whispercpp 1.3.3 → 1.3.5

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

@@ -30,19 +30,29 @@
 #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-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/norm.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
+#include "ggml-sycl/set_rows.hpp"
+#include "ggml-sycl/set.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
 #include "ggml-sycl/getrows.hpp"
+#include "ggml-sycl/repeat_back.hpp"
+#include "ggml-sycl/quantize.hpp"
+#include "ggml-sycl/ssm_conv.hpp"
 #include "ggml.h"
 
 static bool g_sycl_loaded = false;
@@ -51,6 +61,7 @@ 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;
 
 static ggml_sycl_device_info ggml_sycl_init() {
     ggml_sycl_device_info info = {};
@@ -83,7 +94,10 @@ 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].opt_feature.reorder = !device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
+        info.devices[i].nsm = prop.get_max_compute_units();
+        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.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
 
@@ -231,7 +245,20 @@ 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;
@@ -1372,120 +1399,6 @@ typedef void (*ggml_sycl_op_mul_mat_t)(
 
 
 
-template<int QUANT_BLOCK_TILE>
-static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded,
-                          const sycl::nd_item<3> &item_ct1) {
-    const int ix = (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2)) * QUANT_BLOCK_TILE;
-
-    if (ix >= kx_padded) {
-        return;
-    }
-
-    const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                   item_ct1.get_local_id(1);
-
-    const int i_padded = iy*kx_padded + ix;
-
-    block_q8_1 * y = (block_q8_1 *) vy;
-
-    const int ib = i_padded / QK8_1; // block index
-    const int iqs = i_padded % QK8_1; // quant index
-    typedef  sycl::vec<float, QUANT_BLOCK_TILE> TC;
-    typedef  sycl::vec<int8_t, QUANT_BLOCK_TILE> TQ;
-    TC zeros;
-    TQ qzeros;
-#pragma unroll
-    for (int i = 0; i < QUANT_BLOCK_TILE; i++)
-    {
-        zeros[i] = 0.f;
-        qzeros[i] = 0;
-    }
-    const TC xi = ix < kx ? *(const TC *)&x[iy * kx + ix] : zeros;
-    float sum = xi[0];
-    float amax = sycl::fabs(xi[0]);
-#pragma unroll
-    for (int i = 1; i < QUANT_BLOCK_TILE; i++)
-    {
-        sum += xi[i];
-        amax = sycl::fmax(sycl::fabs(xi[i]), amax);
-    }
-    sum = warp_reduce_sum(sum, item_ct1);
-    amax = warp_reduce_max(amax, item_ct1);
-
-    const float d = amax / 127;
-    TQ q = qzeros;
-    if (amax != 0.0f)
-    {
-#pragma unroll
-        for (int i = 0; i < QUANT_BLOCK_TILE; i++) {
-            q[i] = sycl::round(xi[i] / d);
-        }
-    }
-
-    *(TQ *)&y[ib].qs[iqs] = q;
-
-    if (iqs > 0) {
-        return;
-    }
-
-    reinterpret_cast<sycl::half &>(y[ib].ds.x()) = d;
-    reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
-}
-
-template <int ElementsPerWI>
-static __dpct_inline__ void quantize_and_reorder_q8_1(const float * __restrict__ x, void * reordered_q8_tensor,
-                                                      const int kx, const int kx_padded, const sycl::nd_item<1> & it) {
-    /*
-        Quantizes and reorders the resultant q8 tensor in a per row fashion
-        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
-    */
-
-    auto subgroup_id = it.get_group(0);
-    auto wi_id       = it.get_local_id(0);
-
-    const int num_blocks_per_row = kx / QK8_1;
-    auto      row                = subgroup_id / num_blocks_per_row;
-    auto      col                = subgroup_id % num_blocks_per_row;
-
-    auto row_offset = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
-    auto col_offset = QK8_1 * col + wi_id * ElementsPerWI;
-
-    auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
-    auto ds_ptr    = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
-
-    sycl::vec<float, ElementsPerWI>  wi_f32_vals;
-    sycl::vec<int8_t, ElementsPerWI> quantized_values;
-
-    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
-    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
-
-    float sum  = 0.0f;
-    float amax = 0.0f;
-
-#pragma unroll(ElementsPerWI)
-    for (int i = 0; i < ElementsPerWI; i++) {
-        sum += wi_f32_vals[i];
-        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
-        quantized_values[i] = 0;
-    }
-    sum     = sycl::reduce_over_group(it.get_group(), sum, sycl::plus<float>());
-    amax    = sycl::reduce_over_group(it.get_group(), amax, sycl::maximum<float>());
-    float d = amax == 0 ? 1 : amax / 127;
-
-#pragma unroll(ElementsPerWI)
-    for (int i = 0; i < ElementsPerWI; i++) {
-        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
-    }
-
-    d = amax == 0 ? 0 : d;
-
-    *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
-    if (wi_id == 0) {
-        *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
-    }
-}
-
 static void mul_mat_p021_f16_f32(
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y,
@@ -1545,7 +1458,7 @@ static void mul_mat_p021_f16_f32(
 
