npm - @fugood/llama.node - Versions diffs - 0.3.3 → 0.3.4 - Mend

@fugood/llama.node 0.3.3 → 0.3.4

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 (225) hide show

package/CMakeLists.txt +5 -0
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-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 +18 -1
package/package.json +1 -1
package/src/EmbeddingWorker.cpp +15 -5
package/src/EmbeddingWorker.h +2 -1
package/src/LlamaCompletionWorker.cpp +1 -1
package/src/LlamaContext.cpp +81 -18
package/src/LlamaContext.h +2 -0
package/src/llama.cpp/.github/workflows/build.yml +197 -159
package/src/llama.cpp/.github/workflows/docker.yml +5 -8
package/src/llama.cpp/.github/workflows/python-lint.yml +8 -1
package/src/llama.cpp/.github/workflows/server.yml +21 -14
package/src/llama.cpp/CMakeLists.txt +11 -6
package/src/llama.cpp/Sources/llama/llama.h +4 -0
package/src/llama.cpp/cmake/common.cmake +33 -0
package/src/llama.cpp/cmake/x64-windows-llvm.cmake +11 -0
package/src/llama.cpp/common/CMakeLists.txt +6 -2
package/src/llama.cpp/common/arg.cpp +426 -245
package/src/llama.cpp/common/common.cpp +143 -80
package/src/llama.cpp/common/common.h +81 -24
package/src/llama.cpp/common/sampling.cpp +53 -19
package/src/llama.cpp/common/sampling.h +22 -1
package/src/llama.cpp/common/speculative.cpp +274 -0
package/src/llama.cpp/common/speculative.h +28 -0
package/src/llama.cpp/docs/build.md +101 -148
package/src/llama.cpp/examples/CMakeLists.txt +32 -13
package/src/llama.cpp/examples/batched/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/batched/batched.cpp +5 -4
package/src/llama.cpp/examples/batched-bench/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/cvector-generator/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp +1 -1
package/src/llama.cpp/examples/embedding/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/eval-callback/CMakeLists.txt +3 -2
package/src/llama.cpp/examples/export-lora/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gbnf-validator/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp +4 -7
package/src/llama.cpp/examples/gen-docs/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf-hash/CMakeLists.txt +8 -1
package/src/llama.cpp/examples/gguf-split/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf-split/gguf-split.cpp +2 -2
package/src/llama.cpp/examples/gritlm/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gritlm/gritlm.cpp +1 -1
package/src/llama.cpp/examples/imatrix/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/imatrix/imatrix.cpp +11 -2
package/src/llama.cpp/examples/infill/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/infill/infill.cpp +1 -1
package/src/llama.cpp/examples/llama-bench/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +405 -316
package/src/llama.cpp/examples/llama.android/llama/build.gradle.kts +1 -0
package/src/llama.cpp/examples/llava/CMakeLists.txt +10 -3
package/src/llama.cpp/examples/llava/clip.cpp +262 -66
package/src/llama.cpp/examples/llava/clip.h +8 -2
package/src/llama.cpp/examples/llava/llava-cli.cpp +1 -1
package/src/llama.cpp/examples/llava/llava.cpp +46 -19
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +1 -1
package/src/llama.cpp/examples/llava/qwen2vl-cli.cpp +581 -0
package/src/llama.cpp/examples/lookahead/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/lookahead/lookahead.cpp +1 -1
package/src/llama.cpp/examples/lookup/CMakeLists.txt +4 -4
package/src/llama.cpp/examples/lookup/lookup-stats.cpp +2 -1
package/src/llama.cpp/examples/lookup/lookup.cpp +2 -2
package/src/llama.cpp/examples/main/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/main/main.cpp +9 -5
package/src/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/parallel/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/parallel/parallel.cpp +1 -1
package/src/llama.cpp/examples/passkey/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/perplexity/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/quantize/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/quantize/quantize.cpp +0 -3
package/src/llama.cpp/examples/quantize-stats/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/retrieval/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/retrieval/retrieval.cpp +4 -4
package/src/llama.cpp/examples/run/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/run/run.cpp +911 -0
package/src/llama.cpp/examples/save-load-state/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +4 -4
package/src/llama.cpp/examples/server/CMakeLists.txt +3 -7
package/src/llama.cpp/examples/server/server.cpp +1758 -886
package/src/llama.cpp/examples/server/tests/requirements.txt +2 -2
package/src/llama.cpp/examples/server/utils.hpp +94 -304
package/src/llama.cpp/examples/simple/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple/simple.cpp +4 -0
package/src/llama.cpp/examples/simple-chat/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +3 -0
package/src/llama.cpp/examples/speculative/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/speculative/speculative.cpp +16 -15
package/src/llama.cpp/examples/speculative-simple/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +265 -0
package/src/llama.cpp/examples/tokenize/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/tokenize/tokenize.cpp +1 -1
package/src/llama.cpp/examples/tts/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/tts/tts.cpp +932 -0
package/src/llama.cpp/ggml/CMakeLists.txt +46 -34
