whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-cuda/binbcast.cu

@@ -39,13 +39,16 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
                                    const uint3            ne11,
                                    const uint3            ne12,
                                    const uint3            ne13,
-                                   /*int s0, */ const int s1,
+                                 /*const int              s0,*/
+                                   const int              s1,
                                    const int              s2,
                                    const int              s3,
-                                   /*int s00,*/ const int s01,
+                                   const int              s00,
+                                   const int              s01,
                                    const int              s02,
                                    const int              s03,
-                                   /*int s10,*/ const int s11,
+                                   const int              s10,
+                                   const int              s11,
                                    const int              s12,
                                    const int              s13,
                                    src1_ptrs... src1s) {
@@ -72,11 +75,11 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
     for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
         const uint32_t i10 = fastmodulo(i0, ne10);
 
-        float result = src0_row ? (float) src0_row[i0] : 0.0f;
+        float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
         if constexpr (sizeof...(src1_ptrs) > 0) {
-            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
         } else {
-            result = bin_op(result, (float)src1[i_src1 + i10]);
+            result = bin_op(result, (float)src1[i_src1 + i10*s10]);
         }
 
         dst_row[i0] = (dst_t) result;
@@ -101,13 +104,16 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
                                            const uint3            ne11,
                                            const uint3            ne12,
                                            const uint3            ne13,
-                                           /*int s0, */ const int s1,
+                                         /*const int              s0,*/
+                                           const int              s1,
                                            const int              s2,
                                            const int              s3,
-                                           /*int s00,*/ const int s01,
+                                           const int              s00,
+                                           const int              s01,
                                            const int              s02,
                                            const int              s03,
-                                           /*int s10,*/ const int s11,
+                                           const int              s10,
+                                           const int              s11,
                                            const int              s12,
                                            const int              s13,
                                            src1_ptrs... src1s) {
@@ -135,11 +141,11 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
 
     const int i10 = fastmodulo(i0, ne10);
 
-    float result = src0_row ? (float) src0_row[i0] : 0.0f;
+    float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
     if constexpr (sizeof...(src1_ptrs) > 0) {
-        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
     } else {
-        result = bin_op(result, (float)src1[i_src1 + i10]);
+        result = bin_op(result, (float)src1[i_src1 + i10*s10]);
     }
 
     dst_row[i0] = (dst_t) result;
@@ -179,7 +185,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         cnb[3] *= cne[3];
     };
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
         for (int i = 0; i < 4; i++) {
             if (nr[i] != 1) {
                 break;
@@ -221,7 +227,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         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);
@@ -251,10 +257,6 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
         GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
-        GGML_ASSERT(s0 == 1);
-        GGML_ASSERT(s00 == 1);
-        GGML_ASSERT(s10 == 1);
-
         const int block_size = 128;
 
         int64_t hne0 = std::max(ne0 / 2LL, 1LL);
@@ -284,31 +286,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
                 k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
                     ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
                 k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
                     <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
                                                            ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
-                                                           /* s0, */ s1, s2, s3,
-                                                           /* s00,*/ s01, s02, s03,
-                                                           /* s10,*/ s11, s12, s13);
+                                                         /*s0,*/ s1,  s2,  s3,
+                                                           s00, s01, s02, s03,
+                                                           s10, s11, s12, s13);
             }
         } else {
             const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
             if constexpr (sizeof...(I) > 0) {
                 k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1, s2,  s3,
+                    s00 ,s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
                 k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13);
             }
         }
     }