 static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor,
+    const int row_stride_x, const int channel_stride_x,const int channel_stride_y, const int channel_x_divisor,
     const sycl::nd_item<3> &item_ct1) {
 
     const sycl::half *x = (const sycl::half *)vx;
@@ -1556,7 +1469,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
                         item_ct1.get_local_id(0);
     const int channel_x = channel / channel_x_divisor;
 
-    const int nrows_y   = ncols_x;
     const int nrows_dst = nrows_x;
     const int row_dst   = row_x;
 
@@ -1575,7 +1487,7 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
         const int row_y = col_x;
 
         const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
+        const int iy = channel * channel_stride_y + row_y;
 
         const float xi =
             sycl::vec<sycl::half, 1>(x[ix])
@@ -1624,60 +1536,70 @@ 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) +
@@ -1695,7 +1617,7 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-static void scale_f32(const float * x, float * dst, const float scale, const int k,
+static void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k,
                       const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -1704,7 +1626,7 @@ static void scale_f32(const float * x, float * dst, const float scale, const int
         return;
     }
 
-    dst[i] = scale * x[i];
+    dst[i] = scale * x[i] + bias;
 }
 
 
@@ -1770,32 +1692,6 @@ static  void pool2d_nchw_kernel(
         o_ptr[cur_oh * ow + cur_ow] = res;
 }
 
-static void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
-                                   bool reorder_q8_tensor, queue_ptr stream) {
-    if (reorder_q8_tensor) {
-        auto local_range      = std::size_t(WARP_SIZE);
-        auto num_quant_blocks = ky * (kx / QK8_1);
-        auto global_range     = num_quant_blocks * local_range;
-        stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
-                             [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                 quantize_and_reorder_q8_1<QK8_1 / WARP_SIZE>(x, vy, kx, kx_padded, it);
-                             });
-    } else {
-        const int            block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
-        const sycl::range<3> num_blocks(1, ky, block_num_x);
-        int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
-        static_assert(QK8_1 % WARP_SIZE == 0);
-        const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
-        {
-            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
-
-            stream->parallel_for(sycl::nd_range<3>(num_blocks * block_size, block_size),
-                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                     quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
-                                 });
-        }
-    }
-}
 
 static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
                                            float *dst, const int ncols_x,
@@ -1822,7 +1718,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
 static void ggml_mul_mat_vec_nc_f16_f32_sycl(
     const void *vx, const float *y, float *dst, const int ncols_x,
     const int nrows_x, const int row_stride_x, const int nchannels_x,
-    const int nchannels_y, const int channel_stride_x, queue_ptr stream) {
+    const int nchannels_y, const int channel_stride_x, const int channel_stride_y, queue_ptr stream) {
 
     const sycl::range<3> block_nums(nchannels_y, nrows_x, 1);
     const sycl::range<3> block_dims(1, 1, WARP_SIZE);
@@ -1834,7 +1730,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x,
-                                       row_stride_x, channel_stride_x,
+                                       row_stride_x, channel_stride_x, channel_stride_y,
                                        nchannels_y / nchannels_x, item_ct1);
             });
     }
@@ -1842,7 +1738,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
 
