@fugood/llama.node 0.3.17 → 0.4.0

package/src/llama.cpp/ggml/src/ggml-quants.c

@@ -19,12 +19,6 @@
 #define GROUP_MAX_EPS_IQ1_M 1e-7f
 #define GROUP_MAX_EPS_IQ1_S 1e-12f
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid warnings for hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-#endif
-
 #define UNUSED GGML_UNUSED
 
 // reference implementation for deterministic creation of model files

package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -151,6 +151,12 @@ struct rpc_msg_buffer_clear_req {
     uint8_t value;
 };
 
+struct rpc_msg_set_tensor_hash_req {
+    rpc_tensor tensor;
+    uint64_t offset;
+    uint64_t hash;
+};
+
 struct rpc_msg_set_tensor_hash_rsp {
     uint8_t result;
 };
@@ -518,6 +524,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
     result.view_offs = tensor->view_offs;
     result.data = reinterpret_cast<uint64_t>(tensor->data);
+
+    // Avoid sending uninitialized data over the wire
+    memset(result.name, 0, sizeof(result.name));
+    memset(result.padding, 0, sizeof(result.padding));
+
     snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
     return result;
 }
@@ -543,15 +554,12 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_tensor rpc_tensor = serialize_tensor(tensor);
     if (size > HASH_THRESHOLD) {
-        // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes)
-        size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t);
-        std::vector<uint8_t> input(input_size, 0);
-        uint64_t hash = fnv_hash((const uint8_t*)data, size);
-        memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
-        memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
-        memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash));
+        rpc_msg_set_tensor_hash_req request;
+        request.tensor = rpc_tensor;
+        request.offset = offset;
+        request.hash = fnv_hash((const uint8_t*)data, size);
         rpc_msg_set_tensor_hash_rsp response;
-        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response));
+        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
         GGML_ASSERT(status);
         if (response.result) {
             // the server has the same data, no need to send it
@@ -859,7 +867,7 @@ public:
     bool free_buffer(const rpc_msg_free_buffer_req & request);
     bool buffer_clear(const rpc_msg_buffer_clear_req & request);
     bool set_tensor(const std::vector<uint8_t> & input);
-    bool set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response);
+    bool set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response);
     bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
     bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
     bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
@@ -1096,18 +1104,10 @@ bool rpc_server::get_cached_file(uint64_t hash, std::vector<uint8_t> & data) {
     return true;
 }
 
-bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response)
+bool rpc_server::set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response)
 {
-    // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) |
-    if (input.size() != sizeof(rpc_tensor) + 16) {
-        return false;
-    }
-    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
-    uint64_t offset;
-    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset));
     std::vector<uint8_t> cached_file;
-    if (!get_cached_file(*hash, cached_file)) {
+    if (!get_cached_file(request.hash, cached_file)) {
         response.result = 0;
         return true;
     }
@@ -1120,25 +1120,28 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
     ggml_context_ptr ctx_ptr { ggml_init(params) };
     GGML_ASSERT(ctx_ptr != nullptr);
     ggml_context * ctx = ctx_ptr.get();
-    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
+    ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
         return false;
     }
-    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash);
+    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n",
+        __func__, (void*)tensor->buffer, tensor->data, request.offset, size, request.hash);
 
