@fugood/llama.node 0.3.2 → 0.3.4

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

@@ -0,0 +1,84 @@
+if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
+    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
+endif()
+
+check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
+
+if (DEFINED ENV{ONEAPI_ROOT})
+    message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
+elseif(SUPPORTS_SYCL)
+    message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
+        If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
+        source /opt/intel/oneapi/setvars.sh")
+else()
+    message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+endif()
+message(STATUS "SYCL found")
+#todo: AOT
+
+ggml_add_backend_library(ggml-sycl
+                         ggml-sycl.cpp
+                         ../../include/ggml-sycl.h
+                        )
+
+if (GGML_SYCL_F16)
+    if (GGML_SYCL_TARGET STREQUAL "AMD")
+        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
+    endif()
+    add_compile_definitions(GGML_SYCL_F16)
+endif()
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+
+if (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()
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+endif()
+
+file(GLOB   GGML_HEADERS_SYCL "*.hpp")
+file(GLOB   GGML_SOURCES_SYCL "*.cpp")
+target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
+
+find_package(DNNL)
+message("-- DNNL found:" ${DNNL_FOUND})
+
+if (GGML_SYCL_TARGET STREQUAL "INTEL")
+    add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
+else()
+    add_compile_definitions(GGML_SYCL_DNNL=0)
+endif()
+
+if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+    target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+endif()
+
+if (WIN32)
+    find_package(IntelSYCL REQUIRED)
+    find_package(MKL REQUIRED)
+    target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+else()
+    if (GGML_SYCL_TARGET STREQUAL "INTEL")
+        target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+    elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+        add_compile_definitions(GGML_SYCL_NVIDIA)
+        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas)
+    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+        if (NOT GGML_SYCL_DEVICE_ARCH)
+            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
+        endif()
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
+        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
+    endif()
+
+    if (GGML_SYCL_DEVICE_ARCH)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
+  endif()
+endif()

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

@@ -26,5 +26,8 @@
 #include "softmax.hpp"
 #include "tsembd.hpp"
 #include "im2col.hpp"
+#include "wkv6.hpp"
+#include "outprod.hpp"
+#include "element_wise.hpp"
 
 #endif // GGML_SYCL_BACKEND_HPP

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

@@ -11,6 +11,7 @@
 //
 
 #include "common.hpp"
+#include "ggml-impl.h"
 
 int get_current_device_id() {
   return dpct::dev_mgr::instance().current_device_id();
@@ -28,11 +29,7 @@ void* ggml_sycl_host_malloc(size_t size) try {
 
   if (err != 0) {
     // clear the error
-    fprintf(
-        stderr,
-        "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
-        size / 1024.0 / 1024.0,
-        "syclGetErrorString is not supported");
+    GGML_LOG_ERROR("WARNING: failed to allocate %.2f MB of pinned memory: %s\n", size / 1024.0 / 1024.0,    "syclGetErrorString is not supported");
     return nullptr;
   }
 
@@ -62,3 +59,37 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
   }
   return sycl_down_blk_size;
 }
+
+void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const ggml_sycl_op_flatten_t op) try {
+
+    const bool use_src1 = src1 != nullptr;
+
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+
+    // dd = data device
+    float * src0_ddf = (float *) src0->data;
+    float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
+    float *  dst_ddf = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> src0_f(ctx.pool());
+    ggml_sycl_pool_alloc<float> src1_f(ctx.pool());
+    ggml_sycl_pool_alloc<float>  dst_f(ctx.pool());
+
+    ggml_sycl_set_device(ctx.device);
+    queue_ptr main_stream = ctx.stream();
+    // GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n",
+        // ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device);
+
+    // do the computation
+    op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
+    // print_ggml_tensor("tensor", dst);
+}
+catch (sycl::exception const &exc) {
+
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}

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

@@ -404,4 +404,263 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
 
 int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
 
+typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                       const ggml_tensor *src1,
+                                       ggml_tensor *dst, const float *src0_dd,
+                                       const float *src1_dd, float *dst_dd,
+                                       const queue_ptr &main_stream);
+
+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 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*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    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 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 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;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    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 <typename src0_t, typename src1_t, typename dst_t>
+    void operator()(ggml_backend_sycl_context & ctx,
+                    const struct ggml_tensor *src0,
+                    const struct ggml_tensor *src1, struct ggml_tensor *dst,
+                    const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd,
+                    queue_ptr stream) {
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        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 cne0[] = {ne0, ne1, ne2, ne3};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+        size_t cnb0[] = {nb0, nb1, nb2, nb3};
+        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];
+        };
+
+        for (int i = 0; i < 4; i++) {
+            if (nr[i] != 1) {
+                break;
+            }
+            if (i > 0) {
+                collapse_nb(cnb0, cne0);
+                collapse_nb(cnb1, cne1);
+                collapse(cne0);
+                collapse(cne1);
+            }
+        }
+        {
+            int64_t ne0 = cne0[0];
+            int64_t ne1 = cne0[1];
+            int64_t ne2 = cne0[2];
+            int64_t ne3 = cne0[3];
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb0[0];
+            size_t nb1 = cnb0[1];
+            size_t nb2 = cnb0[2];
+            size_t nb3 = 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);
+
+            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, 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, s11, s12, s13,
+                                            item_ct1);
+                    });
+            }
+        }
+        GGML_UNUSED(ctx);
+    }
+};
+
+template <class op>
+inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                   const ggml_tensor *src1, ggml_tensor *dst,
+                                   const float *src0_dd, const float *src1_dd,
+                                   float *dst_dd,
+                                   const queue_ptr &main_stream) {
+
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd,
+             (sycl::half *)dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd,
+             main_stream);
+    } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+        op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd,
+             main_stream);
+    } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
+        op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd,
+             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));
+        GGML_ABORT("fatal error");
+    }
+}
+
+
+void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const ggml_sycl_op_flatten_t op);
+
 #endif // GGML_SYCL_COMMON_HPP