 
 
-static void scale_f32_sycl(const float *x, float *dst, const float scale,
+static void scale_f32_sycl(const float *x, float *dst, const float scale, const float bias,
                            const int k, queue_ptr stream) {
     const int num_blocks = (k + SYCL_SCALE_BLOCK_SIZE - 1) / SYCL_SCALE_BLOCK_SIZE;
     stream->parallel_for(
@@ -1850,7 +1746,7 @@ static void scale_f32_sycl(const float *x, float *dst, const float scale,
                               sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE)),
         [=](sycl::nd_item<3> item_ct1) {
-            scale_f32(x, dst, scale, k, item_ct1);
+            scale_f32(x, dst, scale, bias, k, item_ct1);
         });
 }
 
@@ -1876,37 +1772,51 @@ 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) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler &cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                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());
                 });
         });
     } else if (order == GGML_SORT_ORDER_DESC) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler &cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                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());
                 });
         });
@@ -1921,47 +1831,50 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = 256 * sizeof(float);
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
+    stream->submit([&](sycl::handler &cgh) {
         sycl::local_accessor<float, 1> shared_data(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
         sycl::local_accessor<int, 1> shared_indices(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
 
-        sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            const int tid = item_ct1.get_local_id(2);
-            const int row = item_ct1.get_global_id(1);
-
-            float max_val = -INFINITY;
-            int   max_idx = -1;
-
-            for (int col = tid; col < ncols; col += 256) {
-                float val = x[row * ncols + col];
-                if (val > max_val) {
-                    max_val = val;
-                    max_idx = col;
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                const int tid = item_ct1.get_local_id(2);
+                const int row = item_ct1.get_global_id(1);
+
+                float max_val = -INFINITY;
+                int max_idx = -1;
+
+                for (int col = tid; col < ncols; col += 256) {
+                    float val = x[row * ncols + col];
+                    if (val > max_val) {
+                        max_val = val;
+                        max_idx = col;
+                    }
                 }
-            }
 
-            shared_data[tid]    = max_val;
-            shared_indices[tid] = max_idx;
-            item_ct1.barrier(sycl::access::fence_space::local_space);
-
-            for (int stride = 256 / 2; stride > 0; stride >>= 1) {
-                if (tid < stride) {
-                    float val1 = shared_data[tid];
-                    float val2 = shared_data[tid + stride];
-                    if (val2 > val1) {
-                        shared_data[tid]    = val2;
-                        shared_indices[tid] = shared_indices[tid + stride];
+                shared_data[tid] = max_val;
+                shared_indices[tid] = max_idx;
+                item_ct1.barrier(sycl::access::fence_space::local_space);
+
+                for (int stride = 256/2; stride > 0; stride >>= 1) {
+                    if (tid < stride) {
+                        float val1 = shared_data[tid];
+                        float val2 = shared_data[tid + stride];
+                        if (val2 > val1) {
+                            shared_data[tid] = val2;
+                            shared_indices[tid] = shared_indices[tid + stride];
+                        }
                     }
+                    item_ct1.barrier(sycl::access::fence_space::local_space);
                 }
-                item_ct1.barrier(sycl::access::fence_space::local_space);
-            }
 
-            if (tid == 0) {
-                dst[row] = shared_indices[0];
-            }
-        });
+
+                if (tid == 0) {
+                    dst[row] = shared_indices[0];
+                }
+            });
     });
 }
 static void diag_mask_inf_f32_sycl(const float *x, float *dst,
@@ -2123,8 +2036,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #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>(),
+                DnnlGemmWrapper::row_gemm(ctx,row_diff, src1_ncols , ne10, src0_ptr,
+                                     DnnlGemmWrapper::to_dt<sycl::half>(), src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2170,8 +2083,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #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>(),
+            DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
+                                      DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2261,6 +2174,30 @@ 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);
@@ -2275,7 +2212,8 @@ 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);
 }
 
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -2319,9 +2257,11 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     float *       dst_dd  = static_cast<float *>(dst->data);
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float bias;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&bias,  (float *) dst->op_params + 1, sizeof(float));
 
