npm - @fugood/llama.node - Versions diffs - 0.3.13 → 0.3.15 - Mend

@fugood/llama.node 0.3.13 → 0.3.15

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (184) hide show

package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-cuda/arm64/llama-node.node +0 -0
package/bin/linux-cuda/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/lib/binding.ts +1 -1
package/package.json +1 -1
package/src/LlamaContext.cpp +98 -76
package/src/LlamaContext.h +1 -1
package/src/common.hpp +1 -2
package/src/llama.cpp/.github/workflows/build.yml +89 -10
package/src/llama.cpp/.github/workflows/server.yml +2 -0
package/src/llama.cpp/CMakeLists.txt +9 -1
package/src/llama.cpp/cmake/common.cmake +2 -0
package/src/llama.cpp/common/CMakeLists.txt +3 -3
package/src/llama.cpp/common/arg.cpp +132 -13
package/src/llama.cpp/common/chat.cpp +960 -266
package/src/llama.cpp/common/chat.h +135 -0
package/src/llama.cpp/common/common.cpp +33 -174
package/src/llama.cpp/common/common.h +27 -67
package/src/llama.cpp/common/json-schema-to-grammar.cpp +4 -5
package/src/llama.cpp/common/json-schema-to-grammar.h +0 -1
package/src/llama.cpp/common/{minja.hpp → minja/minja.hpp} +37 -5
package/src/llama.cpp/common/ngram-cache.cpp +1 -0
package/src/llama.cpp/common/sampling.cpp +45 -7
package/src/llama.cpp/common/speculative.cpp +10 -9
package/src/llama.cpp/common/speculative.h +1 -1
package/src/llama.cpp/docs/build.md +45 -7
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +2 -2
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +4 -2
package/src/llama.cpp/examples/embedding/embedding.cpp +2 -1
package/src/llama.cpp/examples/export-lora/export-lora.cpp +4 -2
package/src/llama.cpp/examples/gritlm/gritlm.cpp +2 -2
package/src/llama.cpp/examples/imatrix/imatrix.cpp +3 -4
package/src/llama.cpp/examples/infill/infill.cpp +2 -2
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +2 -2
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +5 -5
package/src/llama.cpp/examples/llava/CMakeLists.txt +7 -0
package/src/llama.cpp/examples/llava/clip.cpp +373 -107
package/src/llama.cpp/examples/llava/clip.h +19 -3
package/src/llama.cpp/examples/llava/gemma3-cli.cpp +341 -0
package/src/llama.cpp/examples/llava/llava.cpp +4 -2
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +30 -11
package/src/llama.cpp/examples/lookahead/lookahead.cpp +7 -6
package/src/llama.cpp/examples/lookup/lookup.cpp +1 -1
package/src/llama.cpp/examples/main/main.cpp +79 -34
package/src/llama.cpp/examples/parallel/parallel.cpp +6 -5
package/src/llama.cpp/examples/passkey/passkey.cpp +15 -14
package/src/llama.cpp/examples/perplexity/perplexity.cpp +6 -6
package/src/llama.cpp/examples/quantize/quantize.cpp +1 -0
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +2 -2
package/src/llama.cpp/examples/retrieval/retrieval.cpp +1 -1
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.cpp +882 -237
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.h +35 -26
package/src/llama.cpp/examples/run/run.cpp +196 -108
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +2 -2
package/src/llama.cpp/examples/server/server.cpp +113 -101
package/src/llama.cpp/examples/server/utils.hpp +94 -105
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +2 -2
package/src/llama.cpp/examples/speculative/speculative.cpp +14 -14
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +1 -1
package/src/llama.cpp/examples/sycl/run-llama2.sh +2 -2
package/src/llama.cpp/examples/tts/tts.cpp +263 -151
package/src/llama.cpp/ggml/CMakeLists.txt +14 -1
package/src/llama.cpp/ggml/cmake/common.cmake +26 -0
package/src/llama.cpp/ggml/include/ggml-alloc.h +1 -1
package/src/llama.cpp/ggml/include/ggml-backend.h +3 -3
package/src/llama.cpp/ggml/include/ggml-cpu.h +3 -0
package/src/llama.cpp/ggml/include/ggml.h +29 -1
package/src/llama.cpp/ggml/src/CMakeLists.txt +15 -34
package/src/llama.cpp/ggml/src/ggml-alloc.c +24 -15
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +58 -54
package/src/llama.cpp/ggml/src/ggml-backend.cpp +10 -8
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +6 -2
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +3 -7
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +3 -5
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +139 -16
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +4 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +151 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +1546 -387
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +1645 -113
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +22 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +259 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +61 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +288 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +15 -2
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +2 -1
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +3 -1
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +14 -0
package/src/llama.cpp/ggml/src/ggml-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +4 -5
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +242 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +6 -6
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +1 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +315 -138
package/src/llama.cpp/ggml/src/ggml-quants.c +114 -114
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +5 -0
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +17 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +117 -36
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +33 -4
package/src/llama.cpp/ggml/src/ggml-sycl/convert.hpp +2 -2
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.cpp +701 -0
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.hpp +11 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dequantize.hpp +55 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +147 -16
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +40 -40
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.cpp +307 -0
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +262 -746
package/src/llama.cpp/ggml/src/ggml-sycl/mmq.cpp +0 -1