data/ext/sources/ggml/src/ggml-cuda/common.cuh

@@ -53,6 +53,7 @@
 // While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see
 // https://docs.nvidia.com/cutlass/media/docs/cpp/blackwell_functionality.html#blackwell-sm120-gemms
 #define GGML_CUDA_CC_BLACKWELL       1200
+#define GGML_CUDA_CC_DGX_SPARK       1210
 #define GGML_CUDA_CC_RUBIN           1300
 #define GGML_CUDA_CC_OFFSET_AMD      0x1000000
 #define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000
@@ -262,6 +263,10 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
+#if defined(TURING_MMA_AVAILABLE)
+#define LDMATRIX_TRANS_AVAILABLE
+#endif // defined(TURING_MMA_AVAILABLE)
+
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
@@ -526,6 +531,86 @@ static __device__ __forceinline__ half2 warp_prefix_inclusive_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+enum class block_reduce_method {
+    MAX,
+    SUM,
+};
+
+template<block_reduce_method method_t, typename T>
+struct block_reduce_policy;
+
+template <typename T, typename... Ts>
+inline constexpr bool is_any = (std::is_same_v<T, Ts> || ...);
+
+template<typename...>
+inline constexpr bool ggml_cuda_dependent_false_v = false;
+
+template <typename T> struct block_reduce_policy<block_reduce_method::SUM, T> {
+    static __device__ T reduce(T val) {
+        if constexpr(is_any<T, float, float2, half2, int>) {
+            return warp_reduce_sum(val);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce sum");
+        }
+    }
+
+    static __device__ T sentinel() {
+        if constexpr (std::is_same_v<T, float>) {
+            return 0.0f;
+        } else if constexpr (std::is_same_v<T, float2>) {
+            return make_float2(0.0f, 0.0f);
+        } else if constexpr (std::is_same_v<T, half2>) {
+            return make_half2(0.0f, 0.0f);
+        } else if constexpr (std::is_same_v<T, int>) {
+            return 0;
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce sum");
+        }
+    }
+};
+
+template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
+    static __device__ T reduce(T val) {
+        if constexpr (is_any<T, float, half2>) {
+            return warp_reduce_max(val);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce max");
+        }
+    }
+
+    static __device__ T sentinel() {
+        if constexpr (std::is_same_v<T, float>) {
+            return -INFINITY;
+        } else if constexpr (std::is_same_v<T, half2>) {
+            return make_half2(-INFINITY, -INFINITY);
+        } else {
+            static_assert(ggml_cuda_dependent_false_v<T>, "Unsupported type for block reduce max");
+        }
+    }
+};
+
+template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
+static __device__ T block_reduce(T val, T * shared_vals) {
+    val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
+    const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+    if (block_size > WARP_SIZE) {
+        assert((block_size <= 1024) && (block_size % WARP_SIZE) == 0);
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            shared_vals[warp_id] = val;
+        }
+        __syncthreads();
+        val = block_reduce_policy<reduce_method_t, T>::sentinel();
+        if (lane_id < (static_cast<int>(block_size) / WARP_SIZE)) {
+            val = shared_vals[lane_id];
+        }
+        return block_reduce_policy<reduce_method_t, T>::reduce(val);
+    }
+
+    return val;
+}
+
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
@@ -1037,14 +1122,18 @@ struct ggml_tensor_extra_gpu {
 #endif
 
 struct ggml_cuda_graph_node_properties {
-    void * node_address;
+    void * node_data;
     ggml_op node_op;
+    enum ggml_type node_type;
+    int32_t flags;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
-    void * src_address[GGML_MAX_SRC];
+    void * src_data[GGML_MAX_SRC];
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 
+static_assert(std::is_trivial<ggml_cuda_graph_node_properties>::value, "ggml_cuda_graph_node_properties must be trivial");
+
 struct ggml_cuda_graph {
 #ifdef USE_CUDA_GRAPH
     ~ggml_cuda_graph() {
@@ -1060,25 +1149,18 @@ struct ggml_cuda_graph {
     size_t num_nodes = 0;
     std::vector<cudaGraphNode_t> nodes;
     bool disable_due_to_gpu_arch = false;
-    bool disable_due_to_too_many_updates = false;
-    int number_consecutive_updates = 0;
+    bool warmup_complete = false;
     std::vector<ggml_cuda_graph_node_properties> props;
 
-    void record_update(bool use_graph, bool update_required) {
-        if (use_graph && update_required) {
-            number_consecutive_updates++;
-        } else {
-            number_consecutive_updates = 0;
-        }
-        if (number_consecutive_updates >= 4) {
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
-            disable_due_to_too_many_updates = true;
-        }
-    }
+    // these are extra tensors (inputs) that participate in the ggml graph but are not nodes
+    // they properties also have to match in order to be able to safely reuse a CUDA graph
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18583
+    // ref: https://github.com/ggml-org/llama.cpp/pull/19165
+    std::vector<ggml_cuda_graph_node_properties> extra;
 
     bool is_enabled() const {
         static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
-        return !(disable_due_to_gpu_arch || disable_cuda_graphs_due_to_env || disable_due_to_too_many_updates);
+        return !(disable_due_to_gpu_arch || disable_cuda_graphs_due_to_env);
     }
 #endif
 };
@@ -1242,10 +1324,44 @@ struct ggml_backend_cuda_context {
     cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
     cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-    std::unique_ptr<ggml_cuda_graph> cuda_graph;
-
     int curr_stream_no = 0;
 