-    scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(dst->src[0]), main_stream);
+    scale_f32_sycl(src0_dd, dst_dd, scale, bias, ggml_nelements(dst->src[0]), main_stream);
     /*
     DPCT1010:87: SYCL uses exceptions to report errors and does not use the
     error codes. The call was replaced with 0. You need to rewrite this code.
@@ -2370,10 +2310,10 @@ static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
     peer_access_enabled = enable_peer_access;
 }
 
+template <template <int> typename quantize_f>
 static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                  const ggml_tensor *src1, ggml_tensor *dst,
-                                 ggml_sycl_op_mul_mat_t op,
-                                 const bool convert_src1_to_q8_1) try {
+                                 ggml_sycl_op_mul_mat_t op) try {
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
@@ -2468,6 +2408,8 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
         }
     }
 
+    constexpr bool quantize_enabled = !std::is_same_v<quantize_f<QK8_1 / WARP_SIZE>,
+                                                      no_quantize_q8_1<QK8_1 / WARP_SIZE>>;
     for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
         if ((!split && i != ctx.device) || dev[i].row_low == dev[i].row_high) {
             continue;
@@ -2493,20 +2435,19 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
             dev[i].src1_ddf = dev[i].src1_ddf_alloc.alloc(ctx.pool(i), ggml_nelements(src1));
         }
 
-        if (convert_src1_to_q8_1) {
+        if constexpr(quantize_enabled) {
             dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
 
             if (src1_on_device && src1_is_contiguous) {
-                bool reorder_q8_tensor = src0->extra && ((ggml_tensor_extra_gpu *)src0->extra)->optimized_feature.reorder;
                 scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                      /*num_src=*/2, " : converting src1 to Q8_1");
-                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, reorder_q8_tensor, stream);
-                /*
-                DPCT1010:90: SYCL uses exceptions to report errors and does not
-                use the error codes. The call was replaced with 0. You need to
-                rewrite this code.
-                */
-                SYCL_CHECK(0);
+                try {
+                    quantize_row_q8_1_sycl<quantize_f>(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                } catch (sycl::exception const &exc) {
+                    std::cerr << "Quantize_row_q8_1_sycl error" << exc.what() << "Exception caught at file:" << __FILE__
+                              << ", line:" << __LINE__ << std::endl;
+                    std::exit(1);
+                }
             }
         }
 
@@ -2522,11 +2463,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
     // here an event is recorded that signals that the main device has finished calculating the input data
     if (split && used_devices > 1) {
         ggml_sycl_set_device(ctx.device);
-        /*
-        DPCT1024:91: The original code returned the error code that was further
-        consumed by the program logic. This original code was replaced with 0.
-        You may need to rewrite the program logic consuming the error code.
-        */
         SYCL_CHECK(CHECK_TRY_ERROR(
             *src0_extra->events[ctx.device][0] =
                 ctx.stream()->ext_oneapi_submit_barrier()));
@@ -2550,11 +2486,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
             // wait for main GPU data if necessary
             if (split && (i != ctx.device || is != 0)) {
-                /*
-                DPCT1009:163: 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.
-                */
                 SYCL_CHECK(CHECK_TRY_ERROR(stream->ext_oneapi_submit_barrier(
                     {*src0_extra->events[ctx.device][0]})));
             }
@@ -2580,39 +2511,42 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 // copy src0, src1 to device if necessary
                 if (src1_is_contiguous) {
                     if (i != ctx.device) {
-                        if (convert_src1_to_q8_1) {
+                        if constexpr (quantize_enabled) {
                             char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
-                          SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
-                                src1_ddq_i, src1_ddq_i_source,
-                                src1_ncols * src1_padded_col_size * q8_1_ts /
-                                    q8_1_bs).wait()));
+                            SYCL_CHECK(
+                                CHECK_TRY_ERROR(stream
+                                                    ->memcpy(src1_ddq_i, src1_ddq_i_source,
+                                                             src1_ncols * src1_padded_col_size * q8_1_ts / q8_1_bs)
+                                                    .wait()));
                         } else {
-
                             float * src1_ddf_i_source = (float *) src1_extra->data_device[ctx.device];
-                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
+                            src1_ddf_i_source += (i0 * ne11 + src1_col_0) * ne10;
 