package/src/llama.cpp/ggml/include/ggml-backend.h +16 -0
package/src/llama.cpp/ggml/include/ggml-cpu.h +7 -49
package/src/llama.cpp/ggml/include/ggml-opencl.h +26 -0
package/src/llama.cpp/ggml/include/ggml.h +106 -24
package/src/llama.cpp/ggml/src/CMakeLists.txt +73 -24
package/src/llama.cpp/ggml/src/ggml-alloc.c +0 -1
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +51 -11
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +379 -22
package/src/llama.cpp/ggml/src/ggml-backend.cpp +4 -4
package/src/llama.cpp/ggml/src/ggml-blas/CMakeLists.txt +3 -7
package/src/llama.cpp/ggml/src/ggml-blas/ggml-blas.cpp +5 -2
package/src/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +33 -3
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +456 -111
package/src/llama.cpp/ggml/src/ggml-cann/common.h +6 -3
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +95 -35
package/src/llama.cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt +2 -5
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +22 -9
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp +24 -13
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp +23 -13
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +11 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +10 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +10 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +17 -0
package/src/llama.cpp/ggml/src/ggml-common.h +42 -42
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +288 -213
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp +220 -0
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.h +8 -0
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/common.h +19 -22
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/mmq.cpp +93 -92
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/mmq.h +2 -9
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +323 -0
package/src/llama.cpp/ggml/src/ggml-cpu/{ggml-cpu-aarch64.c → ggml-cpu-aarch64.cpp} +892 -190
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +2 -24
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp +55 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-hbm.h +8 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +15 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +38 -25
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-traits.cpp +36 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-traits.h +38 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +552 -399
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +101 -136
package/src/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +2 -2
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +7 -10
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +8 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +4 -6
package/src/llama.cpp/ggml/src/ggml-impl.h +32 -11
package/src/llama.cpp/ggml/src/ggml-kompute/CMakeLists.txt +13 -9
package/src/llama.cpp/ggml/src/ggml-kompute/ggml-kompute.cpp +131 -64
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +3 -6
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +39 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +14 -7
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +147 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +4004 -0
package/src/llama.cpp/ggml/src/ggml-opt.cpp +67 -80
package/src/llama.cpp/ggml/src/ggml-quants.c +0 -9
package/src/llama.cpp/ggml/src/ggml-rpc/CMakeLists.txt +3 -5
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +5 -2
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +13 -10
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +2 -11
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/concat.cpp +2 -2
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +1 -1
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +5 -5
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +32 -13
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +80 -61
package/src/llama.cpp/ggml/src/ggml-sycl/gemm.hpp +4 -4
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +159 -114
package/src/llama.cpp/ggml/src/ggml-sycl/im2col.cpp +3 -2
package/src/llama.cpp/ggml/src/ggml-sycl/mmq.cpp +6 -6
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +6 -20
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +4 -3
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +8 -8
package/src/llama.cpp/ggml/src/ggml-sycl/rope.cpp +4 -3
package/src/llama.cpp/ggml/src/ggml-sycl/softmax.cpp +7 -7
package/src/llama.cpp/ggml/src/ggml-sycl/tsembd.cpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +4 -1
package/src/llama.cpp/ggml/src/ggml-threading.h +4 -2
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +21 -7
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1718 -399
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +3 -1
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +105 -31
package/src/llama.cpp/ggml/src/ggml.c +367 -207
package/src/llama.cpp/include/llama-cpp.h +25 -0
package/src/llama.cpp/include/llama.h +26 -19
package/src/llama.cpp/models/ggml-vocab-roberta-bpe.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-roberta-bpe.gguf.out +46 -0
package/src/llama.cpp/pocs/CMakeLists.txt +3 -1
package/src/llama.cpp/pocs/vdot/CMakeLists.txt +2 -2
package/src/llama.cpp/src/CMakeLists.txt +2 -7
package/src/llama.cpp/src/llama-grammar.cpp +15 -15
package/src/llama.cpp/src/llama-grammar.h +2 -5
package/src/llama.cpp/src/llama-sampling.cpp +35 -90