     // sanitize tensor->data
     {
         const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
-        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+        if (request.tensor.data + request.offset < p0
+         || request.tensor.data + request.offset >= p1
+         || size > (p1 - request.tensor.data - request.offset)) {
             GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu, hash=0x%" PRIx64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
-                           __func__, in_tensor->data, offset, size, *hash, p0, p1);
+                           __func__, request.tensor.data, request.offset, size, request.hash, p0, p1);
             return false;
         }
     }
-    ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
+    ggml_backend_tensor_set(tensor, cached_file.data(), request.offset, size);
     response.result = 1;
     return true;
 }
@@ -1498,12 +1501,12 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                 break;
             }
             case RPC_CMD_SET_TENSOR_HASH: {
-                std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
+                rpc_msg_set_tensor_hash_req request;
+                if (!recv_msg(sockfd, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_set_tensor_hash_rsp response;
-                if (!server.set_tensor_hash(input, response)) {
+                if (!server.set_tensor_hash(request, response)) {
                     return;
                 }
                 if (!send_msg(sockfd, &response, sizeof(response))) {
@@ -1589,6 +1592,14 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
 void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
                                    const char * cache_dir,
                                    size_t free_mem, size_t total_mem) {
+    printf("Starting RPC server v%d.%d.%d\n",
+        RPC_PROTO_MAJOR_VERSION,
+        RPC_PROTO_MINOR_VERSION,
+        RPC_PROTO_PATCH_VERSION);
+    printf("  endpoint       : %s\n", endpoint);
+    printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
+    printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));
+
     std::string host;
     int port;
     if (!parse_endpoint(endpoint, host, port)) {
@@ -1748,6 +1759,9 @@ static void * ggml_backend_rpc_get_proc_address(ggml_backend_reg_t reg, const ch
     if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) {
         return (void *)ggml_backend_rpc_add_device;
     }
+    if (std::strcmp(name, "ggml_backend_rpc_start_server") == 0) {
+        return (void *)ggml_backend_rpc_start_server;
+    }
     return NULL;
 
     GGML_UNUSED(reg);

package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt

@@ -49,35 +49,38 @@ endif()
 target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
 
 # Link against oneDNN
-find_package(DNNL)
 set(GGML_SYCL_DNNL 0)
-if(DNNL_FOUND)
-    if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR)
-        # Assuming oneDNN packaged with oneapi release is used which
-        # supports only intel target
-        set(DNNL_GPU_VENDOR "INTEL")
-        if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
-            message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+if(GGML_SYCL_DNN)
+    find_package(DNNL)
+    if(DNNL_FOUND)
+        if (NOT DEFINED DNNL_GPU_VENDOR)
+            # default to intel target
+            set(DNNL_GPU_VENDOR "INTEL")
+            if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
+                message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+            endif()
         endif()
-    endif()
 
-    # Verify oneDNN was compiled for the same target as llama
-    if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
-        target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
-        set(GGML_SYCL_DNNL 1)
-        get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
-        foreach(CONFIG ${CONFIGS})
-            get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
-            message(STATUS "Found oneDNN: ${DNNL_LIB}")
-        endforeach()
+        # Verify oneDNN was compiled for the same target as llama
+        if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
+            target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+            set(GGML_SYCL_DNNL 1)
+            get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
+            foreach(CONFIG ${CONFIGS})
+                get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
+                message(STATUS "Found oneDNN: ${DNNL_LIB}")
+            endforeach()
+        else()
+            message(WARNING
+                "oneDNN must be compiled for the same target as llama.cpp.
+                 llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
+                 Disabling oneDNN support.")
+        endif()
     else()
-        message(WARNING
-            "oneDNN must be compiled for the same target as llama.cpp.
-             llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
-             Disabling oneDNN support.")
+        message(STATUS "oneDNN not found, disabling oneDNN support")
     endif()
 else()
-    message(STATUS "oneDNN not found, disabling oneDNN support")
+    message(STATUS "oneDNN support disabled by the user")
 endif()
 target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
 
@@ -108,6 +111,9 @@ endif()
 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
+    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)

package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp

@@ -14,23 +14,24 @@
 #define GGML_SYCL_BACKEND_HPP
 
 #include "binbcast.hpp"
-#include "concat.hpp"
 #include "common.hpp"
+#include "concat.hpp"
 #include "conv.hpp"
 #include "convert.hpp"
+#include "cpy.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"
+#include "element_wise.hpp"
+#include "gla.hpp"
+#include "im2col.hpp"
 #include "mmq.hpp"
 #include "mmvq.hpp"
-#include "rope.hpp"
 #include "norm.hpp"
+#include "outprod.hpp"
+#include "quants.hpp"
+#include "rope.hpp"
 #include "softmax.hpp"
 #include "tsembd.hpp"
-#include "im2col.hpp"
 #include "wkv.hpp"
-#include "outprod.hpp"
-#include "element_wise.hpp"
-#include "cpy.hpp"
-#include "gla.hpp"
 
