@fugood/llama.node 0.3.3 → 0.3.5

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

@@ -14,51 +14,51 @@
 #include <vector>
 
 struct ggml_opt_dataset {
-    struct ggml_context   * ctx;
-    ggml_backend_buffer_t   buf;
-    struct ggml_tensor    * data;
-    struct ggml_tensor    * labels;
+    struct ggml_context   * ctx    = nullptr;
+    ggml_backend_buffer_t   buf    = nullptr;
+    struct ggml_tensor    * data   = nullptr;
+    struct ggml_tensor    * labels = nullptr;
 
-    int64_t ndata;
-    int64_t ndata_shard;
-    size_t  nbs_data;
-    size_t  nbs_labels;
+    int64_t ndata       = -1;
+    int64_t ndata_shard = -1;
+    size_t  nbs_data    = -1;
+    size_t  nbs_labels  = -1;
 
     std::vector<int64_t> permutation;
 };
 
 struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched;
-    ggml_cgraph           * allocated_graph;
-    ggml_cgraph           * allocated_graph_copy;
-    struct ggml_context   * ctx_static;
-    struct ggml_context   * ctx_static_cpu;
-    struct ggml_context   * ctx_compute;
-    struct ggml_context   * ctx_copy;
-    ggml_backend_buffer_t   buf_static;
-    ggml_backend_buffer_t   buf_static_cpu;
+    ggml_backend_sched_t    backend_sched        = nullptr;
+    ggml_cgraph           * allocated_graph      = nullptr;
+    ggml_cgraph           * allocated_graph_copy = nullptr;
+    struct ggml_context   * ctx_static           = nullptr;
+    struct ggml_context   * ctx_static_cpu       = nullptr;
+    struct ggml_context   * ctx_compute          = nullptr;
+    struct ggml_context   * ctx_copy             = nullptr;
+    ggml_backend_buffer_t   buf_static           = nullptr;
+    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
     std::mt19937            rng;
 
-    struct ggml_tensor * inputs;
-    struct ggml_tensor * outputs;
-    struct ggml_tensor * labels;
+    struct ggml_tensor * inputs  = nullptr;
+    struct ggml_tensor * outputs = nullptr;
+    struct ggml_tensor * labels  = nullptr;
 
-    struct ggml_tensor * loss;
-    struct ggml_tensor * pred;
-    struct ggml_tensor * ncorrect;
+    struct ggml_tensor * loss     = nullptr;
+    struct ggml_tensor * pred     = nullptr;
+    struct ggml_tensor * ncorrect = nullptr;
 
-    struct ggml_cgraph * gf;
-    struct ggml_cgraph * gb_grad;
-    struct ggml_cgraph * gb_opt;
+    struct ggml_cgraph * gf      = nullptr;
+    struct ggml_cgraph * gb_grad = nullptr;
+    struct ggml_cgraph * gb_opt  = nullptr;
 
-    int64_t iter;
-    int32_t opt_period;
-    int32_t opt_i;
-    bool    loss_per_datapoint;
+    int64_t iter               = 1;
+    int32_t opt_period         = 1;
+    int32_t opt_i              = 0;
+    bool    loss_per_datapoint = false;
 
-    ggml_opt_get_optimizer_params get_opt_pars;
-    void * get_opt_pars_ud;
-    struct ggml_tensor * adamw_params;
+    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
+    void * get_opt_pars_ud                     = nullptr;
+    struct ggml_tensor * adamw_params          = nullptr;
 };
 
 struct ggml_opt_result {
@@ -67,8 +67,8 @@ struct ggml_opt_result {
     std::vector<int32_t> pred;
     int64_t              ncorrect = 0;
 
-    bool loss_per_datapoint = false;
-    int64_t opt_period = -1;
+    int64_t opt_period         = -1;
+    bool    loss_per_datapoint = false;
 };
 
 // ====== Dataset ======
@@ -188,11 +188,11 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
 }
 
 struct ggml_opt_params ggml_opt_default_params(
-        ggml_backend_sched_t backend_sched,
-        struct ggml_context * ctx_compute,
-        struct ggml_tensor * inputs,
-        struct ggml_tensor * outputs,
-        enum ggml_opt_loss_type loss_type) {
+        ggml_backend_sched_t      backend_sched,
+        struct ggml_context     * ctx_compute,
+        struct ggml_tensor      * inputs,
+        struct ggml_tensor      * outputs,
+        enum ggml_opt_loss_type   loss_type) {
     return {
         /*backend_sched   =*/ backend_sched,
         /*ctx_compute     =*/ ctx_compute,
@@ -237,25 +237,33 @@ static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_
     return new_tensor;
 }
 
-static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) {
+static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
     std::map<ggml_tensor *, ggml_tensor *> tensor_map;
 
-    ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true);
+    ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true);
 
-    for (int i = 0; i < graph->n_leafs; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i]));
+    for (int i = 0; i < src->n_leafs; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i]));
     }
-    for (int i = 0; i < graph->n_nodes; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i]));
+    GGML_ASSERT(dst->n_leafs == src->n_leafs);
+    for (int i = 0; i < src->n_nodes; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i]));
     }
-    for (int i = 0; i < graph->n_nodes; ++i) {
-        const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]);
-        const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]);
-        graph->grads[igrad_dst]     = new_graph->grads[igrad_src];
-        graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
+    GGML_ASSERT(dst->n_nodes == src->n_nodes);
+    for (int i = 0; i < src->n_nodes; ++i) {
+        const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
+        const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+
+        GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
+        GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
+
+        dst->grads[igrad_dst]     = src->grads[igrad_src];
+        dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
     }
 