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +75 -78
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +114 -6
package/src/llama.cpp/ggml/src/ggml-sycl/norm.hpp +6 -0
package/src/llama.cpp/ggml/src/ggml-sycl/softmax.cpp +4 -1
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.cpp +305 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.hpp +10 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +498 -188
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +0 -4
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +16 -3
package/src/llama.cpp/ggml/src/ggml.c +93 -5
package/src/llama.cpp/include/llama.h +105 -27
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.out +46 -0
package/src/llama.cpp/requirements/requirements-all.txt +1 -0
package/src/llama.cpp/requirements/requirements-tool_bench.txt +12 -0
package/src/llama.cpp/requirements.txt +1 -0
package/src/llama.cpp/src/CMakeLists.txt +5 -2
package/src/llama.cpp/src/llama-adapter.cpp +19 -20
package/src/llama.cpp/src/llama-adapter.h +11 -9
package/src/llama.cpp/src/llama-arch.cpp +123 -16
package/src/llama.cpp/src/llama-arch.h +19 -0
package/src/llama.cpp/src/llama-batch.h +2 -2
package/src/llama.cpp/src/llama-chat.cpp +1 -0
package/src/llama.cpp/src/llama-context.cpp +2253 -1222
package/src/llama.cpp/src/llama-context.h +214 -77
package/src/llama.cpp/src/llama-cparams.h +1 -0
package/src/llama.cpp/src/llama-grammar.cpp +182 -182
package/src/llama.cpp/src/llama-grammar.h +12 -3
package/src/llama.cpp/src/llama-graph.cpp +1662 -0
package/src/llama.cpp/src/llama-graph.h +574 -0
package/src/llama.cpp/src/llama-hparams.cpp +8 -0
package/src/llama.cpp/src/llama-hparams.h +9 -0
package/src/llama.cpp/src/llama-io.cpp +15 -0
package/src/llama.cpp/src/llama-io.h +35 -0
package/src/llama.cpp/src/llama-kv-cache.cpp +1006 -291
package/src/llama.cpp/src/llama-kv-cache.h +178 -109
package/src/llama.cpp/src/llama-memory.cpp +1 -0
package/src/llama.cpp/src/llama-memory.h +21 -0
package/src/llama.cpp/src/llama-mmap.cpp +11 -1
package/src/llama.cpp/src/llama-model.cpp +8230 -122
package/src/llama.cpp/src/llama-model.h +34 -1
package/src/llama.cpp/src/llama-quant.cpp +10 -1
package/src/llama.cpp/src/llama-sampling.cpp +43 -10
package/src/llama.cpp/src/llama-vocab.cpp +12 -0
package/src/llama.cpp/src/llama.cpp +51 -9837
package/src/llama.cpp/tests/test-backend-ops.cpp +247 -112
package/src/llama.cpp/tests/test-chat-template.cpp +32 -22
package/src/llama.cpp/tests/test-chat.cpp +593 -395
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +63 -63
package/src/llama.cpp/tests/test-quantize-fns.cpp +1 -9
package/src/llama.cpp/Sources/llama/llama.h +0 -4
package/src/llama.cpp/common/chat.hpp +0 -55
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +0 -143
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +0 -9
/package/src/llama.cpp/common/{chat-template.hpp → minja/chat-template.hpp} +0 -0

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

@@ -39,8 +39,14 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
+#include "ggml-sycl/sycl_hw.hpp"
+#include "ggml-sycl/getrows.hpp"
+#include "ggml.h"
 static bool g_sycl_loaded = false;
+int g_ggml_sycl_debug = 0;
+int g_ggml_sycl_disable_optimize = 0;
+int g_ggml_sycl_disable_graph = 0;
 static ggml_sycl_device_info ggml_sycl_init() {
     ggml_sycl_device_info info = {};
@@ -63,14 +69,18 @@ static ggml_sycl_device_info ggml_sycl_init() {
     for (int i = 0; i < info.device_count; ++i) {
         info.devices[i].vmm = 0;
         dpct::device_info prop;
+        sycl::device device = dpct::dev_mgr::instance().get_device(i);
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
-            prop, dpct::dev_mgr::instance().get_device(i))));
+            prop, device)));
         info.default_tensor_split[i] = total_vram;
         total_vram += prop.get_global_mem_size();
         info.devices[i].cc =
             100 * prop.get_major_version() + 10 * prop.get_minor_version();
+        info.devices[i].hw_info = get_device_hw_info(&device);
+        info.devices[i].opt_feature = check_gpu_optimize_feature(info.devices[i].hw_info.arch);
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
@@ -86,7 +96,7 @@ const ggml_sycl_device_info & ggml_sycl_info() {
     return info;
 }
-void print_device_detail(int id, sycl::device &device, std::string device_type) {
+static void print_device_detail(int id, sycl::device &device, std::string device_type) {
     dpct::device_info prop;
     SYCL_CHECK(CHECK_TRY_ERROR(
@@ -109,6 +119,27 @@ void print_device_detail(int id, sycl::device &device, std::string device_type)
             global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
 }
+static void print_device_opt_feature(int device_count) {
+    GGML_LOG_INFO("SYCL Optimization Feature:\n");
+    GGML_LOG_INFO(
+        "|ID|        Device Type|Reorder|\n");
+    GGML_LOG_INFO(
+        "|--|-------------------|-------|\n");
+    std::map<std::string, size_t> DeviceNums;
+    for (int id = 0; id < device_count; ++id) {
+      sycl::device device = dpct::dev_mgr::instance().get_device(id);
+      std::string backend_type = get_device_backend_and_type(device);
+      int type_id = DeviceNums[backend_type]++;
+      std::stringstream device_type;
+      device_type << "[" << backend_type << ":" << std::to_string(type_id)
+                  << "]";
+      std::string device_type_s = device_type.str();
+      device_type_s = std::regex_replace(device_type_s, std::regex("ext_oneapi_"), "");
+      GGML_LOG_INFO("|%2d|%19s|%7s|\n", id, device_type_s.c_str(),
+        ggml_sycl_info().devices[id].opt_feature.reorder ? "Y": "N");
+    }
+}
 void ggml_backend_sycl_print_sycl_devices() {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n");
     int device_count = dpct::dev_mgr::instance().device_count();
@@ -137,6 +168,8 @@ void ggml_backend_sycl_print_sycl_devices() {
                   << "]";
       print_device_detail(id, device, device_type.str());
     }
+    print_device_opt_feature(device_count);
 }
 static inline int get_sycl_env(const char *env_name, int default_val) {