package/src/llama.cpp/src/llama-vocab.cpp +6 -1
package/src/llama.cpp/src/llama.cpp +1748 -640
package/src/llama.cpp/src/unicode.cpp +62 -51
package/src/llama.cpp/src/unicode.h +9 -10
package/src/llama.cpp/tests/CMakeLists.txt +48 -37
package/src/llama.cpp/tests/test-arg-parser.cpp +2 -2
package/src/llama.cpp/tests/test-backend-ops.cpp +140 -21
package/src/llama.cpp/tests/test-chat-template.cpp +50 -4
package/src/llama.cpp/tests/test-gguf.cpp +1303 -0
package/src/llama.cpp/tests/test-grammar-integration.cpp +3 -6
package/src/llama.cpp/tests/test-llama-grammar.cpp +2 -4
package/src/llama.cpp/tests/test-quantize-fns.cpp +3 -3
package/src/llama.cpp/tests/test-rope.cpp +61 -20
package/src/llama.cpp/tests/test-sampling.cpp +2 -2
package/src/llama.cpp/.github/workflows/nix-ci-aarch64.yml +0 -72
package/src/llama.cpp/.github/workflows/nix-ci.yml +0 -79
package/src/llama.cpp/.github/workflows/nix-flake-update.yml +0 -22
package/src/llama.cpp/.github/workflows/nix-publish-flake.yml +0 -36
package/src/llama.cpp/ggml/include/ggml-amx.h +0 -25
package/src/llama.cpp/ggml/src/ggml-aarch64.c +0 -129
package/src/llama.cpp/ggml/src/ggml-aarch64.h +0 -19
package/src/llama.cpp/ggml/src/ggml-amx/CMakeLists.txt +0 -107
package/src/llama.cpp/ggml/src/ggml-amx/ggml-amx.cpp +0 -446

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

@@ -47,7 +47,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
     info.device_count = dpct::dev_mgr::instance().device_count();
     if (info.device_count == 0) {
-        fprintf(stderr, "%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__);
+        GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
         return info;
     }
@@ -55,16 +55,16 @@ static ggml_sycl_device_info ggml_sycl_init() {
     int64_t total_vram = 0;
 #if defined(GGML_SYCL_FORCE_MMQ)
-    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   yes\n", __func__);
+    GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ:   yes\n", __func__);
 #else
-    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   no\n", __func__);
+    GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ:   no\n", __func__);
 #endif
 #if defined(SYCL_USE_XMX)
-    fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__);
+    GGML_LOG_INFO("%s: SYCL_USE_XMX: yes\n", __func__);
 #else
-    fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
+    GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__);
 #endif
-    fprintf(stderr, "%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count);
+    GGML_LOG_INFO("%s: found %d %s devices:\n", __func__, info.device_count, GGML_SYCL_NAME);
     for (int i = 0; i < info.device_count; ++i) {
         info.devices[i].vmm = 0;
@@ -110,7 +110,7 @@ void print_device_detail(int id, sycl::device &device, std::string device_type)
     auto global_mem_size = prop.get_global_mem_size()/1000000;
-    fprintf(stderr, "|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
+    GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
             name.c_str(), version.c_str(), prop.get_max_compute_units(),
             prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
             global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
@@ -120,19 +120,29 @@ 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();
     std::map<std::string, size_t> DeviceNums;
-    fprintf(stderr, "found %d SYCL devices:\n", device_count);
-    fprintf(stderr, "|  |                   |                                       |       |Max    |        |Max  |Global |                     |\n");
-    fprintf(stderr, "|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |\n");
-    fprintf(stderr, "|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|\n");
-    fprintf(stderr, "|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|\n");
+    GGML_LOG_INFO("Found %d SYCL devices:\n", device_count);
+    GGML_LOG_INFO(
+        "|  |                   |                                       |      "
+        " |Max    |        |Max  |Global |                     |\n");
+    GGML_LOG_INFO(
+        "|  |                   |                                       |      "
+        " |compute|Max work|sub  |mem    |                     |\n");
+    GGML_LOG_INFO(
+        "|ID|        Device Type|                                   "
+        "Name|Version|units  |group   |group|size   |       Driver version|\n");
+    GGML_LOG_INFO(
+        "|--|-------------------|---------------------------------------|------"
+        "-|-------|--------|-----|-------|---------------------|\n");
     for (int id = 0; id < device_count; ++id) {
-        sycl::device device = dpct::dev_mgr::instance().get_device(id);
-        sycl::backend backend = device.get_backend();
-        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) << "]";
-        print_device_detail(id, device, device_type.str());
+      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)
+                  << "]";
+      print_device_detail(id, device, device_type.str());
     }
 }
@@ -154,15 +164,14 @@ static void ggml_check_sycl() try {
     static bool initialized = false;
     if (!initialized) {
-        fprintf(stderr, "[SYCL] call ggml_check_sycl\n");
+        GGML_LOG_INFO("[SYCL] call ggml_check_sycl\n");
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
+        GGML_LOG_INFO("%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
