(
+    const float * src,
+    int32_t * dst_indices,
+    const int64_t ncols,
+    const int64_t nrows,
+    const int k,
+    dpct::queue_ptr main_stream
+) {
+    const int block_size = 128;
+
+    const sycl::range<1> block_dims(block_size);
+    const sycl::range<1> grid_dims(nrows);
+
+    main_stream->submit([&](sycl::handler &cgh) {
+        sycl::local_accessor<float, 1> shared_vals(sycl::range<1>(block_size * k), cgh);
+        sycl::local_accessor<int, 1> shared_idx(sycl::range<1>(block_size * k), cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<1>(grid_dims * block_dims, block_dims),
+            [=](sycl::nd_item<1> item_ct1) {
+                const int row = item_ct1.get_group(0);
+                const int tid = item_ct1.get_local_id(0);
+
+                if (row >= nrows) return;
+
+                const float * src_row = src + row * ncols;
+                int32_t * dst_idx_row = dst_indices + row * k;
+
+                float local_vals[32];
+                int local_idx[32];
+
+                for (int i = 0; i < k; i++) {
+                    local_vals[i] = -FLT_MAX;
+                    local_idx[i] = -1;
+                }
+
+                for (int col = tid; col < ncols; col += block_size) {
+                    float val = src_row[col];
+
+                    if (val > local_vals[k-1]) {
+                        int pos = k - 1;
+                        while (pos > 0 && val > local_vals[pos - 1]) {
+                            pos--;
+                        }
+
+                        for (int i = k - 1; i > pos; i--) {
+                            local_vals[i] = local_vals[i - 1];
+                            local_idx[i] = local_idx[i - 1];
+                        }
+                        local_vals[pos] = val;
+                        local_idx[pos] = col;
+                    }
+                }
+
+                for (int i = 0; i < k; i++) {
+                    shared_vals[tid * k + i] = local_vals[i];
+                    shared_idx[tid * k + i] = local_idx[i];
+                }
+                item_ct1.barrier(sycl::access::fence_space::local_space);
+
+                if (tid == 0) {
+                    float final_vals[32];
+                    int final_idx[32];
+
+                    for (int i = 0; i < k; i++) {
+                        final_vals[i] = -FLT_MAX;
+                        final_idx[i] = -1;
+                    }
+
+                    for (int t = 0; t < block_size; t++) {
+                        for (int i = 0; i < k; i++) {
+                            float val = shared_vals[t * k + i];
+                            int idx = shared_idx[t * k + i];
+
+                            if (val > final_vals[k-1]) {
+                                int pos = k - 1;
+                                while (pos > 0 && val > final_vals[pos - 1]) {
+                                    pos--;
+                                }
+
+                                for (int j = k - 1; j > pos; j--) {
+                                    final_vals[j] = final_vals[j - 1];
+                                    final_idx[j] = final_idx[j - 1];
+                                }
+                                final_vals[pos] = val;
+                                final_idx[pos] = idx;
+                            }
+                        }
+                    }
+
+                    for (int i = 0; i < k; i++) {
+                        dst_idx_row[i] = final_idx[i];
+                    }
+
+                    if (k > 1) {
+                        int32_t temp = dst_idx_row[0];
+                        dst_idx_row[0] = dst_idx_row[1];
+                        dst_idx_row[1] = temp;
+                    }
+                }
+            });
+    });
+}
+
 static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols,
                                const int nrows, queue_ptr stream) {
     const sycl::range<3> block_dims(1, 1, SYCL_ARGMAX_BLOCK_SIZE);
@@ -2001,8 +2200,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
             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,
+                *stream, oneapi::mkl::transpose::trans,
+                oneapi::mkl::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,
@@ -2045,8 +2244,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
         {
             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,
+            SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+                *stream, oneapi::mkl::transpose::trans, oneapi::mkl::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)));
         }
@@ -2127,6 +2326,30 @@ inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor *
     sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
 }
 