@@ -157,18 +190,24 @@ static void ggml_check_sycl() try {
     static bool initialized = false;
     if (!initialized) {
-        GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        GGML_LOG_INFO("GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
+        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
+        g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
+        GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
+        GGML_LOG_INFO("Running with Environment Variables:\n");
+        GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
+        GGML_LOG_INFO("Build with Macros:\n");
 #if defined(GGML_SYCL_FORCE_MMQ)
-        GGML_LOG_INFO("GGML_SYCL_FORCE_MMQ:   yes\n");
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
 #else
-        GGML_LOG_INFO("GGML_SYCL_FORCE_MMQ:   no\n");
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: no\n");
 #endif
 #if defined(GGML_SYCL_F16)
-        GGML_LOG_INFO("GGML_SYCL_F16: yes\n");
+        GGML_LOG_INFO("  GGML_SYCL_F16: yes\n");
 #else
-        GGML_LOG_INFO("GGML_SYCL_F16: no\n");
+        GGML_LOG_INFO("  GGML_SYCL_F16: no\n");
 #endif
 /* NOT REMOVE, keep it for next optimize for XMX.
@@ -240,19 +279,27 @@ struct ggml_backend_sycl_buffer_context {
     void * dev_ptr = nullptr;
     queue_ptr stream;
     std::string name;
+    optimize_feature opt_feature;
+    std::vector<ggml_tensor_extra_gpu *> tensor_extras;
-     ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) :
+    ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) :
         device(device), dev_ptr(dev_ptr), stream(stream) {
             check_allow_gpu_index(device);
             name = (GGML_SYCL_NAME + std::to_string(device));
+            opt_feature = ggml_sycl_info().devices[device].opt_feature;
         }
     ~ggml_backend_sycl_buffer_context() {
         if (dev_ptr != nullptr) {
             ggml_sycl_set_device(device);
             SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(dev_ptr, *stream)));
         }
+        //release extra used by tensors
+        for (ggml_tensor_extra_gpu * extra : tensor_extras) {
+            release_extra_gpu(extra);
+        }
     }
 };
@@ -280,16 +327,20 @@ static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) {
     return ctx->dev_ptr;
 }
-static void
+static enum ggml_status
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
     if (tensor->view_src != NULL) {
         assert(tensor->view_src->buffer->buft == buffer->buft);
-        return;
+        return GGML_STATUS_SUCCESS;
+    }
+    if (tensor->type == GGML_TYPE_Q4_0) {
+        ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
+        tensor->extra                 = extra;
+        ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
     }
     if (ggml_is_quantized(tensor->type)) {
         // initialize padding to 0 to avoid possible NaN values
@@ -302,6 +353,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                 padded_size - original_size).wait()));
         }
     }
+    return GGML_STATUS_SUCCESS;
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -315,7 +367,6 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 size_t size) try {
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
     ggml_sycl_set_device(ctx->device);
     auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
     SYCL_CHECK(
@@ -353,7 +404,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
-void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst,
+static void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst,
                     const void *ptr_src, size_t size) {
     char *host_buf = (char *)malloc(size);
     q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
@@ -439,6 +490,22 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
+static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) {
+    GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__);
+    if (buffer == nullptr) {
+        return;
+    }
+    ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
+    if (ctx != nullptr) {
+        for (ggml_tensor_extra_gpu * extra : ctx->tensor_extras) {
+            release_extra_gpu(extra);
+        }
+        ctx->tensor_extras.clear();  // reset the tensor_extras vector
+    }
+}
 static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_sycl_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_sycl_buffer_get_base,
@@ -448,7 +515,7 @@ static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
     /* .get_tensor      = */ ggml_backend_sycl_buffer_get_tensor,
     /* .cpy_tensor      = */ ggml_backend_sycl_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_sycl_buffer_clear,
-    /* .reset           = */ NULL,
+    /* .reset           = */ ggml_backend_sycl_buffer_reset,
 };
 // sycl buffer type
@@ -529,7 +596,6 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
-    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
     auto dev_count = ggml_backend_sycl_get_device_count();
@@ -557,7 +623,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
     return &ggml_backend_sycl_buffer_types[device];
 }
-ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) {
+static ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
     int device = ctx->device;
@@ -659,32 +725,7 @@ struct ggml_backend_sycl_split_buffer_type_context {
 struct ggml_backend_sycl_split_buffer_context {
     ~ggml_backend_sycl_split_buffer_context() try {
         for (ggml_tensor_extra_gpu * extra : tensor_extras) {
-            for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
-                for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
-                    if (extra->events[i][is] != nullptr) {
-                        /*
-                        DPCT1009:206: 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(
-                            dpct::destroy_event(extra->events[i][is])));
-                    }
-                }
-                if (extra->data_device[i] != nullptr) {
-                    /*
-                    DPCT1009:207: 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.