-                            SYCL_CHECK(CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream,
-                                src1_ddf_i, src1_ddf_i_source,
-                                src1_ncols * ne10 * sizeof(float))));
+                            SYCL_CHECK(
+                                CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream, src1_ddf_i, src1_ddf_i_source,
+                                                               src1_ncols * ne10 * sizeof(float))));
                         }
                     }
-                } else if (src1_on_device && !src1_is_contiguous) {
-                    SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
-                                   src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
-                    GGML_ABORT("fatal error");
-                }
+                    if (src1_on_device) {
+                        SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, src1_col_0,
+                                                           src1_col_0 + src1_ncols, stream));
+                    } else {
+                        GGML_ABORT("src1 is non-contiguous and not on device");
+                    }
 
-                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
-                    scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
-                                                         /*num_src=*/2, " : converting src1 to Q8_1");
-                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, false, stream);
-                    /*
-                    DPCT1010:92: SYCL uses exceptions to report errors and does
-                    not use the error codes. The call was replaced with 0. You
-                    need to rewrite this code.
-                    */
-                    SYCL_CHECK(0);
+                    if constexpr (quantize_enabled) {
+                        scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
+                                                             /*num_src=*/2, " : converting src1 to Q8_1");
+                        try {
+                            quantize_row_q8_1_sycl<quantize_q8_1>(src1_ddf_i, src1_ddq_i, ne10, src1_ncols,
+                                                                  src1_padded_col_size, stream);
+                        } catch (const sycl::exception & exc) {
+                            std::cerr << "Quantize_row_q8_1_sycl error" << exc.what()
+                                      << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+                            std::exit(1);
+                        }
+                    }
                 }
 
                 if (src1_col_0 == 0 && !src0_is_contiguous && i02 % i02_divisor == 0) {
@@ -2624,12 +2558,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 // do the computation
                 SYCL_CHECK(CHECK_TRY_ERROR(op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
                     dev[i].row_low, dev[i].row_high, src1_ncols, src1_padded_col_size, stream)));
-                /*
-                DPCT1010:93: SYCL uses exceptions to report errors and does not
-                use the error codes. The call was replaced with 0. You need to
-                rewrite this code.
-                */
-                SYCL_CHECK(0);
 
                 // copy dst to host or other device if necessary
                 if (!dst_on_device) {
@@ -2660,12 +2588,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
                 // add event for the main device to wait on until other device is done
                 if (split && (i != ctx.device || is != 0)) {
-                    /*
-                    DPCT1024:94: The original code returned the error code that
-                    was further consumed by the program logic. This original
-                    code was replaced with 0. You may need to rewrite the
-                    program logic consuming the error code.
-                    */
                     SYCL_CHECK(CHECK_TRY_ERROR(
                         *src0_extra->events[i][is] =
                             stream->ext_oneapi_submit_barrier()));
@@ -2698,6 +2620,10 @@ 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);
@@ -2714,6 +2640,11 @@ 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);
@@ -2764,6 +2695,8 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->ne[1] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -2773,6 +2706,7 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     const int64_t nb02 = src0->nb[2];
 
     const int64_t ne12 = src1->ne[2];
+    const int64_t nb11 = src1->nb[1];
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();
@@ -2783,8 +2717,9 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
 
     const int64_t row_stride_x = nb01 / sizeof(sycl::half);
     const int64_t channel_stride_x = nb02 / sizeof(sycl::half);
+    const int64_t channel_stride_y = nb11 / sizeof(float);
 
-    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
+    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x,channel_stride_y, main_stream);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2838,8 +2773,11 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     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_src0 = ggml_type_size(src0->type);
     const size_t       type_size_src1 = ggml_type_size(src1->type);
-    GGML_ASSERT(nb10 == type_size_src1);
+
+    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
+    bool is_src1_cont_2 = ggml_is_contiguous_2(src1);
 