+inline void ggml_sycl_op_mean(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
+    float *       dst_dd  = static_cast<float *>(dst->data);
+
+    const int64_t ncols = dst->src[0]->ne[0];
+    const int64_t nrows = ggml_nrows(dst->src[0]);
+
+    sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
+
+    main_stream->parallel_for(
+        sycl::range<1>(nrows),
+        [=](sycl::id<1> row) {
+            dst_dd[row] /= ncols;
+        }
+    );
+}
+
+
 inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_I32);
@@ -2141,7 +2364,32 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor *
 
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
-    argsort_f32_i32_sycl(src0_dd, (int *) dst_dd, ncols, nrows, order, main_stream);
+    argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order,
+                         main_stream, ctx.device);
+}
+
+static void ggml_sycl_op_top_k(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float * src0_dd = static_cast<const float *>(src0->data);
+    int32_t * dst_dd = static_cast<int32_t *>(dst->data);
+
+    const int k = dst->ne[0];
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    GGML_ASSERT(k > 0 && k <= 32);
+    GGML_ASSERT(k <= ncols);
+
+    top_k_f32_sycl(src0_dd, dst_dd, ncols, nrows, k, main_stream);
 }
 
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -2176,6 +2424,65 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_ten
     diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
 }
 
+static void tri_f32_sycl(
+    const float * src,
+    float * dst,
+    const int64_t ne0,
+    const int64_t ne1,
+    const int64_t ne2,
+    const int64_t ne3,
+    const ggml_tri_type ttype,
+    dpct::queue_ptr main_stream
+) {
+    const size_t total = (size_t) ne0 * (size_t) ne1 * (size_t) ne2 * (size_t) ne3;
+
+    main_stream->parallel_for(sycl::range<1>(total), [=](sycl::id<1> tid) {
+        const int64_t idx = (int64_t) tid[0];
+
+        const int64_t i0 = idx % ne0;
+        const int64_t t1 = idx / ne0;
+        const int64_t i1 = t1 % ne1;
+
+        bool keep = false;
+        switch (ttype) {
+            case GGML_TRI_TYPE_LOWER:      keep = (i0 <  i1); break;
+            case GGML_TRI_TYPE_LOWER_DIAG: keep = (i0 <= i1); break;
+            case GGML_TRI_TYPE_UPPER:      keep = (i0 >  i1); break;
+            case GGML_TRI_TYPE_UPPER_DIAG: keep = (i0 >= i1); break;
+            default: keep = false; break;
+        }
+
+        dst[idx] = keep ? src[idx] : 0.0f;
+    });
+}
+
+static void ggml_sycl_op_tri(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    GGML_ASSERT(src0);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float * src0_dd = static_cast<const float *>(src0->data);
+    float *       dst_dd  = static_cast<float *>(dst->data);
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    const int64_t ne0 = src0->ne[0];
+    const int64_t ne1 = src0->ne[1];
+    const int64_t ne2 = src0->ne[2];
+    const int64_t ne3 = src0->ne[3];
+
+    tri_f32_sycl(src0_dd, dst_dd, ne0, ne1, ne2, ne3, ttype, main_stream);
+}
+
+
 inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -2548,6 +2855,10 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
+static void ggml_sycl_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_repeat_back(ctx, dst);
+}
 
 static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
@@ -2564,6 +2875,11 @@ static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     ggml_sycl_op_rms_norm(ctx, dst);
 }
 
+static void ggml_sycl_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_rms_norm_back(ctx, dst);
+}
+
 static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_l2_norm(ctx, dst);
@@ -2729,7 +3045,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 
         }
 #if GGML_SYCL_DNNL
-        // oneDNN handles strided data and does not need overhead of get_to_fp16_nc_sycl
+        // oneDNN handles strided data and does not need overhead of ggml_get_to_fp16_nc_sycl
         const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
         src1_f16_alloc.alloc(ne_src1);
         const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
@@ -2738,7 +3054,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 # else
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
-        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
+        const to_fp16_nc_sycl_t to_fp16_nc_sycl = ggml_get_to_fp16_nc_sycl(src1->type);
         GGML_ASSERT(to_fp16_nc_sycl != nullptr);
         to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue);
 #endif
@@ -2882,8 +3198,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             const int64_t smb = ne12 == 1 ? s13       : s12;
 