-                    */
-                    ggml_sycl_set_device(i);
-                    SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(
-                        extra->data_device[i], *(streams[i]))));
-                }
-            }
-            delete extra;
+            release_extra_gpu(extra, streams);
         }
     }
     catch (sycl::exception const &exc) {
@@ -709,7 +750,7 @@ static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buff
     GGML_UNUSED(buffer);
 }
-static void
+static enum ggml_status
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
@@ -722,7 +763,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
     ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
     ctx->tensor_extras.push_back(extra);
-        ctx->streams.push_back(&(dpct::get_current_device().default_queue()));
+    ctx->streams.push_back(&(dpct::get_current_device().default_queue()));
     for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
         int64_t row_low, row_high;
@@ -784,6 +825,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
         }
     }
     tensor->extra = extra;
+    return GGML_STATUS_SUCCESS;
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -1263,8 +1305,6 @@ std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_device(q
 // struct ggml_sycl_pool_vmm : public ggml_sycl_pool
 /// kernels
-typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 typedef void (*ggml_sycl_op_mul_mat_t)(
     ggml_backend_sycl_context & ctx,
     const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
@@ -1336,83 +1376,6 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
     reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
 }
-template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static void k_get_rows(
-            const void * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12,
-            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
-    const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) +
-                     item_ct1.get_local_id(2)) *
-                    2;
-    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1);
-    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) /
-                    ne12;
-    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) %
-                    ne12;
-    if (i00 >= ne00) {
-        return;
-    }
-    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
-    const int ib = i00/qk; // block index
-    const int iqs = (i00%qk)/qr; // quant index
-    const int iybs = i00 - i00%qk; // dst block start index
-    const int y_offset = qr == 1 ? 1 : qk/2;
-    // dequantize
-    dfloat2 v;
-    dequantize_kernel(src0_row, ib, iqs, v);
-    dst_row[iybs + iqs + 0] = v.x();
-    dst_row[iybs + iqs + y_offset] = v.y();
-}
-template<typename src0_t, typename dst_t>
-static void k_get_rows_float(
-            const src0_t * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12,
-            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
-    const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) +
-                    item_ct1.get_local_id(2);
-    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1);
-    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) /
-                    ne12;
-    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) %
-                    ne12;
-    if (i00 >= ne00) {
-        return;
-    }
-    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
-    dst_row[i00] = src0_row[i00];
-}
 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,
@@ -1523,193 +1486,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     }
 }
-static void cpy_1_f32_f32(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    float * dsti = (float *) cdsti;
-    *dsti = *xi;
-}
-static void cpy_1_f32_f16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    sycl::half *dsti = (sycl::half *)cdsti;
-    *dsti = sycl::vec<float, 1>(*xi)
-                .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
-}
-static void cpy_1_f16_f16(const char * cxi, char * cdsti) {
-    const sycl::half *xi = (const sycl::half *)cxi;
-    sycl::half *dsti = (sycl::half *)cdsti;
-    *dsti = *xi;
-}
-static void cpy_1_f16_f32(const char * cxi, char * cdsti) {
-    const sycl::half *xi = (const sycl::half *)cxi;
-    float * dsti = (float *) cdsti;
-    *dsti = *xi;
-}
-static void cpy_1_i16_i16(const char * cxi, char * cdsti) {
-    const int16_t *xi = (const int16_t *)cxi;
-    int16_t *dsti = (int16_t *)cdsti;
-    *dsti = *xi;
-}
-static void cpy_1_i32_i32(const char * cxi, char * cdsti) {
-    const int32_t *xi = (const int32_t *)cxi;
-    int32_t *dsti = (int32_t *)cdsti;
-    *dsti = *xi;
-}
-template <cpy_kernel_t cpy_1>
-static void cpy_f32_f16(const char * cx, char * cdst, const int ne,
-                        const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-                        const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                        const int nb12, const int nb13, 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);
-    if (i >= ne) {
-        return;
-    }
-    // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
-    // then combine those indices with the corresponding byte offsets to get the total offsets
-    const int i03 = i/(ne00 * ne01 * ne02);
-    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
-    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
-    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
-    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
-    const int i13 = i/(ne10 * ne11 * ne12);
-    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
-    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
-    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
-    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
-    cpy_1(cx + x_offset, cdst + dst_offset);
-}
-static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q8_0 * dsti = (block_q8_0 *) cdsti;
-    float amax = 0.0f; // absolute max
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax = sycl::fmax(amax, sycl::fabs((float)v));
-    }
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-    dsti->d = d;
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j]*id;
-        dsti->qs[j] = sycl::round((float)x0);
-    }
-}
-static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_0 * dsti = (block_q4_0 *) cdsti;
-    float amax = 0.0f;
-    float vmax = 0.0f;
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < sycl::fabs((float)v)) {
-            amax = sycl::fabs((float)v);
-            vmax = v;
-        }
-    }
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f/d : 0.0f;
-    dsti->d = d;
-    for (int j = 0; j < QK4_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK4_0/2 + j]*id;
-        const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f));
-        const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f));
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_1 * dsti = (block_q4_1 *) cdsti;
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-        if (v < vmin) vmin = v;
-        if (v > vmax) vmax = v;
-    }
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-    dsti->dm.x() = d;
-    dsti->dm.y() = vmin;