-    return new_graph;
+    return dst;
 }
 
 static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
@@ -284,18 +292,13 @@ static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph
 
 ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched        = params.backend_sched;
-    result->allocated_graph      = nullptr;
-    result->allocated_graph_copy = nullptr;
-    result->ctx_compute          = params.ctx_compute;
-    result->ctx_copy             = nullptr;
-    result->inputs               = params.inputs;
-    result->outputs              = params.outputs;
-    result->iter                 = 1;
-    result->opt_period           = params.opt_period;
-    result->opt_i                = 0;
-    result->get_opt_pars         = params.get_opt_pars;
-    result->get_opt_pars_ud      = params.get_opt_pars_ud;
+    result->backend_sched   = params.backend_sched;
+    result->ctx_compute     = params.ctx_compute;
+    result->inputs          = params.inputs;
+    result->outputs         = params.outputs;
+    result->opt_period      = params.opt_period;
+    result->get_opt_pars    = params.get_opt_pars;
+    result->get_opt_pars_ud = params.get_opt_pars_ud;
 
     GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
     GGML_ASSERT(result->opt_period >= 1);
@@ -348,7 +351,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
 
     switch (params.loss_type) {
         case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->labels = nullptr;
             result->loss = ggml_sum(result->ctx_static, result->outputs);
             ggml_set_name(result->loss, "loss_sum");
             const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
@@ -358,7 +360,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
             break;
         }
         case GGML_OPT_LOSS_TYPE_SUM: {
-            result->labels = nullptr;
             result->loss = ggml_sum(result->ctx_static, result->outputs);
             ggml_set_name(result->loss, "loss_sum");
             result->loss_per_datapoint = false;
@@ -413,14 +414,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     }
 
     if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->gb_grad = nullptr;
-        result->gb_opt  = nullptr;
-
         result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gf);
-
         return result;
     }
 
@@ -429,14 +423,8 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
 
     if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->gb_opt  = nullptr;
-
         result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gb_grad);
         ggml_graph_reset(result->gb_grad);
-
         return result;
     }
 
@@ -466,7 +454,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
 
     result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
 
-    ggml_opt_alloc_graph(result, result->gb_opt);
     ggml_graph_reset(result->gb_opt);
 
     return result;

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

@@ -5220,15 +5220,6 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
             } break;
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-            {
-                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
-            } break;
-        case GGML_TYPE_Q4_0_8_8:
-            {
-                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
-            } break;
 
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:

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

@@ -1,10 +1,8 @@
 message(STATUS "Using RPC backend")
 
-add_library(ggml-rpc
-            ggml-rpc.cpp)
-
-target_link_libraries(ggml-rpc PRIVATE ggml-base)
-target_include_directories(ggml-rpc PRIVATE . ..)
+ggml_add_backend_library(ggml-rpc
+                         ggml-rpc.cpp
+                        )
 
 if (WIN32)
     target_link_libraries(ggml-rpc PRIVATE ws2_32)

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

@@ -1369,8 +1369,9 @@ static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
 
 ggml_backend_reg_t ggml_backend_rpc_reg(void) {
     static struct ggml_backend_reg ggml_backend_rpc_reg = {
-        /* .iface   = */ ggml_backend_rpc_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_rpc_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_rpc_reg;
@@ -1401,3 +1402,5 @@ ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
 
     return dev;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_rpc_reg)

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

@@ -16,12 +16,10 @@ endif()
 message(STATUS "SYCL found")
 #todo: AOT
 
-add_library(ggml-sycl
-            ggml-sycl.cpp
-            ../../include/ggml-sycl.h)
-
-target_link_libraries(ggml-sycl PRIVATE ggml-base)
-target_include_directories(ggml-sycl PRIVATE . ..)
+ggml_add_backend_library(ggml-sycl
+                         ggml-sycl.cpp
+                         ../../include/ggml-sycl.h
+                        )
 
 if (GGML_SYCL_F16)
     if (GGML_SYCL_TARGET STREQUAL "AMD")
@@ -70,12 +68,17 @@ else()
         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 "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
+        add_compile_definitions(GGML_SYCL_NVIDIA)
+        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas)
     elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        if (GGML_SYCL_HIP_TARGET STREQUAL "")
-            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_HIP_TARGET has not been set.")
+        if (NOT GGML_SYCL_DEVICE_ARCH)
+            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
         endif()
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=${GGML_SYCL_HIP_TARGET}")
+        set(CMAKE_CXX_FLAGS "${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/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;
   }
 
@@ -66,18 +63,12 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
 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 int64_t nrows0 = ggml_nrows(src0);
 
     const bool use_src1 = src1 != nullptr;
-    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
 
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
 
-    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
-
     // dd = data device
     float * src0_ddf = (float *) src0->data;
     float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;

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

@@ -626,6 +626,7 @@ struct bin_bcast_sycl {
                     });
             }
         }
+        GGML_UNUSED(ctx);
     }
 };
 

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];

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

@@ -1237,7 +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.
@@ -1689,9 +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>
@@ -1754,14 +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)
                      {
@@ -1783,10 +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