whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-openvino/utils.h

@@ -0,0 +1,123 @@
+#include "ggml-backend-impl.h"
+#include "ggml-decoder.h"
+#include "ggml-impl.h"
+
+#include <algorithm>
+#include <cstddef>
+#include <memory>
+#include <openvino/runtime/core.hpp>
+#include <openvino/runtime/infer_request.hpp>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+struct graph_key {
+    int n_nodes;
+    std::string first_node_name;
+    std::string last_node_name;
+
+    graph_key(const ggml_cgraph * cgraph) : n_nodes(cgraph->n_nodes) {
+        if (n_nodes > 0) {
+            first_node_name = cgraph->nodes[0]->name;
+            last_node_name = cgraph->nodes[n_nodes - 1]->name;
+        }
+    }
+
+    bool operator==(const graph_key & other) const {
+        return n_nodes == other.n_nodes && first_node_name == other.first_node_name &&
+               last_node_name == other.last_node_name;
+    }
+};
+
+struct graph_key_hash {
+    size_t operator()(const graph_key & key) const {
+        size_t h = std::hash<int>{}(key.n_nodes);
+        if (key.n_nodes > 0) {
+            h ^= std::hash<std::string>{}(key.first_node_name) + 0x9e3779b9 + (h << 6) + (h >> 2);
+            h ^= std::hash<std::string>{}(key.last_node_name) + 0x9e3779b9 + (h << 6) + (h >> 2);
+        }
+        return h;
+    }
+};
+
+struct ov_runtime_context {
+    std::mutex ov_compute_mutex;
+    std::string device;
+    bool stateful;
+    std::unordered_map<graph_key, std::shared_ptr<GgmlOvDecoder>, graph_key_hash> decoder_cache;
+    std::unordered_map<graph_key, std::shared_ptr<ov::InferRequest>, graph_key_hash> infer_request_cache;
+    std::unordered_map<graph_key, std::shared_ptr<ov::InferRequest>, graph_key_hash> infer_request_cache_prefill;
+    std::unordered_map<graph_key, std::vector<std::string>, graph_key_hash> ov_input_names_cache;
+    std::unordered_map<graph_key, std::vector<std::string>, graph_key_hash> ov_output_names_cache;
+    //TODO: Stateful is only supported for single request at a time.
+    //      Simultanous stateful inference request support to be added.
+    size_t stateful_kv_size;
+    std::map<std::string, std::string> kv_state_input_name_map;
+
+    ov_runtime_context() :
+        device("CPU"),
+        stateful(false),
+        stateful_kv_size(0) {}
+};
+
+enum ggml_status ov_graph_compute(struct ggml_cgraph * cgraph, ggml_backend_t backend);
+
+enum ggml_status ov_graph_compute_dynamic(struct ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx);
+enum ggml_status ov_graph_compute_static(struct ggml_cgraph * cgraph, std::shared_ptr<ov_runtime_context> r_ctx);
+
+size_t checksum(const void * data, size_t size);
+
+void print_input_tensor_info(const std::string & name, const ov::Tensor & tensor);
+
+void print_output_tensor_info(const std::string & name, const ov::Tensor & tensor, const void * output_dst);
+
+template <typename T>
+std::vector<T> pad_input(const T * data,
+                         size_t rows,
+                         size_t cols,
+                         size_t padded_rows,
+                         size_t padded_cols,
+                         T pad_value) {
+    std::vector<T> padded(padded_rows * padded_cols, pad_value);
+
+    for (size_t i = 0; i < std::min(rows, padded_rows); ++i) {
+        for (size_t j = 0; j < std::min(cols, padded_cols); ++j) {
+            padded[i * padded_cols + j] = data[i * cols + j];
+        }
+    }
+
+    return padded;
+}
+
+template <typename T>
+std::vector<T> pad_input(const ggml_tensor * tensor, size_t padded_rows, size_t padded_cols, T pad_value) {
+    return pad_input<T>(reinterpret_cast<const T *>(tensor->data),
+                        static_cast<size_t>(tensor->ne[1]),  // rows
+                        static_cast<size_t>(tensor->ne[0]),  // cols
+                        padded_rows, padded_cols, pad_value);
+}
+
+void set_zero_diagonal(std::vector<float> & matrix, size_t rows, size_t cols);
+
+const ggml_tensor * get_inp_pos_tensor(struct ggml_cgraph * cgraph);
+
+bool get_is_prefill(const ggml_tensor * inp_pos);
+
+ov::Tensor get_ov_input_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder, const std::string & param_name);
+ov::Tensor get_ov_input_tensor_static_decode(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
+                                             const std::string & param_name);
+ov::Tensor get_ov_input_tensor_static_prefill(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
+                                              const std::string & param_name,
+                                              int chunk_index);
+
+ov::Tensor create_ov_output_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder,
+                                   std::shared_ptr<ov::InferRequest> infer_request,
+                                   int output_index,
+                                   const ggml_tensor * ggml_tensor);
+
+bool is_naive(struct ggml_cgraph * cgraph);
+
+enum ggml_status naive_compute(struct ggml_cgraph * cgraph,
+                               ov::Core & core,
+                               const std::string & device,
+                               const ov::AnyMap & config);