-    for (int j = 0; j < QK4_1/2; ++j) {
-        const float x0 = (xi[0       + j] - vmin)*id;
-        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
-        const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f));
-        const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f));
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-template <cpy_kernel_t cpy_blck, int qk>
-static void cpy_f32_q(const char * cx, char * cdst, const int ne,
-                      const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-                      const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                      const int nb12, const int nb13, 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)) *
-                  qk;
-    if (i >= ne) {
-        return;
-    }
-    const int i03 = i/(ne00 * ne01 * ne02);
-    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
-    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
-    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
-    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
-    const int i13 = i/(ne10 * ne11 * ne12);
-    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
-    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
-    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
-    const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
-    cpy_blck(cx + x_offset, cdst + dst_offset);
-}
 static void k_sum_rows_f32(const float * x, float * dst, const int ncols,
                            const sycl::nd_item<3> &item_ct1) {
     const int row = item_ct1.get_group(1);
@@ -1895,81 +1671,6 @@ static  void pool2d_nchw_kernel(
         o_ptr[cur_oh * ow + cur_ow] = res;
 }
-template <int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
-                          ggml_tensor *dst, const void *src0_dd,
-                          const int32_t *src1_dd, float *dst_dd,
-                          queue_ptr stream) {
-    GGML_TENSOR_BINARY_OP_LOCALS
-    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
-    const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE);
-    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
-    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
-    GGML_ASSERT(ne00 % 2 == 0);
-    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             k_get_rows<qk, qr, dq>(
-                                 src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
-                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
-                         });
-    GGML_UNUSED(dst);
-    GGML_UNUSED(ctx);
-}
-template <typename src0_t>
-static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                const ggml_tensor *src1, ggml_tensor *dst,
-                                const src0_t *src0_dd, const int32_t *src1_dd,
-                                float *dst_dd, queue_ptr stream) {
-    GGML_TENSOR_BINARY_OP_LOCALS
-    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
-    const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE;
-    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
-    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
-                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
-            });
-    }
-    GGML_UNUSED(dst);
-    GGML_UNUSED(ctx);
-}
 static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
                                    const int ky, const int kx_padded,
                                    queue_ptr stream) {
@@ -1984,7 +1685,7 @@ static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
         stream->parallel_for(
             sycl::nd_range<3>(num_blocks * block_size, block_size),
-            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_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);
             });
     }
@@ -2005,7 +1706,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
         stream->parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x,
                                      nchannels_y, item_ct1);
             });
@@ -2025,7 +1726,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
         stream->parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+            [=](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,
                                        nchannels_y / nchannels_x, item_ct1);
@@ -2033,231 +1734,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
     }
 }
-static void
-ggml_cpy_f16_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00,
-                      const int ne01, const int ne02, const int nb00,
-                      const int nb01, const int nb02, const int nb03,
-                      const int ne10, const int ne11, const int ne12,
-                      const int nb10, const int nb11, const int nb12,
-                      const int nb13, queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_f16_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00,
-                                           nb01, nb02, nb03, ne10, ne11, ne12,
-                                           nb10, nb11, nb12, nb13, item_ct1);
-            });
-    }
-}
-static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne,
-                                  const int ne00, const int ne01,
-                                  const int ne02, const int nb00,
-                                  const int nb01, const int nb02,
-                                  const int nb03, const int ne10,
-                                  const int ne11, const int ne12,
-                                  const int nb10, const int nb11,
-                                  const int nb12, const int nb13,
-                                  queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_f32_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                           item_ct1);
-            });
-    }
-}
-static void ggml_cpy_f32_f16_sycl(const char *cx, char *cdst, const int ne,
-                                  const int ne00, const int ne01,
-                                  const int ne02, const int nb00,
-                                  const int nb01, const int nb02,
-                                  const int nb03, const int ne10,
-                                  const int ne11, const int ne12,
-                                  const int nb10, const int nb11,
-                                  const int nb12, const int nb13,
-                                  queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_f32_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                           item_ct1);
-            });
-    }
-}
-static void ggml_cpy_f32_q8_0_sycl(const char *cx, char *cdst, const int ne,
-                                   const int ne00, const int ne01,
-                                   const int ne02, const int nb00,
-                                   const int nb01, const int nb02,
-                                   const int nb03, const int ne10,
-                                   const int ne11, const int ne12,
-                                   const int nb10, const int nb11,
-                                   const int nb12, const int nb13,
-                                   queue_ptr stream) {