 #if defined(GGML_SYCL_F16)
-        fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__);
+        GGML_LOG_INFO("%s: GGML_SYCL_F16: yes\n", __func__);
 #else
-        fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
+        GGML_LOG_INFO("%s: GGML_SYCL_F16: no\n", __func__);
 #endif
 /* NOT REMOVE, keep it for next optimize for XMX.
@@ -180,9 +189,10 @@ static void ggml_check_sycl() try {
             return;
         }
         GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
-        ggml_backend_sycl_print_sycl_devices();
         initialized = true;
         g_sycl_loaded = true;
+        ggml_backend_sycl_print_sycl_devices();
     }
 }
 catch (sycl::exception const &exc) {
@@ -205,7 +215,7 @@ inline void check_allow_gpu_index(const int device_index) {
         __func__,
         device_index,
         ggml_sycl_info().device_count - 1);
-    fprintf(stderr, "%s\n", error_buf);
+    GGML_LOG_ERROR("%s\n", error_buf);
     assert(false);
   }
 }
@@ -409,14 +419,12 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
         return true;
     }
     return false;
-}
-catch (sycl::exception const &exc) {
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-            << ", line:" << __LINE__ << std::endl;
-  std::exit(1);
+    GGML_UNUSED(buffer);
+} catch (const sycl::exception & exc) {
+    std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+    std::exit(1);
 }
 static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
                                            uint8_t value) try {
      ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@@ -475,8 +483,8 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
     SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device(
                                     size, *stream)));
     if (!dev_ptr) {
-        fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size);
-        return nullptr;
+      GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+      return nullptr;
     }
     ggml_backend_sycl_buffer_context * ctx = new  ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream);
     return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size);
@@ -752,7 +760,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                         size, *stream)));
         if (!buf) {
             char err_buf[1024];
-            snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+            snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
             throw std::runtime_error(err_buf);
         }
         // set padding to 0 to avoid possible NaN values
@@ -1081,10 +1089,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
     ggml_sycl_buffer buffer_pool[MAX_SYCL_BUFFERS] = {};
     size_t pool_size = 0;
-    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) :
-        qptr(qptr_),
-        device(device_) {
-    }
+    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) : device(device_), qptr(qptr_) {}
     ~ggml_sycl_pool_leg() {
         for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
@@ -1142,17 +1147,18 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
             CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device(
                                 look_ahead_size, *qptr)));
         if (!ptr) {
-            fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size);
+            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
             return nullptr;
         }
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
-    #ifdef DEBUG_SYCL_MALLOC
-        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
+#ifdef DEBUG_SYCL_MALLOC
+        GGML_LOG_DEBUG("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
                 (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
-    #endif
+#endif
         // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
         return ptr;
     }
@@ -1166,7 +1172,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
                 return;
             }
         }
-        fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n");
+        GGML_LOG_WARN("WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n");
         SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
         pool_size -= size;
     }
@@ -1226,7 +1232,7 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
         zeros[i] = 0.f;
         qzeros[i] = 0;
     }
-    const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros;
+    const TC xi = ix < kx ? *(const TC *)&x[iy * kx + ix] : zeros;
     float sum = xi[0];
     float amax = sycl::fabs(xi[0]);
 #pragma unroll
@@ -1787,6 +1793,9 @@ static  void pool2d_nchw_kernel(
         switch (op) {
             case GGML_OP_POOL_AVG: res = 0; break;
             case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+            default:
+                res      = (To) sycl::nan(uint32_t(0));
+                break;
         }
         for (int i = bh; i < eh; i += 1) {
@@ -1805,6 +1814,9 @@ static  void pool2d_nchw_kernel(
                 switch (op) {
                     case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
                     case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
+                    default:
+                        res = (To) sycl::nan(uint32_t(0));
+                        break;
                 }
             }
         }