data/ext/sources/ggml/src/ggml-quants.c

@@ -304,6 +304,41 @@ void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RE
     }
 }
 
+void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_NVFP4;
+    static const int qk_sub = QK_NVFP4_SUB;
+    static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        for (int s = 0; s < n_sub; s++) {
+            const float * xb = x + i*qk + s*qk_sub;
+
+            float amax = 0.0f;
+            for (int j = 0; j < qk_sub; j++) {
+                if (amax < fabsf(xb[j])) {
+                    amax = fabsf(xb[j]);
+                }
+            }
+
+            // UE4M3 scale: amax / 6.0 maps the max E2M1 value (6.0) to amax
+            const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
+            y[i].d[s] = ue;
+            const float d = ggml_ue4m3_to_fp32(ue);
+
+            for (int j = 0; j < qk_sub/2; ++j) {
+                const uint8_t x0 = best_index_mxfp4(xb[0        + j], d);
+                const uint8_t x1 = best_index_mxfp4(xb[qk_sub/2 + j], d);
+
+                y[i].qs[s*(qk_sub/2) + j] = x0 | (x1 << 4);
+            }
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
@@ -434,6 +469,31 @@ void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
+void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_NVFP4;
+    static const int qk_sub = QK_NVFP4_SUB;
+    static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        for (int s = 0; s < n_sub; s++) {
+            const float d = ggml_ue4m3_to_fp32(x[i].d[s]);
+            float * yb = y + i*qk + s*qk_sub;
+
+            for (int j = 0; j < qk_sub/2; ++j) {
+                const int8_t v0 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] & 0x0F];
+                const int8_t v1 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] >>   4];
+
+                yb[j + 0       ] = v0*d;
+                yb[j + qk_sub/2] = v1*d;
+            }
+        }
+    }
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -2098,6 +2158,12 @@ size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
     return nrow * ggml_row_size(GGML_TYPE_MXFP4, n_per_row);
 }
 