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

@@ -47,7 +47,7 @@ static void concat_f32_dim1(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(1) < ne01) { // src0
+  if (item_ct1.get_group(1) < (size_t) ne01) { // src0
     int offset_src =
         nidx + item_ct1.get_group(1) * ne0 + item_ct1.get_group(0) * ne0 * ne01;
     dst[offset_dst] = x[offset_src];
@@ -70,7 +70,7 @@ static void concat_f32_dim2(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(0) < ne02) { // src0
+  if (item_ct1.get_group(0) < (size_t) ne02) { // src0
     int offset_src = nidx + item_ct1.get_group(1) * ne0 +
                      item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
     dst[offset_dst] = x[offset_src];
@@ -106,6 +106,7 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst,
           concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
         });
     break;
+  // dim >=2 will be dispatched to the default path
   default:
     stream->parallel_for(
         sycl::nd_range<3>(gridDim *

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

@@ -424,7 +424,7 @@ static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y,
     const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
-    const src_t * x = (src_t *) vx;
+    const src_t * x = (const src_t *) vx;
     for (int64_t i = global_id; i < k; i += work_group_size * item_ct1.get_group_range(2)) {
         y[i] = x[i];
     }

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

@@ -1015,9 +1015,9 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             break;
     }
 
-    (void) src1;
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_ncols;
-    (void) src1_padded_row_size;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
 }

package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp

@@ -15,6 +15,7 @@
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
+#include <syclcompat/math.hpp>
 #include <oneapi/mkl.hpp>
 #include <map>
 
@@ -1236,7 +1237,7 @@ namespace dpct
 
             std::map<byte_t *, allocation>::iterator get_map_iterator(const void *ptr)
             {
-                auto it = m_map.upper_bound((byte_t *)ptr);
+                auto it = m_map.upper_bound(const_cast<byte_t *>(reinterpret_cast<const byte_t *>(ptr)));
                 if (it == m_map.end())
                 {
                     // Not a virtual pointer.
@@ -1688,9 +1689,14 @@ namespace dpct
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
-            oneapi::mkl::blas::column_major::gemm(
-                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
-                data_b, ldb, beta_value, data_c, ldc);
+#ifdef GGML_SYCL_NVIDIA
+            oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
+                                                  a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
+                                                  beta_value, data_c, ldc);
+#else
+            oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
+                                                  beta_value, data_c, ldc);
+#endif
         }
 
         template <typename VecT, class BinaryOperation, class = void>
@@ -1753,14 +1759,22 @@ namespace dpct
             matrix_info->ld_info[2] = ldc;
             matrix_info->groupsize_info = batch_size;
 
+#ifdef GGML_SYCL_NVIDIA
+            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
+                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, matrix_info->transpose_info,
+                matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1,
+                matrix_info->size_info + 2, matrix_info->value_info, reinterpret_cast<const Ta **>(a),
+                matrix_info->ld_info, reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
+                matrix_info->value_info + 1, reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1,
+                &(matrix_info->groupsize_info));
+#else
             sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
-                q, matrix_info->transpose_info, matrix_info->transpose_info + 1,
-                matrix_info->size_info, matrix_info->size_info + 1,
-                matrix_info->size_info + 2, matrix_info->value_info,
-                reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
-                reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
-                matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
+                q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
+                matrix_info->size_info + 1, matrix_info->size_info + 2, matrix_info->value_info,
+                reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
+                matrix_info->ld_info + 1, matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
                 matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
+#endif
 
             q.submit([&](sycl::handler &cgh)
                      {
@@ -1782,10 +1796,16 @@ namespace dpct
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
+#ifdef GGML_SYCL_NVIDIA
             oneapi::mkl::blas::column_major::gemm_batch(
-                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
-                stride_a, data_b, ldb, stride_b, beta_value,
-                data_c, ldc, stride_c, batch_size);
+                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, a_trans, b_trans, m, n, k,
+                alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c,
+                batch_size);
+#else
+            oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
+                                                        stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc,
+                                                        stride_c, batch_size);
+#endif
         }
 
     } // namespace detail
@@ -1830,31 +1850,10 @@ namespace dpct
                                            : id);
     }
 
-    template <typename T>
-    sycl::vec<T, 4> extract_and_sign_or_zero_extend4(T val)
-    {
-        return sycl::vec<T, 1>(val)
-            .template as<sycl::vec<
-                std::conditional_t<std::is_signed_v<T>, int8_t, uint8_t>, 4>>()
-            .template convert<T>();
-    }
-
-    template <typename T1, typename T2>
-    using dot_product_acc_t =
-        std::conditional_t<std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,
-                           uint32_t, int32_t>;
-
     template <typename T1, typename T2, typename T3>
     inline auto dp4a(T1 a, T2 b, T3 c)
     {
-        dot_product_acc_t<T1, T2> res = c;
-        auto va = extract_and_sign_or_zero_extend4(a);
-        auto vb = extract_and_sign_or_zero_extend4(b);
-        res += va[0] * vb[0];
-        res += va[1] * vb[1];
-        res += va[2] * vb[2];
-        res += va[3] * vb[3];
-        return res;
+        return syclcompat::dp4a(a, b, c);
     }
 
     struct sub_sat