@@ -1843,7 +1855,8 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
                          });
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 template <typename src0_t>
@@ -1881,10 +1894,10 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
             });
     }
-    (void) dst;
+    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) {
@@ -2437,7 +2450,7 @@ static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_te
             break;
         default:
             // TODO: k-quants
-            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
             GGML_ABORT("fatal error");
             break;
     }
@@ -2452,8 +2465,8 @@ static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tens
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
-    (void) src1;
-    (void) src1_d;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src1_d);
 }
@@ -2472,17 +2485,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
     const int64_t ne00 = src0->ne[0];
     const int64_t ne10 = src1->ne[0];
-    const int64_t ne0 = dst->ne[0];
     const int64_t row_diff = row_high - row_low;
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
+#if !GGML_SYCL_DNNL
+    const int64_t ne0 = dst->ne[0];
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
     int ldc = id == ctx.device ? ne0 : row_diff;
+#endif
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -2519,9 +2533,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
         ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
-        const sycl::half alpha_f16 = 1.0f;
-        const sycl::half beta_f16 = 0.0f;
 #if !GGML_SYCL_DNNL
+        const sycl::half alpha_f16 = 1.0f;
+        const sycl::half beta_f16  = 0.0f;
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
             *stream, oneapi::mkl::transpose::trans,
             oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
@@ -2558,24 +2572,29 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
 #if !GGML_SYCL_DNNL
+        const float alpha = 1.0f;
+        const float beta  = 0.0f;
+#    ifdef GGML_SYCL_NVIDIA
         SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
-            *stream, oneapi::mkl::transpose::trans,
-            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
-            dpct::get_value(&alpha, *stream), src0_ddf_i, ne00,
-            src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
+            oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream }, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i,
+            ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
+#    else
+        SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+            *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
+            dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
             dst_dd_i, ldc)));
+#    endif
 #else
         auto dnnl_stream = ctx.stream_dnnl(stream);
          DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
             src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
 #endif
     }
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_padded_row_size;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2621,8 +2640,9 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
                                item_ct1);
         });
-    (void) src1;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2637,9 +2657,10 @@ inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor
     sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2656,9 +2677,10 @@ inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_te
     sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2677,9 +2699,10 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_ten
     argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2696,9 +2719,10 @@ inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tens
     argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2718,9 +2742,10 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const gg
     diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2741,9 +2766,10 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2766,9 +2792,10 @@ inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
@@ -2845,7 +2872,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
     ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
     const bool src1_is_contiguous = ggml_is_contiguous(src1);
@@ -3272,7 +3298,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
     GGML_TENSOR_BINARY_OP_LOCALS
-    const int64_t ne_dst = ggml_nelements(dst);
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();;
@@ -3380,6 +3405,7 @@ catch (sycl::exception const &exc) {
 inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     // TODO: accuracy issues in MMQ
+    GGML_UNUSED(type);
     return false;
 }
@@ -3447,8 +3473,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
     if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // KQ single-batch
-        ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
+        // TODO: Refactor and cleanup of mul mat dispatching.
+        if (src0->ne[3] == 1 && src1->ne[3] == 1) {
+            // KQ single-batch
+            // mmv p021 was specific for these dimensions
+            ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
+        } else {
+            // The kernel from the if path is faster for that specific case, but does not support all mul mats.