+size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_UNUSED(quant_weights);
+    quantize_row_nvfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * ggml_row_size(GGML_TYPE_NVFP4, n_per_row);
+}
+
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
 void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
@@ -3104,6 +3170,11 @@ static void quantize_row_iq2_xxs_impl(const float * GGML_RESTRICT x, void * GGML
             }
             float scale = make_qp_quants(32, kMaxQ+1, xval, (uint8_t*)L, weight);
             float eff_max = scale*kMaxQ;
+            if (eff_max <= 0) {
+                scales[ib] = 0;
+                memset(L, 0, 32);
+                continue;
+            }
             float best = 0;
             for (int is = -6; is <= 6; ++is) {
                 float id = (2*kMaxQ-1+is*0.1f)/eff_max;
@@ -3273,9 +3344,9 @@ static void quantize_row_iq2_xs_impl(const float * GGML_RESTRICT x, void * GGML_
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+            memset(L, 0, 16);
             if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
-                memset(L, 0, 16);
                 continue;
             }
             float best = 0;
@@ -3714,9 +3785,9 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * GGML_RESTRICT
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+            memset(L, 0, 32);
             if (max < GROUP_MAX_EPS_IQ3_XXS) {
                 scales[ib] = 0;
-                memset(L, 0, 32);
                 continue;
             }
             float best = 0;
@@ -3922,6 +3993,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * GGML_RESTRICT
             }
             float max = xval[0];
             for (int i = 1; i < block_size; ++i) max = MAX(max, xval[i]);
+            memset(L, 0, block_size);
             if (!max) {
                 scales[ib] = 0;
                 continue;
@@ -4245,6 +4317,7 @@ static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_R
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
             if (max < GROUP_MAX_EPS_IQ1_S) {
                 scales[ib] = 0;
+                shifts[ib] = 1;
                 memset(L, 1, block_size);
                 continue;
             }
@@ -4285,7 +4358,12 @@ static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_R
                     }
                 }
             }
-            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_shift != 0);
+            if (besti1 < 0 || besti2 < 0 || best_shift == 0) {
+                scales[ib] = 0;
+                shifts[ib] = 1;
+                memset(L, 1, block_size);
+                continue;
+            }
             for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
             for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
             for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
@@ -4429,6 +4507,7 @@ static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_R
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
             if (max < GROUP_MAX_EPS_IQ1_M) {
                 scales[ib] = 0;
+                shifts[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
             }
@@ -4527,7 +4606,12 @@ static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_R
                     }
                 }
             }
-            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_k >= 0);
+            if (besti1 < 0 || besti2 < 0 || best_k < 0) {
+                scales[ib] = 0;
+                shifts[ib] = 0;
+                memset(L, 1, block_size);
+                continue;
+            }
             for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
             for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
             for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
@@ -4683,7 +4767,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
             sumqx += w*q*xb[j];
             sumq2 += w*q*q;
         }
-        d = sumqx/sumq2;
+        d = sumq2 > 0 ? sumqx/sumq2 : 0.f;
         float best = d*sumqx;
         for (int itry = -ntry; itry <= ntry; ++itry) {
             id = (itry + values[0])/max;
@@ -4874,6 +4958,7 @@ static void quantize_row_iq2_s_impl(const float * GGML_RESTRICT x, void * GGML_R
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+            memset(L, 0, 16);
             if (max < GROUP_MAX_EPS_IQ2_S) {
                 scales[ib] = 0;
                 continue;
@@ -5225,6 +5310,12 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_E_E8M0_IMPL(block_mxfp4, data, nb);
             } break;
+        case GGML_TYPE_NVFP4:
+            {
+                // UE4M3 scales are uint8_t — all byte values are valid
+                GGML_UNUSED(data);
+                GGML_UNUSED(nb);
+            } break;
         case GGML_TYPE_Q2_K:
             {
                 VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);

data/ext/sources/ggml/src/ggml-quants.h

@@ -22,6 +22,7 @@ GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 *
 GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
 
 GGML_API void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k);
 
 GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
@@ -48,6 +49,7 @@ GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GG
 //GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 GGML_API void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -95,6 +97,7 @@ GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);

data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt

@@ -1,7 +1,7 @@
 message(STATUS  "GGML_SYCL_TARGET=${GGML_SYCL_TARGET}")
 
-if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
-    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
+if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL)$")
+    message(FATAL_ERROR "GGML_SYCL_TARGET: Invalid target, the supported options are [INTEL]")
 endif()
 
 check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
@@ -25,6 +25,11 @@ ggml_add_backend_library(ggml-sycl
 
 file(GLOB   GGML_HEADERS_SYCL "*.hpp")
 file(GLOB   GGML_SOURCES_SYCL "*.cpp")
+file(GLOB   SRCS "template-instances/fattn-tile*.cpp")
+list(APPEND GGML_SOURCES_SYCL ${SRCS})
+file(GLOB   SRCS "template-instances/fattn-vec*.cpp")
+list(APPEND GGML_SOURCES_SYCL ${SRCS})
+
 target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
 
 if (WIN32)
@@ -125,106 +130,28 @@ endif()
 target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
 
 if (GGML_SYCL_F16)
-    if (GGML_SYCL_TARGET STREQUAL "AMD")
-        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
-    endif()
     add_compile_definitions(GGML_SYCL_F16)
 endif()
 
 if (GGML_SYCL_TARGET STREQUAL "INTEL")
     add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
     target_link_options(ggml-sycl PRIVATE  -Xs   -ze-intel-greater-than-4GB-buffer-required)
-elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-    # INFO: Allowed Sub_group_sizes are not consistent through all
-    # hip targets. For example, 64 is used for certain models, but the backend
-    # does not support it.
-    # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-else()
-    # default for other target
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-endif()
-
-if (GGML_SYCL_GRAPH)
-    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
-endif()
 