             // 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,
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::mkl::transpose::trans,
+                                                        oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
                                                         src0_f16, dpct::library_data_t::real_half, nb01 / nb00, sma,
                                                         src1_f16, dpct::library_data_t::real_half, s11, smb, beta, dst_ddf,
                                                         mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
@@ -2907,7 +3223,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             });
 
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
-                *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                *queue, oneapi::mkl::transpose::trans, oneapi::mkl::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())));
@@ -2981,19 +3297,51 @@ static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     }
 }
 
+// Helper functions to unify device memory allocation for both async and sync paths
+static inline void * sycl_ext_malloc_device(dpct::queue_ptr stream, size_t size) {
+    bool use_async = g_ggml_sycl_use_async_mem_op;
+#if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
+    if (use_async) {
+        return syclex::async_malloc(*stream, sycl::usm::alloc::device, size);
+    }
+#else
+    // If async allocation extension is not available, use_async should always be false.
+    GGML_ASSERT(!use_async);
+#endif
+    return sycl::malloc(size, *stream, sycl::usm::alloc::device);
+}
+
+static inline void sycl_ext_free(dpct::queue_ptr stream, void * ptr) {
+    bool use_async = g_ggml_sycl_use_async_mem_op;
+#if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
+    if (use_async) {
+        syclex::async_free(*stream, ptr);
+        return;
+    }
+#else
+    // If async allocation extension is not available, use_async should always be false.
+    GGML_ASSERT(!use_async);
+#endif
+    sycl::free(ptr, *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()));
+    uint8_t * tmp_buf = static_cast<uint8_t *>(sycl_ext_malloc_device(stream, size));
+
+    sycl::event copy_event;
+    SYCL_CHECK(CHECK_TRY_ERROR(copy_event = stream->memcpy(tmp_buf, data_device, size)));
+    if (!g_ggml_sycl_use_async_mem_op) {
+        copy_event.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      = data_device + offset_blks * QK4_0 / 2;
     auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
 
-    stream->parallel_for(
+    auto reorder_event = stream->parallel_for(
         size / sizeof(block_q4_0),
             [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             const block_q4_0* x = (const block_q4_0*)tmp_buf;
@@ -3004,9 +3352,11 @@ static void reorder_qw_q4_0(uint8_t * data_device, const int ncols, const int nr
                 *(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);
+        });
+    if (!g_ggml_sycl_use_async_mem_op) {
+        reorder_event.wait_and_throw();
+    }
+    sycl_ext_free(stream, tmp_buf);
 }
 
 static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
@@ -3015,14 +3365,19 @@ static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, d
 
     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()));
+    uint8_t * tmp_buf = static_cast<uint8_t *>(sycl_ext_malloc_device(stream, size));
+
+    sycl::event copy_event;
+    SYCL_CHECK(CHECK_TRY_ERROR(copy_event = stream->memcpy(tmp_buf, data_device, size)));
+    if (!g_ggml_sycl_use_async_mem_op) {
+        copy_event.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) {
+    auto reorder_event = stream->parallel_for(nblocks, [=](auto i) {
         const block_q4_K * x  = (const block_q4_K *) tmp_buf;
         const int          ib = i;
 
@@ -3035,9 +3390,11 @@ static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, d
         }
 
         dm_ptr[ib] = x[ib].dm;
-    }).wait_and_throw();
-
-    sycl::free(tmp_buf, *stream);
+    });
+    if (!g_ggml_sycl_use_async_mem_op) {
+        reorder_event.wait_and_throw();
+    }
+    sycl_ext_free(stream, tmp_buf);
 }
 
 static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
@@ -3046,42 +3403,46 @@ static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, d
 
     const int nblocks = size / sizeof(block_q6_K);
 
-    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
-    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+    uint8_t * tmp_buf = static_cast<uint8_t *>(sycl_ext_malloc_device(stream, size));
+
+    sycl::event copy_event;
+    SYCL_CHECK(CHECK_TRY_ERROR(copy_event = stream->memcpy(tmp_buf, data_device, size)));
+    if (!g_ggml_sycl_use_async_mem_op) {
+        copy_event.wait();
+    }
 
     auto *       ql_ptr     = data_device;
     auto *       qh_ptr     = ql_ptr + (QK_K / 2) * nblocks;
     auto *       scales_ptr = qh_ptr + (QK_K / 4) * nblocks;
     sycl::half * dm_ptr     = (sycl::half *) (scales_ptr + (QK_K / 16) * nblocks);
 