     // SRC1 strides
     int64_t                          s11 = nb11 / type_size_src1;
@@ -2851,16 +2789,47 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     if (src1->type != GGML_TYPE_F16) {
         scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                 " : converting src1 to fp16");
-        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
-        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
+
+        // iterate tensor dims and find the slowest moving dim and stride
+        int last_dim=0;
+        int last_str=0;
+        size_t largest_str=0;
+        for(int i = 0; i< 4; i++){
+            // last stride is always the largest
+            if(src1->nb[i] == largest_str){
+                if(src1->ne[last_dim] == 1){
+                    last_str = i;
+                    last_dim = i;
+                }
+            }
+            if(src1->nb[i] > largest_str){
+                largest_str = src1->nb[i];
+                last_str = i;
+                last_dim = i;
+            }
+
+        }
+#if GGML_SYCL_DNNL
+        // oneDNN handles strided data and does not need overhead of 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);
+        GGML_ASSERT(to_fp16_sycl != nullptr);
+        to_fp16_sycl(src1_f16, src1_f16_alloc.get(), ne_src1, queue);
+# 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);
+        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
 
         src1_f16 = src1_f16_alloc.get();
         s11      = ne10;
         s12      = ne11 * s11;
         s13      = ne12 * s12;
+
+        is_src1_cont_2 = true;
     }
 
     ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool());
@@ -2889,48 +2858,115 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 
 #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);
+            int64_t str_a0 = nb00 / type_size_src0;
+            int64_t str_a1 = nb01 / type_size_src0;
+            int64_t str_a2 = nb02 / type_size_src0;
+
+            int64_t str_b0 = nb10 / type_size_src1;
+            int64_t str_b1 = nb11 / type_size_src1;
+            int64_t str_b2 = nb12 / type_size_src1;
+
+            auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
+                                                const sycl::half *src1, float *dst,
+                                                int64_t a0, int64_t a1, int64_t batcha,
+                                                int64_t /*b0*/, int64_t b1, int64_t batchb,
+                                                int64_t sa0, int64_t sa1, int64_t sa2,
+                                                int64_t sb0, int64_t sb1, int64_t sb2,
+                                                int64_t sd2) {
+                bool supported_broadcast = batchb == batcha ? true
+                        : batchb == 1 || batcha == 1        ? true
+                                                            : false;
+                if (supported_broadcast) {
+                    DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2, src1,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sb0, sb1, sb2, dst,
+                            DnnlGemmWrapper::to_dt<float>(), queue, batcha, batchb);
+                } else {
+                    // iterate over batches from smaller set of matrices (matrix 0)
+                    int64_t batches0 = batcha;
+                    int64_t batches1 = batchb;
+
+                    if (batches0 > batches1) {
+                        int64_t num_mul_mats = batches1;
+                        int64_t sub_batch = batches0 / num_mul_mats;
+                        // src0 is batched and bigger, shift and multiply with src1
+                        for (int64_t i0 = 0; i0 < num_mul_mats; i0++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i0 * sub_batch);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i0);
+                            float *dst_shifted = dst + (sd2 * i0 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, sub_batch, 1);
+                        }
+                    } else {
+                        int64_t num_mul_mats = batches0;
+                        int64_t sub_batch = batches1 / num_mul_mats;
+                        // src1 is batched and bigger, shift and multiply with src0
+                        for (int64_t i1 = 0; i1 < num_mul_mats; i1++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i1);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i1 * sub_batch);
+                            float *dst_shifted = dst + (sd2 * i1 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, 1, sub_batch);
+                        }
+                    }
                 }
-            }
-        } 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);
+            };
+
+            const bool cont_batches_dim2_a = nb02 * ne02 == nb03;
+            const bool cont_batches_dim2_b = nb12 * ne12 == nb13;
+            const bool cont_batches_dim3_a = ne02 == 1 && nb02 * ne01 == nb03;
+            const bool cont_batches_dim3_b = ne12 == 1 && nb12 * ne11 == nb13;
+            if (cont_batches_dim2_a && cont_batches_dim2_b) {
+                // A batch is considered contiguous if the dimension 2 is not strided
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
+                        str_b2, nb2 / sizeof(float));
+            } else if (cont_batches_dim3_a && cont_batches_dim3_b) {
+                // This case is similar to the one above with the difference that only the batch in dimension 3 is used and the dimension 2 is of size 1.
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                int64_t str_a3 = nb03 / type_size_src0;
+                int64_t str_b3 = nb13 / type_size_src1;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a3, str_b0, str_b1,
+                        str_b3, nb2 / sizeof(float));
+            } else {
+                for (int64_t b_a = 0; b_a < ne03; b_a++) {
+                    const sycl::half *src0_f16_shifted
+                            = src0_f16 + (nb03 * b_a / type_size_src0);
+                    const sycl::half *src1_f16_shifted
+                            = src1_f16 + (nb13 * b_a / type_size_src1);
+                    float *dst_shifted = dst_ddf + (nb3 * b_a / sizeof(float));
+                    int64_t batches0 = ne02;
+                    int64_t batches1 = ne12;
+                    launch_gemm_for_batches(src0_f16_shifted, src1_f16_shifted, dst_shifted,
+                            ne00, ne01, batches0, ne10, ne11, batches1, str_a0, str_a1,
+                            str_a2, str_b0, str_b1, str_b2, nb2 / sizeof(float));
                 }
             }
-        }
+
     }
     else
 #endif
     {
-        if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+        if (r2 == 1 && r3 == 1 && is_src0_cont_2 && is_src1_cont_2) {
+            // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+            const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+            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,
-                                                        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,
+                                                        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)));
         } else {
             const int ne23 = ne12 * ne13;
@@ -2945,7 +2981,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 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);
-                sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                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);
                 });
@@ -3026,19 +3062,51 @@ 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;
@@ -3049,9 +3117,11 @@ 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) {
@@ -3060,14 +3130,19 @@ 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;
 