-#endif // GGML_SYCL_BACKEND_HPP
+#endif  // GGML_SYCL_BACKEND_HPP

package/src/llama.cpp/ggml/src/ggml-sycl/binbcast.cpp

@@ -1,93 +1,74 @@
 #include "binbcast.hpp"
 
+#include <array>
 #include <cstddef>
 #include <cstdint>
 #include <sycl/sycl.hpp>
 
+#include "dpct/helper.hpp"
 #include "ggml.h"
 
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-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,
-        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);
-    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1));
-    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) /
-                   ne3;
-    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) %
-                   ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_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]);
+template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __dpct_inline__ void k_bin_bcast_contiguous(const src0_t * __restrict__ src0, const src1_t * __restrict__ src1,
+                                                   dst_t * dst, std::size_t num_elements, const sycl::nd_item<1> & it) {
+    auto element_id   = it.get_global_id(0);
+    auto global_range = it.get_global_range(0);
+    for (; element_id < num_elements; element_id += global_range) {
+        auto  src0_float_val = sycl::vec(src0[element_id]).template convert<float, sycl::rounding_mode::rte>();
+        auto  src1_float_val = sycl::vec(src1[element_id]).template convert<float, sycl::rounding_mode::rte>();
+        float dst_val        = bin_op(src0_float_val[0], src1_float_val[0]);
+        auto  val_to_store   = sycl::vec(dst_val).template convert<dst_t, sycl::rounding_mode::rte>();
+        dst[element_id]      = val_to_store;
     }
 }
 
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast_unravel(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,
-        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);
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
+template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __dpct_inline__ void k_bin_bcast(const src0_t * __restrict__ src0, const src1_t * __restrict__ 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, std::size_t num_dst_elements,
+                                        const sycl::nd_item<1> & item_ct1) {
+    auto calculate_logical_index =
+        [](const std::array<int, 4> & dims, std::size_t element_id) __attribute__((always_inline))->std::array<int, 4> {
+        std::array<int, 4> logical_index;
+#pragma unroll(4)
+        for (int i = 3; i >= 0; i--) {
+            logical_index[i] = element_id % dims[i];
+            element_id /= dims[i];
+        }
+        return logical_index;
+    };
+
+    auto calculate_index = [](const std::array<int, 4> & dims, const std::array<int, 4> & strides,
+                              const std::array<int, 4> & indices) __attribute__((always_inline))
+                               ->std::size_t {
+        std::size_t index = 0;
+#pragma unroll(4)
+        for (int i = 0; i < 4; i++) {
+            auto index_i = indices[i];
+            if (indices[i] >= dims[i]) {
+                index_i = indices[i] % dims[i];
+            }
+            index += strides[i] * index_i;
+        }
+        return index;
+    };
+
+    auto element_id = item_ct1.get_global_id(0);
+    for (; element_id < num_dst_elements; element_id += item_ct1.get_global_range(0)) {
+        auto  logical_index  = calculate_logical_index({ ne3, ne2, ne1, ne0 }, element_id);
+        auto  src_0_index    = calculate_index({ ne3, ne2, ne1, ne0 }, { s03, s02, s01, s00 }, logical_index);
+        auto  src_1_index    = calculate_index({ ne13, ne12, ne11, ne10 }, { s13, s12, s11, s10 }, logical_index);
+        auto  dst_index      = calculate_index({ ne3, ne2, ne1, ne0 }, { s3, s2, s1, s0 }, logical_index);
+        auto  src0_float_val = sycl::vec(src0[src_0_index]).template convert<float, sycl::rounding_mode::rte>();
+        auto  src1_float_val = sycl::vec(src1[src_1_index]).template convert<float, sycl::rounding_mode::rte>();
+        float dst_val        = bin_op(src0_float_val[0], src1_float_val[0]);
+        auto  val_to_store   = sycl::vec(dst_val).template convert<dst_t, sycl::rounding_mode::rte>();
+        dst[dst_index]       = val_to_store;
     }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    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]);
 }
 