-    stream
-        ->parallel_for(nblocks,
-                       [=](auto i) {
-                           const block_q6_K * x  = (const block_q6_K *) tmp_buf;
-                           const int          ib = i;
-
-                           const uint8_t * ql              = x[ib].ql;
-                           const uint8_t * qh              = x[ib].qh;
-                           uint8_t *       base_ql_ptr     = ql_ptr + (QK_K / 2) * ib;
-                           uint8_t *       base_qh_ptr     = qh_ptr + (QK_K / 4) * ib;
-                           uint8_t *       base_scales_ptr = scales_ptr + (QK_K / 16) * ib;
+    auto reorder_event = stream->parallel_for(nblocks, [=](auto i) {
+        const block_q6_K * x  = (const block_q6_K *) tmp_buf;
+        const int          ib = i;
 
-                           for (int j = 0; j < QK_K / 2; ++j) {
-                               base_ql_ptr[j] = ql[j];
-                           }
-                           for (int j = 0; j < QK_K / 4; ++j) {
-                               base_qh_ptr[j] = qh[j];
-                           }
+        const uint8_t * ql              = x[ib].ql;
+        const uint8_t * qh              = x[ib].qh;
+        uint8_t *       base_ql_ptr     = ql_ptr + (QK_K / 2) * ib;
+        uint8_t *       base_qh_ptr     = qh_ptr + (QK_K / 4) * ib;
+        uint8_t *       base_scales_ptr = scales_ptr + (QK_K / 16) * ib;
 
-                           for (int j = 0; j < QK_K / 16; ++j) {
-                               base_scales_ptr[j] = x[ib].scales[j];
-                           }
+        for (int j = 0; j < QK_K / 2; ++j) {
+            base_ql_ptr[j] = ql[j];
+        }
+        for (int j = 0; j < QK_K / 4; ++j) {
+            base_qh_ptr[j] = qh[j];
+        }
 
-                           dm_ptr[ib] = x[ib].d;
-                       })
-        .wait_and_throw();
+        for (int j = 0; j < QK_K / 16; ++j) {
+            base_scales_ptr[j] = x[ib].scales[j];
+        }
 
-    sycl::free(tmp_buf, *stream);
+        dm_ptr[ib] = x[ib].d;
+    });
+    if (!g_ggml_sycl_use_async_mem_op) {
+        reorder_event.wait_and_throw();
+    }
+    sycl_ext_free(stream, tmp_buf);
 }
 
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
@@ -3188,20 +3549,19 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
+
     // mmvq and mmq need the __dp4a instruction which is available for gen12+
-    // Workaround in https://github.com/ggerganov/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e
+    // Workaround in https://github.com/ggml-org/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e
     use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS);
 #ifdef SYCL_USE_XMX
     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 (!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)))) {
+    // 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
+    if (!g_ggml_sycl_prioritize_dmmv && ((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;
     }
 
@@ -3510,6 +3870,12 @@ static void ggml_sycl_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     ggml_sycl_op_sum_rows(ctx, dst);
 }
 