-# Link against Intel oneMKL or oneMath
-if (GGML_SYCL_TARGET STREQUAL "INTEL")
-    # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically
-    # See https://github.com/uxlfoundation/oneMath/issues/654
+    # Link against Intel oneMKL
     if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         set(SYCL_COMPILER ON)
     endif()
     find_package(MKL REQUIRED)
     target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
-    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL)
 else()
-    find_package(oneMath QUIET)
-    if (NOT oneMath_FOUND)
-        message(STATUS "oneMath not found: oneMath will be automatically downloaded")
-        # Use FetchContent to automatically pull and build oneMath
-        include(FetchContent)
-        set(BUILD_FUNCTIONAL_TESTS False)
-        set(BUILD_EXAMPLES False)
-        set(TARGET_DOMAINS blas)
-        if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-            set(ENABLE_MKLCPU_BACKEND False)
-            set(ENABLE_MKLGPU_BACKEND False)
-            set(ENABLE_CUBLAS_BACKEND True)
-        elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-            set(ENABLE_MKLCPU_BACKEND False)
-            set(ENABLE_MKLGPU_BACKEND False)
-            set(ENABLE_ROCBLAS_BACKEND True)
-            # Ensure setting a string variable here is not overriden by oneMath CACHE variables
-            cmake_policy(SET CMP0126 NEW)
-            # Setting the device architecture is only needed and useful for AMD devices in oneMath
-            set(HIP_TARGETS ${GGML_SYCL_DEVICE_ARCH} CACHE STRING "oneMath HIP target" FORCE)
-        endif()
-        FetchContent_Declare(
-            ONEMATH
-            GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
-        )
-        FetchContent_MakeAvailable(ONEMATH)
-        # Create alias to match with find_package targets name
-        function(onemath_alias target)
-            if (TARGET ${target}_obj)
-                # Silence verbose warnings from external libraries
-                target_compile_options(${target}_obj PRIVATE -w)
-            endif()
-            if (TARGET ${target})
-                add_library(ONEMATH::${target} ALIAS ${target})
-            endif()
-        endfunction()
-        onemath_alias(onemath)
-        onemath_alias(onemath_blas_mklcpu)
-        onemath_alias(onemath_blas_mklgpu)
-        onemath_alias(onemath_blas_cublas)
-        onemath_alias(onemath_blas_rocblas)
-    endif()
+    # default for other target
+    message(FATAL_ERROR "GGML_SYCL_TARGET is not supported")
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+endif()
 
-    # Below oneMath compile-time dispatching is used for better performance
-    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_cublas)
-        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
-        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
-    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        if (NOT GGML_SYCL_DEVICE_ARCH)
-            message(FATAL_ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
-        endif()
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
-        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
-        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_AMD)
-    else()
-        # Fallback to oneMath runtime dispatcher
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath)
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GENERIC)
-    endif()
+if (GGML_SYCL_GRAPH)
+    message(STATUS "find GGML_SYCL_GRAPH")
+    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
 endif()
 
 if (GGML_SYCL_DEVICE_ARCH)

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

@@ -55,7 +55,11 @@ void ggml_sycl_add_id(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
   const int32_t* src2_d = (const int32_t*)src2->data;
   float* dst_d = (float*)dst->data;
 
-  int threads = std::min((int)ne00, 768);  // cols
+  const unsigned int max_work_group_size = ggml_sycl_info().max_work_group_sizes[ctx.device];
+  assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+
+  int threads = std::min((unsigned int)ne00, max_work_group_size);  // cols
+
   ctx.stream()->parallel_for(
       sycl::nd_range<3>(
           sycl::range<3>(1, ne02, ne01) * sycl::range<3>(1, 1, threads),

data/ext/sources/ggml/src/ggml-sycl/backend.hpp

@@ -23,6 +23,7 @@
 #include "dequantize.hpp"
 #include "dmmv.hpp"
 #include "element_wise.hpp"
+#include "fattn.hpp"
 #include "gla.hpp"
 #include "im2col.hpp"
 #include "mmq.hpp"