@@ -3080,9 +3155,11 @@ 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) {
@@ -3091,42 +3168,46 @@ 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) {
@@ -3233,6 +3314,7 @@ 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
     use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS);
@@ -3240,7 +3322,6 @@ 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 (!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
@@ -3260,26 +3341,27 @@ 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_contiguous(src1) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1 && src1->ne[3] == 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) {
+    } 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) {
-        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);
+        ggml_sycl_op_mul_mat<no_quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec);
     } else if (use_mul_mat_vec_q) {
-        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);
+        ggml_tensor_extra_gpu * extra = static_cast<ggml_tensor_extra_gpu *>(src0->extra);
+        if (extra && extra->optimized_feature.reorder) {
+            ggml_sycl_op_mul_mat<quantize_and_reorder_q8_1_soa>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q);
+        } else {
+            ggml_sycl_op_mul_mat<quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q);
+        }
     } else if (use_mul_mat_q) {
-        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);
+        ggml_sycl_op_mul_mat<quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q);
     } else {
-        constexpr bool convert_src1_to_q8_1 = false;
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1);
+        ggml_sycl_op_mul_mat<no_quantize_q8_1>(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl);
     }
 }
 
@@ -3446,10 +3528,13 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             SYCL_CHECK(CHECK_TRY_ERROR(
                 stream->memset(dev_cur_src1_row.get(), 0, sizeof(int))));
 
+            const unsigned int max_work_group_size = ggml_sycl_info().max_work_group_sizes[ctx.device];
+            assert(max_work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+
             {
-                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, max_work_group_size));
                 sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
-                sycl_launch(stream, [&](sycl::handler & cgh) {
+                stream->submit([&](sycl::handler &cgh) {
                     sycl::local_accessor<int, 0> src1_row_acc(cgh);
 
                     char *__restrict src1_contiguous_get =
@@ -3461,8 +3546,9 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
                     size_t ids_nb_ct6 = ids->nb[1];
                     size_t ids_nb_ct7 = ids->nb[0];
 