+static void ggml_sycl_mean(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
+    ggml_sycl_op_mean(ctx, dst);
+}
+
 static void ggml_sycl_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
@@ -3561,9 +3927,15 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_REPEAT:
             ggml_sycl_repeat(ctx, dst);
             break;
+        case GGML_OP_REPEAT_BACK:
+            ggml_sycl_repeat_back(ctx, dst);
+            break;
         case GGML_OP_GET_ROWS:
             ggml_sycl_get_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_sycl_op_set(ctx, dst);
+            break;
         case GGML_OP_SET_ROWS:
             ggml_sycl_op_set_rows(ctx, dst);
             break;
@@ -3574,6 +3946,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_ADD1: // TODO: more efficient implementation
             ggml_sycl_add(ctx, dst);
             break;
+        case GGML_OP_ADD_ID:
+            ggml_sycl_add_id(ctx, dst);
+            break;
         case GGML_OP_SUB:
             ggml_sycl_sub(ctx, dst);
             break;
@@ -3630,6 +4005,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_sycl_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_sycl_softplus(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_SGN:
                     ggml_sycl_sgn(ctx, dst);
                     break;
@@ -3639,6 +4017,18 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_UNARY_OP_ELU:
                     ggml_sycl_elu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_sycl_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_sycl_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_sycl_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_sycl_trunc(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3654,6 +4044,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_sycl_swiglu(ctx, dst);
                     break;
+                case GGML_GLU_OP_SWIGLU_OAI:
+                    ggml_sycl_swiglu_oai(ctx, dst);
+                    break;
                 case GGML_GLU_OP_GEGLU_ERF:
                     ggml_sycl_geglu_erf(ctx, dst);
                     break;
@@ -3673,6 +4066,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_CONCAT:
             ggml_sycl_op_concat(ctx, dst);
             break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_sycl_op_pad_reflect_1d(ctx,dst);
+            break;
         case GGML_OP_UPSCALE:
             ggml_sycl_upscale(ctx, dst);
             break;
@@ -3682,6 +4078,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_LEAKY_RELU:
             ggml_sycl_leaky_relu(ctx, dst);
             break;
+        case GGML_OP_RMS_NORM_BACK:
+            ggml_sycl_rms_norm_back(ctx, dst);
+            break;
         case GGML_OP_RMS_NORM:
             ggml_sycl_rms_norm(ctx, dst);
             break;
@@ -3735,15 +4134,24 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_TRANSPOSE:
             GGML_SYCL_DEBUG("%s: Tensor NO-OP\n", __func__);
             break;
+        case GGML_OP_TRI:
+            ggml_sycl_op_tri(ctx, dst);
+            break;
         case GGML_OP_DIAG_MASK_INF:
             ggml_sycl_diag_mask_inf(ctx, dst);
             break;
         case GGML_OP_SOFT_MAX:
             ggml_sycl_op_soft_max(ctx, dst);
             break;
+        case GGML_OP_SOFT_MAX_BACK:
+            ggml_sycl_op_soft_max_back(ctx, dst);
+            break;
         case GGML_OP_ROPE:
             ggml_sycl_rope(ctx, dst);
             break;
+        case GGML_OP_ROPE_BACK:
+            ggml_sycl_rope_back(ctx, dst);
+            break;
         case GGML_OP_IM2COL:
             ggml_sycl_im2col(ctx, dst);
             break;
@@ -3756,9 +4164,15 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_SUM_ROWS:
             ggml_sycl_sum_rows(ctx, dst);
             break;
+        case GGML_OP_MEAN:
+            ggml_sycl_mean(ctx, dst);
+            break;
         case GGML_OP_ARGSORT:
             ggml_sycl_argsort(ctx, dst);
             break;
+        case GGML_OP_TOP_K:
+            ggml_sycl_op_top_k(ctx, dst);
+            break;
         case GGML_OP_TIMESTEP_EMBEDDING:
             ggml_sycl_op_timestep_embedding(ctx, dst);
             break;
@@ -3771,6 +4185,21 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_GATED_LINEAR_ATTN:
             ggml_sycl_op_gated_linear_attn(ctx, dst);
             break;
+        case GGML_OP_GATED_DELTA_NET:
+            ggml_sycl_gated_delta_net(ctx, dst);
+            break;
+        case GGML_OP_SSM_CONV:
+            ggml_sycl_ssm_conv(ctx, dst);
+            break;
+        case GGML_OP_ROLL:
+            ggml_sycl_roll(ctx, dst);
+            break;
+        case GGML_OP_ARANGE:
+            ggml_sycl_arange(ctx, dst);
+            break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            ggml_sycl_flash_attn_ext(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3778,6 +4207,7 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
     return true;
 } catch (sycl::exception & e) {
     std::cerr << e.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+    std::cerr << "Error OP "<<ggml_op_name(dst->op)<< std::endl;
     std::exit(1);
 }
 
@@ -3800,16 +4230,6 @@ void ggml_backend_sycl_get_device_memory(int device, size_t *free,
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_memory\n");
     ggml_sycl_set_device(device);
 