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_sycl {
+template <float (*bin_op)(const float, const float)> struct bin_bcast_sycl {
     template <typename src0_t, typename src1_t, typename dst_t>
     void operator()(const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, const int64_t ne00,
                     const int64_t ne01, const int64_t ne02, const int64_t ne03, const int64_t ne10, const int64_t ne11,
@@ -96,165 +77,73 @@ struct bin_bcast_sycl {
                     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) {
-        int nr0 = ne10 / ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (src0_is_contiguous && src1_is_contiguous && dst_is_contiguous) {
+        auto check_bcast_required = [](const std::array<int64_t, 4> & src_dims,
+                                       const std::array<int64_t, 4> & dst_dims) -> bool {
             for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
+                if (dst_dims[i] > src_dims[i]) {
+                    return true;
                 }
             }
-        }
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            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);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_UNUSED(s00);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            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);
-
-            sycl::range<3> block_dims(1, 1, 1);
-            block_dims[2] = std::min<unsigned int>(hne0, block_size);
-            block_dims[1] = std::min<unsigned int>(
-                ne1, block_size / (unsigned int)block_dims[2]);
-            block_dims[0] = std::min(
-                std::min<unsigned int>(
-                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
-                                   (unsigned int)block_dims[1]),
-                64U);
-
-            sycl::range<3> block_nums(
-                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
-                (ne1 + block_dims[1] - 1) / block_dims[1],
-                (hne0 + block_dims[2] - 1) / block_dims[2]);
-
-            if (block_nums[0] > 65535) {
-                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / 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, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
-                                          sycl::range<3>(1, 1, block_size)),
-                        [=](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);
-                        });
-                }
-            } else {
-                /*
-                DPCT1049:16: The work-group size passed to the SYCL kernel may
-                exceed the limit. To get the device limit, query
-                info::device::max_work_group_size. Adjust the work-group size if
-                needed.
-                */
-                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_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,
-                                            item_ct1);
-                    });
-            }
+            return false;
+        };
+
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+        GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
+
+        GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
+
+        GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
+
+        // dst strides in number of elements
+        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);
+
+        // src1 strides in number of elements
+        size_t s10 = nb10 / sizeof(src0_t);
+        size_t s11 = nb11 / sizeof(src1_t);
+        size_t s12 = nb12 / sizeof(src1_t);
+        size_t s13 = nb13 / sizeof(src1_t);
+
+        // src0 strides in number of elements
+        size_t s00 = nb00 / sizeof(src0_t);
+        size_t s01 = nb01 / sizeof(src0_t);
+        size_t s02 = nb02 / sizeof(src0_t);
+        size_t s03 = nb03 / sizeof(src0_t);
+
+        std::size_t num_dst_elements = static_cast<std::size_t>(ne0) * static_cast<std::size_t>(ne1) *
+                                       static_cast<std::size_t>(ne2) * static_cast<std::size_t>(ne3);
+        std::size_t local_range  = 256;
+        std::size_t global_range = ceil_div(num_dst_elements, local_range) * local_range;
+
+        bool needs_broadcasting = check_bcast_required({ ne00, ne01, ne02, ne03 }, { ne0, ne1, ne2, ne3 }) ||
+                                  check_bcast_required({ ne10, ne11, ne12, ne13 }, { ne0, ne1, ne2, ne3 });
+        bool all_contiguous = src0_is_contiguous && src1_is_contiguous && dst_is_contiguous;
+
+        if (! needs_broadcasting && all_contiguous) {
+            stream->submit([&](sycl::handler & cgh) {
+                cgh.parallel_for(sycl::nd_range<1>({ global_range }, { local_range }), [=](sycl::nd_item<1> it) {
+                    k_bin_bcast_contiguous<bin_op>(src0_dd, src1_dd, dst_dd, num_dst_elements, it);
+                });
+            });
+        } else {
+            stream->submit([&](sycl::handler & cgh) {
+                cgh.parallel_for(sycl::nd_range<1>({ global_range }, { local_range }), [=](sycl::nd_item<1> it) {
+                    k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12, ne13, s0, s1,
+                                        s2, s3, s00, s01, s02, s03, s10, s11, s12, s13, num_dst_elements, it);
+                });
+            });
         }
     }
 };