-    GGML_ASSERT(ne % QK8_0 == 0);
-    const int num_blocks = ne / QK8_0;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
-                                           sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(
-                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                 item_ct1);
-                         });
-}
-static void ggml_cpy_f32_q4_0_sycl(const char *cx, char *cdst, const int ne,
-                                   const int ne00, const int ne01,
-                                   const int ne02, const int nb00,
-                                   const int nb01, const int nb02,
-                                   const int nb03, const int ne10,
-                                   const int ne11, const int ne12,
-                                   const int nb10, const int nb11,
-                                   const int nb12, const int nb13,
-                                   queue_ptr stream) {
-    GGML_ASSERT(ne % QK4_0 == 0);
-    const int num_blocks = ne / QK4_0;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
-                                           sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(
-                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                 item_ct1);
-                         });
-}
-static void ggml_cpy_f32_q4_1_sycl(const char *cx, char *cdst, const int ne,
-                                   const int ne00, const int ne01,
-                                   const int ne02, const int nb00,
-                                   const int nb01, const int nb02,
-                                   const int nb03, const int ne10,
-                                   const int ne11, const int ne12,
-                                   const int nb10, const int nb11,
-                                   const int nb12, const int nb13,
-                                   queue_ptr stream) {
-    GGML_ASSERT(ne % QK4_1 == 0);
-    const int num_blocks = ne / QK4_1;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
-                                           sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(
-                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                 item_ct1);
-                         });
-}
-static void ggml_cpy_f16_f16_sycl(const char *cx, char *cdst, const int ne,
-                                  const int ne00, const int ne01,
-                                  const int ne02, const int nb00,
-                                  const int nb01, const int nb02,
-                                  const int nb03, const int ne10,
-                                  const int ne11, const int ne12,
-                                  const int nb10, const int nb11,
-                                  const int nb12, const int nb13,
-                                  queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_f16_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                           item_ct1);
-            });
-    }
-}
-static void ggml_cpy_i16_i16_sycl(const char *cx, char *cdst, const int ne,
-                                  const int ne00, const int ne01,
-                                  const int ne02, const int nb00,
-                                  const int nb01, const int nb02,
-                                  const int nb03, const int ne10,
-                                  const int ne11, const int ne12,
-                                  const int nb10, const int nb11,
-                                  const int nb12, const int nb13,
-                                  queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        // dpct::has_capability_or_fail(stream->get_device(),
-        //                              {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_i16_i16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                           item_ct1);
-            });
-    }
-}
-static void ggml_cpy_i32_i32_sycl(const char *cx, char *cdst, const int ne,
-                                  const int ne00, const int ne01,
-                                  const int ne02, const int nb00,
-                                  const int nb01, const int nb02,
-                                  const int nb03, const int ne10,
-                                  const int ne11, const int ne12,
-                                  const int nb10, const int nb11,
-                                  const int nb12, const int nb13,
-                                  queue_ptr stream) {
-    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
-    {
-        // dpct::has_capability_or_fail(stream->get_device(),
-        //                              {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                cpy_f32_f16<cpy_1_i32_i32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
-                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
-                                           item_ct1);
-            });
-    }
-}
 static void scale_f32_sycl(const float *x, float *dst, const float scale,
                            const int k, queue_ptr stream) {
@@ -2290,7 +1767,7 @@ static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols,
     const sycl::range<3> block_nums(1, nrows, 1);
     stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
                          [=](sycl::nd_item<3> item_ct1)
-                             [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                             [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                                  k_sum_rows_f32(x, dst, ncols, item_ct1);
                              });
 }
@@ -2493,52 +1970,6 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
-static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                  const ggml_tensor *src1, ggml_tensor *dst,
-                                  const float *src0_d, const float *src1_d,
-                                  float *dst_d, const queue_ptr &stream) {
-    GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
-    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
-    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
-    const int32_t * src1_i32 = (const int32_t *) src1_d;
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_sycl_float(ctx, src0, src1, dst, (const sycl::half *)src0_d,
-                                src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_sycl_float(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            GGML_ABORT("fatal error");
-            break;
-    }
-}
 static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
                                 const float *src0_d, const float *src1_d,
@@ -2588,11 +2019,10 @@ inline void ggml_sycl_op_mul_mat_sycl(
     if ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         use_fp16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1] &&
         dst->op_params[0] == GGML_PREC_DEFAULT) {
         // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp16 path\n");
         ggml_sycl_pool_alloc<sycl::half> src0_as_f16(ctx.pool());
         if (src0->type != GGML_TYPE_F16) {
-            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type);