-    /*
-    DPCT1009:218: SYCL uses exceptions to report errors and does not use the
-    error codes. The original code was commented out and a warning string was
-    inserted. You need to rewrite this code.
-    */
-    /*
-    DPCT1106:217: 'cudaMemGetInfo' was migrated with the Intel extensions for
-    device information which may not be supported by all compilers or runtimes.
-    You may need to adjust the code.
-    */
     SYCL_CHECK(CHECK_TRY_ERROR(
         dpct::dev_mgr::instance().get_device(device).get_memory_info(*free, *total)));
 }
@@ -3931,6 +4351,9 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
         if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
 #ifndef NDEBUG
         assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
         for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -3972,6 +4395,18 @@ static bool check_graph_compatibility(ggml_cgraph * cgraph) {
                 GGML_LOG_INFO("%s: disabling SYCL graphs due to unsupported node type %s\n", __func__,
                               ggml_op_name(node_op));
                 return false;
+            case GGML_OP_MUL_MAT:
+                // We cannot use graphs with ggml_sycl_mul_mat() when SYCL async memory allocation extensions are not available,
+                // as SYCL malloc / free and host wait calls are not supported when recording to a graph which are all present
+                // in reordering.
+                if (!g_ggml_sycl_use_async_mem_op) {
+                    GGML_LOG_INFO(
+                        "%s: disabling SYCL graphs due to unsupported node type when using a compiler without the "
+                        "oneAPI async memory allocation extension "
+                        "%s\n",
+                        __func__, ggml_op_name(node_op));
+                    return false;
+                }
         }
     }
     return true;
@@ -4096,6 +4531,7 @@ struct ggml_backend_sycl_device_context {
     int device;
     std::string name;
     std::string description;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_sycl_device_get_name(ggml_backend_dev_t dev) {
@@ -4166,6 +4602,9 @@ static ggml_backend_buffer_t ggml_backend_sycl_device_buffer_from_host_ptr(ggml_
 }
 