+            ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
+        }
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
@@ -3743,12 +3776,12 @@ static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
     } 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 {
-        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+        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");
     }
-    (void) dst;
+    GGML_UNUSED(dst);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -3759,7 +3792,7 @@ catch (sycl::exception const &exc) {
 static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     // TODO: why do we pass dst as src1 here?
     ggml_sycl_cpy(ctx, src0, dst, nullptr);
-    (void) src1;
+    GGML_UNUSED(src1);
 }
 static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3804,13 +3837,16 @@ static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor
 }
 static void ggml_sycl_nop(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    (void) src0;
-    (void) src1;
-    (void) dst;
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 void ggml_sycl_set_main_device(const int main_device) try {
-    if (dpct::get_current_device_id() == main_device) return;
+    if (dpct::get_current_device_id() == static_cast<unsigned int> (main_device)) {
+        return;
+    }
     check_allow_gpu_index(main_device);
     dpct::select_device(main_device);
@@ -3818,7 +3854,7 @@ void ggml_sycl_set_main_device(const int main_device) try {
         dpct::device_info prop;
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
             prop, dpct::dev_mgr::instance().get_device(main_device))));
-        fprintf(stderr, "Using device %d (%s) as main device\n",
+        GGML_LOG_INFO("Using device %d (%s) as main device\n",
                 main_device, prop.get_name());
     }
 }
@@ -4165,7 +4201,7 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
 #endif
         bool ok = ggml_sycl_compute_forward(*sycl_ctx, node);
         if (!ok) {
-            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
         }
         GGML_ASSERT(ok);
     }
@@ -4178,6 +4214,7 @@ try
 {
     ggml_backend_sycl_context *sycl_ctx =
         (ggml_backend_sycl_context *)backend->context;
     sycl::event *sycl_event = static_cast<sycl::event *>(event->context);
     const queue_ptr &stream = sycl_ctx->stream(sycl_ctx->device, 0);
@@ -4192,7 +4229,7 @@ catch (sycl::exception const &exc)
 }
 static void ggml_backend_sycl_event_wait(ggml_backend_t backend, ggml_backend_event_t event) try {
-    ggml_backend_sycl_context* sycl_ctx = static_cast<ggml_backend_sycl_context*>(backend->context);
     sycl::event* sycl_event = static_cast<sycl::event*>(event->context);
     if (ggml_backend_is_sycl(backend)) {
@@ -4350,10 +4387,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (op->op == GGML_OP_MUL_MAT) {
                     a = op->src[0];
                     b = op->src[1];
-                    if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
-                        // TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
-                        return false;
-                    }
                 } else {
                     a = op->src[2];
                     b = op->src[1];
@@ -4455,7 +4488,16 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
-            return ggml_is_contiguous(op->src[0]);
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return ggml_is_contiguous(op->src[0]);
+            }
         case GGML_OP_IM2COL:
             // TODO: add support for the new F32 operations
             return op->src[0]->type == GGML_TYPE_F16;
@@ -4490,7 +4532,7 @@ static bool ggml_backend_sycl_device_supports_buft(ggml_backend_dev_t dev, ggml_
 static int64_t get_op_batch_size(const ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_GET_ROWS:
-            return op->ne[1]; // this will increse the speed of prefill in test
+            return 0;
         case GGML_OP_MUL_MAT:
             return op->ne[1];
         case GGML_OP_MUL_MAT_ID:
@@ -4604,13 +4646,14 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"
     // "ggml_backend_unregister_host_buffer"
+    GGML_UNUSED(name);
     return nullptr;
 }
 static const ggml_backend_reg_i ggml_backend_sycl_reg_interface = {
     /* .get_name          = */ ggml_backend_sycl_reg_get_name,
     /* .get_device_count  = */ ggml_backend_sycl_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_sycl_reg_get_device,
+    /* .get_device        = */ ggml_backend_sycl_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_sycl_reg_get_proc_address,
 };
@@ -4641,16 +4684,17 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                 dev_ctx->description = prop.get_name();
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_sycl_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface       = */ ggml_backend_sycl_device_interface,
+                    /* .reg         = */ &reg,
+                    /* .context     = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_sycl_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_sycl_reg_interface,
+                /* .context     = */ ctx
             };
         }
@@ -4668,7 +4712,7 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context(device);
     if (ctx == nullptr) {
-        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return nullptr;
     };
@@ -4682,3 +4726,4 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     return sycl_backend;
 }
+GGML_BACKEND_DL_IMPL(ggml_backend_sycl_reg)

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

@@ -120,6 +120,7 @@ void ggml_sycl_op_im2col(
         im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
     }
-    (void) src0;
-    (void) src0_dd;
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src0_dd);
+    GGML_UNUSED(ctx);
 }