-                    sycl_parallel_for(
-                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
                             k_copy_src1_to_contiguous(
                                 src1_original, src1_contiguous_get,
                                 dev_cur_src1_row_get,
@@ -3491,16 +3577,17 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
             {
-                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, max_work_group_size));
                 sycl::range<3> grid_dims(1, 1, num_src1_rows);
-                sycl_launch(stream, [&](sycl::handler & cgh) {
+                stream->submit([&](sycl::handler &cgh) {
                     const char *__restrict dst_contiguous_get =
                         dst_contiguous.get();
                     const mmid_row_mapping *__restrict dev_row_mapping_get =
                         dev_row_mapping.get();
 
-                    sycl_parallel_for(
-                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
                             k_copy_dst_from_contiguous(dst_original,
                                                        dst_contiguous_get,
                                                        dev_row_mapping_get,
@@ -3549,6 +3636,12 @@ 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]));
@@ -3600,9 +3693,18 @@ 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;
         case GGML_OP_DUP:
             ggml_sycl_dup(ctx, dst);
             break;
@@ -3610,9 +3712,15 @@ 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;
+        case GGML_OP_COUNT_EQUAL:
+            ggml_sycl_count_equal(ctx, dst);
+            break;
         case GGML_OP_ACC:
             ggml_sycl_acc(ctx, dst);
             break;
@@ -3672,6 +3780,18 @@ 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;
             }
@@ -3687,6 +3807,15 @@ 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;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_sycl_geglu_quick(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3700,6 +3829,9 @@ 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;
@@ -3709,6 +3841,9 @@ 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;
@@ -3768,6 +3903,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         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;
@@ -3783,6 +3921,9 @@ 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;
@@ -3798,6 +3939,15 @@ 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_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;
         default:
             return false;
     }
@@ -3805,6 +3955,7 @@ 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);
 }
 
@@ -3999,6 +4150,18 @@ 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;
@@ -4100,6 +4263,7 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
     /* .event_record            = */ ggml_backend_sycl_event_record,
     /* .event_wait              = */ ggml_backend_sycl_event_wait,
+    /* .graph_optimize          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -4122,6 +4286,7 @@ 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) {
@@ -4192,6 +4357,9 @@ 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:
             {
@@ -4204,21 +4372,26 @@ 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_ELU:
+                    return true;
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                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
@@ -4232,6 +4405,9 @@ 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]);
                 default:
                     return false;
@@ -4240,15 +4416,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
-                struct ggml_tensor * a;
-                struct ggml_tensor * b;
-                if (op->op == GGML_OP_MUL_MAT) {
-                    a = op->src[0];
-                    b = op->src[1];
-                } else {
-                    a = op->src[2];
-                    b = op->src[1];
-                }
+                struct ggml_tensor * a = op->src[0];
+                struct ggml_tensor * b = op->src[1];
+
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
@@ -4263,7 +4433,21 @@ 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) {
+                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 && 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) {
                     return false;
                 }
                 return true;
@@ -4285,6 +4469,20 @@ 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 ||
+                         op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q5_0 ||
+                         op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_IQ4_NL) &&
+                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32));
+            }
+            break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -4354,11 +4552,12 @@ 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:
@@ -4366,14 +4565,17 @@ 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:
@@ -4386,35 +4588,62 @@ 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:
-        case GGML_OP_RMS_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);
+        case GGML_OP_RMS_NORM_BACK:
+            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
         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_IM2COL:
             return true;
         case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
-        case GGML_OP_POOL_2D:
+            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_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_POOL_2D:
         case GGML_OP_ACC:
+            return true;
         case GGML_OP_PAD:
+            // 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:
             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;
         default:
             return false;
     }
@@ -4446,9 +4675,8 @@ 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
@@ -4571,6 +4799,7 @@ 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;
@@ -4584,6 +4813,7 @@ 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,
@@ -4619,10 +4849,10 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     };
 
     ggml_backend_t sycl_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_sycl_guid(),
-        /* .interface = */ ggml_backend_sycl_interface,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_sycl_reg(), device),
-        /* .context   = */ ctx
+        /* .guid    = */ ggml_backend_sycl_guid(),
+        /* .iface   = */ ggml_backend_sycl_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_sycl_reg(), device),
+        /* .context = */ ctx
     };
 
     return sycl_backend;