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

@@ -11,8 +11,8 @@ static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
+        int s00, int s01, int s02, int s03,
+        int s10, int s11, int s12, int s13,
         const sycl::nd_item<3> &item_ct1) {
     const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                     item_ct1.get_local_id(2);
@@ -44,7 +44,7 @@ static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
     for (int i0 = i0s; i0 < ne0;
          i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
         const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0*s00] : 0.0f, (float)src1_row[i10*s10]);
     }
 }
 
@@ -53,8 +53,8 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
+        int s00, int s01, int s02, int s03,
+        int s10, int s11, int s12, int s13,
         const sycl::nd_item<3> &item_ct1) {
 
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
@@ -82,7 +82,7 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t
     dst_t * dst_row = dst + i_dst;
 
     const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0*s00] : 0.0f, (float)src1_row[i10*s10]);
 }
 
 
@@ -95,7 +95,8 @@ struct bin_bcast_sycl {
                     const int64_t ne3, const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
                     const size_t nb10, const size_t nb11, const size_t nb12, const size_t nb13, const size_t nb0,
                     const size_t nb1, const size_t nb2, const size_t nb3, const bool src0_is_contiguous,
-                    const bool src1_is_contiguous, const bool dst_is_contiguous, queue_ptr stream) {
+                    const bool src1_is_contiguous, const bool src0_is_permuted, const bool src1_is_permuted,
+                    queue_ptr stream) {
         int nr0 = ne10 / ne0;
         int nr1 = ne11/ne1;
         int nr2 = ne12/ne2;
@@ -123,7 +124,7 @@ struct bin_bcast_sycl {
             cnb[3] *= cne[3];
         };
 
-        if (src0_is_contiguous && src1_is_contiguous && dst_is_contiguous) {
+        if (src0_is_contiguous && src1_is_contiguous && !src0_is_permuted && !src1_is_permuted) {
             for (int i = 0; i < 4; i++) {
                 if (nr[i] != 1) {
                     break;
@@ -164,7 +165,7 @@ struct bin_bcast_sycl {
             size_t nb12 = cnb1[2];
             size_t nb13 = cnb1[3];
 
-            size_t s0 = nb0 / sizeof(dst_t);
+            // size_t s0 = nb0 / sizeof(dst_t);
             size_t s1 = nb1 / sizeof(dst_t);
             size_t s2 = nb2 / sizeof(dst_t);
             size_t s3 = nb3 / sizeof(dst_t);
@@ -196,9 +197,6 @@ struct bin_bcast_sycl {
             GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
             GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
             const int block_size = 128;
 
             int64_t hne0 = std::max(ne0/2LL, 1LL);
@@ -232,8 +230,8 @@ struct bin_bcast_sycl {
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
                                 src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
-                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
-                                s03, s11, s12, s13, item_ct1);
+                                ne10, ne11, ne12, ne13, s1, s2, s3, s00, s01, s02,
+                                s03, s10, s11, s12, s13, item_ct1);
                         });
                 }
             } else {
@@ -251,7 +249,7 @@ struct bin_bcast_sycl {
                     [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
-                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
+                                            s1, s2, s3, s00, s01, s02, s03, s10, s11, s12, s13,
                                             item_ct1);
                     });
             }
@@ -268,24 +266,27 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
     if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
         op()((const float *) src0->data, (const float *) src1->data, (float *) dst->data, ne00, ne01, ne02, ne03, ne10,
              ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3,
-             ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
         op()((const sycl::half *) src0->data, (const sycl::half *) src1->data, (sycl::half *) dst->data, ne00, ne01,
              ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13,
-             nb0, nb1, nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst),
+             nb0, nb1, nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
              main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
         op()((const sycl::half *) src0->data, (const float *) src1->data, (sycl::half *) dst->data, ne00, ne01, ne02,
              ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1,
-             nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb2, nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
         op()((const int32_t *) src0->data, (const int32_t *) src1->data, (int32_t *) dst->data, ne00, ne01, ne02, ne03,
              ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
-             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
         op()((const int16_t *) src0->data, (const int16_t *) src1->data, (int16_t *) dst->data, ne00, ne01, ne02, ne03,
              ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2,
-             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_contiguous(dst), main_stream);
+             nb3, ggml_is_contiguous(src0), ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1),
+             main_stream);
     } else {
         fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
                 ggml_type_name(src0->type), ggml_type_name(src1->type));