 static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_sycl_device_context *sycl_ctx =
+        (ggml_backend_sycl_device_context *)dev->context;
+    int device = sycl_ctx->device;
     switch (op->op) {
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
@@ -4178,21 +4617,27 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             }
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_ABS:
                 case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_SIGMOID:
-                case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_GELU_ERF:
-                case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
-                case GGML_UNARY_OP_SGN:
-                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_CEIL:
+                    return true;
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
 #if defined (GGML_SYCL_F16)
                     return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type);
 #else
@@ -4206,6 +4651,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]);
@@ -4233,15 +4679,18 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     }
                 }
                 ggml_type src0_type = op->src[0]->type;
-                if (src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_MXFP4) {
-                    // TODO: support MXFP4
+                if (src0_type == GGML_TYPE_BF16 ) {
+                    // TODO: support GGML_TYPE_BF16
                     // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
+
                 // TODO: The configuration below needs more work to be supported with oneDNN
-                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) && a->ne[2] > 1 && a->ne[3] > 1) {
-                    return false;
+                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) &&
+                    a->ne[2] > 1 && a->ne[3] > 1 && src0_type == GGML_TYPE_F16) {
+                  return false;
                 }
+
                 // TODO: This specific configuration can fail with oneDNN and needs more debugging
                 if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
                     a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
@@ -4266,6 +4715,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                         return false;
                 }
             }
+         case GGML_OP_SET:
+               return (op->type == GGML_TYPE_F32) &&
+                      (op->src[0] && op->src[1]) &&
+                      (op->src[0]->type == GGML_TYPE_F32) &&
+                      (op->src[1]->type == GGML_TYPE_F32);
+
         case GGML_OP_SET_ROWS:
             {
                 return ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
@@ -4343,11 +4798,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 }
                 return false;
             }
-        case GGML_OP_CONCAT:
+        case GGML_OP_REPEAT_BACK:
             {
                 ggml_type src0_type = op->src[0]->type;
-                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+                return src0_type == GGML_TYPE_F32;
             }
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_ARGMAX:
         case GGML_OP_NONE:
@@ -4355,15 +4811,17 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
-            return true;
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
+        case GGML_OP_ADD_ID:
         case GGML_OP_SUB:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_REPEAT:
             return true;
+        case GGML_OP_PAD_REFLECT_1D:
+            return ggml_is_contiguous(op->src[0]) && op-> type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
@@ -4376,50 +4834,81 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type);
 #endif
         case GGML_OP_NORM:
-            return true;
         case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
-            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_RMS_NORM:
-            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
+            return true;
+        case GGML_OP_RMS_NORM_BACK:
+            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
-        case GGML_OP_SOFT_MAX:
-            // TODO: support batching
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
-            // TODO: support attention sinks [TAG_ATTN_SINKS]
-            if (op->src[2]) {
-                return false;
+        case GGML_OP_TRI:
+            {
+                const ggml_tensor * src0 = op->src[0];
+                return src0 &&
+                       op->type == GGML_TYPE_F32 &&
+                       ggml_is_contiguous(src0);
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_DIAG_MASK_INF:
+            return true;
+        case GGML_OP_SOFT_MAX:
+            return true;
+        case GGML_OP_SOFT_MAX_BACK: {
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
+            return max_bias == 0.0f;
+        }
         case GGML_OP_ROPE:
+        case GGML_OP_ROPE_BACK:
         case GGML_OP_IM2COL:
             return true;
         case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
+            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
-        case GGML_OP_ARGSORT:
+        case GGML_OP_MEAN:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_ARGSORT:
+            return op->src[0]->ne[0] * sizeof(int) <=
+                   ggml_sycl_info().devices[device].smpbo;
+        case GGML_OP_TOP_K: {
+            const ggml_tensor * src0 = op->src[0];
+            const int k = op->ne[0];
+            return src0 &&
+                op->type == GGML_TYPE_I32 &&
+                src0->type == GGML_TYPE_F32 &&
+                ggml_is_contiguous(src0) &&
+                k > 0 && k <= 32;
+        }
         case GGML_OP_POOL_2D:
-        case GGML_OP_ACC:
             return true;
+        case GGML_OP_ACC:
+            return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
         case GGML_OP_PAD:
-            return (ggml_get_op_params_i32(op, 0) == 0) && (ggml_get_op_params_i32(op, 2) == 0) &&
-                   (ggml_get_op_params_i32(op, 4) == 0) && (ggml_get_op_params_i32(op, 6) == 0);
+            // TODO: add circular padding support for syscl, see https://github.com/ggml-org/llama.cpp/pull/16985
+            if (ggml_get_op_params_i32(op, 8) != 0) {
+                return false;
+            }
+            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
         case GGML_OP_GATED_LINEAR_ATTN:
+        case GGML_OP_GATED_DELTA_NET:
             return true;
+        case GGML_OP_SSM_CONV:
+            return op->type == GGML_TYPE_F32 &&
+                   op->src[0]->type == GGML_TYPE_F32 &&
+                   op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_ROLL:
+            return op->type == GGML_TYPE_F32;
+        case GGML_OP_ARANGE:
+            return op->type == GGML_TYPE_F32;
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ggml_sycl_flash_attn_ext_supported(device, op);
         default:
             return false;
     }
@@ -4451,9 +4940,8 @@ static int64_t get_op_batch_size(const ggml_tensor * op) {
 }
 
 static bool ggml_backend_sycl_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-    return get_op_batch_size(op) >= min_batch_size;
-    GGML_UNUSED(dev);
+    ggml_backend_sycl_device_context * sycl_ctx = (ggml_backend_sycl_device_context *)dev->context;
+    return get_op_batch_size(op) >= sycl_ctx->op_offload_min_batch_size;
 }
 
 static ggml_backend_event_t
@@ -4576,6 +5064,7 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
             ggml_backend_sycl_reg_context * ctx = new ggml_backend_sycl_reg_context;
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             for (int i = 0; i < ggml_sycl_info().device_count; i++) {
                 ggml_backend_sycl_device_context * dev_ctx = new ggml_backend_sycl_device_context;
@@ -4589,6 +5078,7 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                     prop, dpct::dev_mgr::instance().get_device(i))));
 
                 dev_ctx->description = prop.get_name();
+                dev_ctx->op_offload_min_batch_size = min_batch_size;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface       = */ ggml_backend_sycl_device_interface,