+#ifdef USE_CUDA_GRAPH
+    // Map from first_node_ptr to cuda_graph - allows multiple graphs per context
+    // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
+    std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
+
+    ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
+        auto it = cuda_graphs.find(first_node_ptr);
+        if (it == cuda_graphs.end()) {
+            cuda_graphs[first_node_ptr] = std::make_unique<ggml_cuda_graph>();
+            return cuda_graphs[first_node_ptr].get();
+        }
+        return it->second.get();
+    }
+
+    // Check if any CUDA graph is enabled for this context (used by kernels that need to know
+    // if graphs are in use without having access to the specific graph key)
+    bool any_cuda_graph_enabled() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->is_enabled()) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    // Check if any CUDA graph has an instance for this context
+    bool any_cuda_graph_has_instance() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->instance != nullptr) {
+                return true;
+            }
+        }
+        return false;
+    }
+#endif // USE_CUDA_GRAPH
+
     explicit ggml_backend_cuda_context(int device) :
         device(device),
         name(GGML_CUDA_NAME + std::to_string(device)) {

data/ext/sources/ggml/src/ggml-cuda/convert.cu

@@ -7,7 +7,8 @@
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);
 
@@ -15,24 +16,28 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
         return;
     }
 
-    const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
+    for (int64_t i01 = blockIdx.y; i01 < ne01; i01 += gridDim.y) {
+        for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+            const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+            const int64_t i02 = dm.y;
+            const int64_t i03 = dm.x;
 
-    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+            const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
 
-    const int64_t ib = ibx0 + i00/qk; // block index
-    const int64_t iqs = (i00%qk)/qr; // quant index
-    const int64_t iybs = i00 - i00%qk; // y block start index
-    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+            const int64_t ib = ibx0 + i00/qk; // block index
+            const int64_t iqs = (i00%qk)/qr; // quant index
+            const int64_t iybs = i00 - i00%qk; // y block start index
+            const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
-    // dequantize
-    float2 v;
-    dequantize_kernel(vx, ib, iqs, v);
+            // dequantize
+            float2 v;
+            dequantize_kernel(vx, ib, iqs, v);
 
-    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
-    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+            const int64_t iy0 = (i0203*ne01 + i01)*ne00 + iybs + iqs;
+            y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+            y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+        }
+    }
 }
 
 template <bool need_check>
@@ -485,9 +490,11 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), (int)std::min(ne01, (int64_t)65535), (int)std::min(ne0203, (int64_t)65535));
     dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -612,7 +619,8 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
 
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
-        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -620,24 +628,30 @@ static __global__ void convert_unary(
         return;
     }
 
-    const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
-
     const src_t * x = (const src_t *) vx;
 
-    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
-    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+    for (int64_t i01 = blockIdx.y; i01 < ne01; i01 += gridDim.y) {
+        for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+            const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+            const int64_t i02 = dm.y;
+            const int64_t i03 = dm.x;
+
+            const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+            const int64_t iy = (i0203*ne01 + i01)*ne00 + i00;
+            y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+        }
+    }
 }
 
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, (int)std::min(ne01, (int64_t)65535), (int)std::min(ne0203, (int64_t)65535));
     convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <typename src_t, typename dst_t>

data/ext/sources/ggml/src/ggml-cuda/cpy.cu

@@ -56,7 +56,8 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
     const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
     const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
 
-    __shared__ float tile[CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
+    __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
+    int cur_tile_buf = 0;
 
 #pragma unroll
     for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
@@ -70,7 +71,7 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
             if(x < ne01 && y + j < ne00){
                 const int row = threadIdx.y+j;
                 const int col = threadIdx.x * sizeof(float)/sizeof(T);
-                T *tile2 = reinterpret_cast<T*>(tile[row]);
+                T *tile2 = reinterpret_cast<T*>(tile[cur_tile_buf][row]);
                 tile2[col] = src[imat*n + (y+j)*ne01 + x];
             }
         }
@@ -81,10 +82,12 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
         for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
             if (ty + j < ne01 && tx < ne00) {
                 const int col = (threadIdx.y+j)*sizeof(float)/sizeof(T);
-                const T *tile2 = reinterpret_cast<const T*>(tile[threadIdx.x]);
+                const T *tile2 = reinterpret_cast<const T*>(tile[cur_tile_buf][threadIdx.x]);
                 dst[imat*n + (ty+j)*ne00 + tx] = tile2[col];
             }
         }
+
+        cur_tile_buf = (cur_tile_buf + 1) % 2;
     }
 
     GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,