+            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type, dst);
             GGML_ASSERT(to_fp16_sycl != nullptr);
             size_t ne = row_diff*ne00;
             src0_as_f16.alloc(ne);
@@ -2604,7 +2034,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
         ggml_sycl_pool_alloc<sycl::half> src1_as_f16(ctx.pool());
         if (src1->type != GGML_TYPE_F16) {
-            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
             GGML_ASSERT(to_fp16_sycl != nullptr);
             size_t ne = src1_ncols*ne10;
             src1_as_f16.alloc(ne);
@@ -2625,13 +2055,13 @@ inline void ggml_sycl_op_mul_mat_sycl(
             src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
             dst_f16.get(), dpct::library_data_t::real_half, ldc,
             dpct::library_data_t::real_half)));
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
         to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
 #else
         auto dnnl_stream = ctx.stream_dnnl(stream);
         DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
             src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
         to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
 #endif
     }
@@ -2640,13 +2070,13 @@ inline void ggml_sycl_op_mul_mat_sycl(
         ggml_sycl_pool_alloc<float> src0_ddq_as_f32(ctx.pool());
         ggml_sycl_pool_alloc<float> src1_ddq_as_f32(ctx.pool());
         if (src0->type != GGML_TYPE_F32) {
-            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type);
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type, dst);
             GGML_ASSERT(to_fp32_sycl != nullptr);
             src0_ddq_as_f32.alloc(row_diff*ne00);
             to_fp32_sycl(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
         }
         if (src1->type != GGML_TYPE_F32) {
-            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type);
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type, dst);
             GGML_ASSERT(to_fp32_sycl != nullptr);
             src1_ddq_as_f32.alloc(src1_ncols*ne10);
             to_fp32_sycl(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
@@ -3084,7 +2514,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
     for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
         const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_SYCL_MAX_STREAMS : 0;
         const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
         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;
@@ -3270,6 +2699,12 @@ static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
+static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_l2_norm);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
 static void ggml_sycl_group_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_SYCL_DEBUG("call %s\n", __func__);
     ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_group_norm);
@@ -3392,7 +2827,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
     // convert src1 to fp16
     ggml_sycl_pool_alloc<sycl::half> src1_f16_alloc(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
         GGML_ASSERT(to_fp16_sycl != nullptr);
@@ -3488,7 +2923,7 @@ inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     return false;
 }
-bool ggml_sycl_supports_dmmv(enum ggml_type type) {
+static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -3508,6 +2943,7 @@ bool ggml_sycl_supports_dmmv(enum ggml_type type) {
 }
 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
     int64_t min_compute_capability = INT_MAX;
@@ -3569,6 +3005,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
+        // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream());
     } else if (use_mul_mat_vec_q) {
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
     } else if (use_mul_mat_q) {
@@ -3701,8 +3138,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
                 const int64_t i2 = i12;
             src0_row.data = src0_original + i02*nb02;
-            src1_row.data = src1_original + + i11*nb11 + i12*nb12;
-            dst_row.data = dst_original + i1*nb1   + i2*nb2;
+            src1_row.data = src1_original + i11*nb11 + i12*nb12;
+            dst_row.data = dst_original + i1*nb1 + i2*nb2;
             ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
             }
@@ -3821,58 +3258,6 @@ static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_clamp);
 }
-static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
-                          ggml_tensor *dst) try {
-    const int64_t ne = ggml_nelements(src0);
-    GGML_ASSERT(ne == ggml_nelements(src1));
-    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
-    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
-    GGML_TENSOR_BINARY_OP_LOCALS01;
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    queue_ptr main_stream = ctx.stream();
-    char * src0_ddc = (char *) src0->data;
-    char * src1_ddc = (char *) src1->data;
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) {
-        ggml_cpy_i16_i16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
-        ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
-    } else {
-        GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__,
-                ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ABORT("fatal error");
-    }
-    GGML_UNUSED(dst);
-}
-catch (sycl::exception const &exc) {
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-            << ", line:" << __LINE__ << std::endl;
-  std::exit(1);
-}
-static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    // TODO: why do we pass dst as src1 here?
-    ggml_sycl_cpy(ctx, dst->src[0], dst, nullptr);
-}
 static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf);
 }
@@ -3911,7 +3296,7 @@ static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
 }
-void ggml_sycl_set_main_device(const int main_device) try {
+static void ggml_sycl_set_main_device(const int main_device) try {
     if (dpct::get_current_device_id() == static_cast<unsigned int> (main_device)) {
         return;
     }
@@ -3932,7 +3317,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
-bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) {
+static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) {
     if (!g_sycl_loaded) return false;
     if (dst->src[0] != nullptr && ggml_backend_buffer_is_sycl_split(dst->src[0]->buffer)) {
@@ -4034,6 +3419,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
         case GGML_OP_RMS_NORM:
             ggml_sycl_rms_norm(ctx, dst);
             break;
+        case GGML_OP_L2_NORM:
+            ggml_sycl_l2_norm(ctx, dst);
+            break;
         case GGML_OP_MUL_MAT:
             if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
                 return false;
@@ -4069,7 +3457,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
             ggml_sycl_clamp(ctx, dst);
             break;
         case GGML_OP_CPY:
-            ggml_sycl_cpy(ctx, dst->src[0], dst->src[1], dst);
+            ggml_sycl_cpy(ctx, dst->src[0], dst->src[1]);
             break;
         case GGML_OP_CONT:
             ggml_sycl_dup(ctx, dst);
@@ -4111,6 +3499,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
         case GGML_OP_RWKV_WKV6:
             ggml_sycl_op_rwkv_wkv6(ctx, dst);
             break;
+        case GGML_OP_RWKV_WKV7:
+            ggml_sycl_op_rwkv_wkv7(ctx, dst);
+            break;
         case GGML_OP_GATED_LINEAR_ATTN:
             ggml_sycl_op_gated_linear_attn(ctx, dst);
             break;
@@ -4250,10 +3641,71 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
-static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
-    ggml_sycl_set_main_device(sycl_ctx->device);
+static void reorder_qw(char *data_device, const int ncols, const int nrows,
+                size_t size, size_t offset, dpct::queue_ptr stream) {
+    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
+    SYCL_CHECK(
+        CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
+            .wait()));
+    GGML_ASSERT((size % sizeof(block_q4_0) == 0));
+    GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
+    int offset_blks = offset / sizeof(block_q4_0);
+    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
+    auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
+    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;
+            const int ib = i;
+            for (int j = 0; j < QK4_0/2; j ++)
+            {
+                *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
+            }
+            *(d_ptr + ib) = x[ib].d;
+        });
+    sycl::free(tmp_buf, *stream);
+}
+static void reorder_qw(ggml_tensor * src0, dpct::queue_ptr stream) {
+    char*data_device = (char*)src0->data;
+    size_t ncols = src0->ne[0];
+    size_t nrows = src0->ne[1];
+    size_t size = ggml_nbytes(src0);
+    reorder_qw(data_device, ncols, nrows, size, 0, stream);
+}
+static void opt_for_reorder(ggml_tensor * dst, dpct::queue_ptr stream) {
+    ggml_tensor *src0 = dst->src[0];
+    ggml_tensor *src1 = dst->src[1];
+    if (dst->op == GGML_OP_MUL_MAT && src0->type == GGML_TYPE_Q4_0 &&
+        src1->ne[2]==1 && src1->ne[3]==1) {
+        reorder_qw(src0, stream);
+        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
+        GGML_ASSERT(extra);
+        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
+    }
+}
+static void optimize_graph_once(ggml_cgraph * cgraph, ggml_backend_sycl_context * ctx) {
+    dpct::queue_ptr stream = ctx->stream();
+    if (ctx->optimized_graph) {
+       return;
+    }
+    ctx->optimized_graph = true;
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        if (ctx->opt_feature.reorder) opt_for_reorder(cgraph->nodes[i], stream);
+    }
+}
+static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * sycl_ctx, ggml_cgraph * cgraph) {
+    ggml_sycl_set_main_device(sycl_ctx->device);
+    if (!g_ggml_sycl_disable_optimize) optimize_graph_once(cgraph, sycl_ctx);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -4274,7 +3726,46 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
         }
         GGML_ASSERT(ok);
     }
+}
+static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    auto * sycl_ctx = static_cast<ggml_backend_sycl_context *>(backend->context);
+#ifdef GGML_SYCL_GRAPH
+    if (!g_ggml_sycl_disable_graph) {
+        if (!sycl_ctx->exec_graph && !dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_graph)) {
+            GGML_SYCL_DEBUG("[SYCL-GRAPH] can not use graphs on device:%d\n", sycl_ctx->device);
+            ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
+            return GGML_STATUS_SUCCESS;
+        }
+        sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream()));
+        model_sycl_graph.begin_recording(*(sycl_ctx->stream()));
+        ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
+        model_sycl_graph.end_recording();
+        if (!sycl_ctx->exec_graph) {
+            auto exec_graph = model_sycl_graph.finalize({sycl_ex::property::graph::updatable{}});
+            sycl_ctx->exec_graph = std::make_unique<
+                sycl_ex::command_graph<sycl_ex::graph_state::executable>>(exec_graph);
+        } else {
+            try {
+                sycl_ctx->exec_graph->update(model_sycl_graph);
+                GGML_SYCL_DEBUG("[SYCL-GRAPH] update success\n");
+            } catch (sycl::exception const & e) {
+                GGML_SYCL_DEBUG("[SYCL-GRAPH] Exception when updating graph, %s\n", e.what());
+                auto exec_graph = model_sycl_graph.finalize({sycl_ex::property::graph::updatable{}});
+                sycl_ctx->exec_graph = std::make_unique<
+                    sycl_ex::command_graph<sycl_ex::graph_state::executable>>(exec_graph);
+            }
+        }
+        sycl_ctx->stream()->ext_oneapi_graph(*(sycl_ctx->exec_graph));
+    } else
+#endif
+    {
+        ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
+    }
     return GGML_STATUS_SUCCESS;
 }
@@ -4339,7 +3830,6 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) {
 }
 int ggml_backend_sycl_get_device_count() {
-    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n");
     return ggml_sycl_info().device_count;
 }
@@ -4429,7 +3919,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     return true;
                 }
                 return false;
-            } break;
+            }
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_NEG:
@@ -4443,11 +3933,10 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
-                    return ggml_is_contiguous(op->src[0]);
+                    return ggml_is_contiguous(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32);
                 default:
                     return false;
             }
-            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -4478,7 +3967,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     return false;
                 }
                 return true;
-            } break;
+            }
         case GGML_OP_OUT_PROD:
             return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
         case GGML_OP_GET_ROWS:
@@ -4495,7 +3984,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -4521,13 +4010,37 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
+                    return true;
+                }
                 return false;
-            } break;
+            }
         case GGML_OP_CONCAT:
             {
                 ggml_type src0_type = op->src[0]->type;
                 return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
-            } break;
+            }
         case GGML_OP_DUP:
         case GGML_OP_ARGMAX:
         case GGML_OP_NONE:
@@ -4536,23 +4049,25 @@ 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_LOG:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
-            return true;
-        case GGML_OP_NORM:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_GROUP_NORM:
-            return ggml_is_contiguous(op->src[0]);
-        case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+        case GGML_OP_LOG:
+            return (op->src[0]->type == GGML_TYPE_F32);
+        case GGML_OP_NORM:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_L2_NORM:
+        case GGML_OP_GROUP_NORM:
+            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;
@@ -4583,6 +4098,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         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;
         default: