llama-cpp-bin 9692.0.0__tar.gz → 9702.0.0__tar.gz

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9692.0.0
+Version: 9702.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/app/llama.cpp

@@ -20,16 +20,21 @@ int llama_fit_params(int argc, char ** argv);
 int llama_quantize(int argc, char ** argv);
 int llama_perplexity(int argc, char ** argv);
 
-// hands the update over to the install script, which downloads and swaps the binary
+// Self-update is only supported for binaries built with llama-install.sh
 static int llama_update(int argc, char ** argv) {
     (void) argc;
     (void) argv;
 
+#ifdef LLAMA_INSTALL_BUILD
 #if defined(_WIN32)
     return system("powershell -NoProfile -ExecutionPolicy Bypass -Command \"irm https://llama.app/install.ps1 | iex\"");
 #else
     return system("curl -fsSL https://llama.app/install.sh | sh");
 #endif
+#else
+    printf("Updates are available only when installed from https://llama.app\n");
+    return 1;
+#endif
 }
 
 static const char * progname;
@@ -46,21 +51,29 @@ struct command {
     int (*func)(int, char **);
 };
 
+#ifdef LLAMA_INSTALL_BUILD
+#define UPDATE_HIDDEN false
+#else
+#define UPDATE_HIDDEN true
+#endif
+
 static const command cmds[] = {
-    {"serve",         "HTTP API server",                                    {"server"},   false, llama_server       },
-    {"cli",           "Command-line interactive interface",                 {"client"},   false, llama_cli          },
-    {"update",        "Update llama to the latest release",                 {},           false, llama_update       },
-    {"completion",    "Text completion",                                    {"complete"}, true,  llama_completion   },
-    {"bench",         "Benchmark prompt processing and text generation",    {},           true,  llama_bench        },
-    {"batched-bench", "Benchmark batched decoding performance",             {},           true,  llama_batched_bench},
-    {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,  llama_fit_params   },
-    {"quantize",      "Quantize a model",                                   {},           true,  llama_quantize     },
-    {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,  llama_perplexity   },
-    {"version",       "Show version",                                       {},           false, version            },
-    {"licenses",      "Show third-party licenses",                          {"credits"},  false, licenses           },
-    {"help",          "Show available commands",                            {},           false, help               },
+    {"serve",         "HTTP API server",                                    {"server"},   false,         llama_server       },
+    {"cli",           "Command-line interactive interface",                 {"client"},   false,         llama_cli          },
+    {"update",        "Update llama to the latest release",                 {},           UPDATE_HIDDEN, llama_update       },
+    {"completion",    "Text completion",                                    {"complete"}, true,          llama_completion   },
+    {"bench",         "Benchmark prompt processing and text generation",    {},           true,          llama_bench        },
+    {"batched-bench", "Benchmark batched decoding performance",             {},           true,          llama_batched_bench},
+    {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,          llama_fit_params   },
+    {"quantize",      "Quantize a model",                                   {},           true,          llama_quantize     },
+    {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,          llama_perplexity   },
+    {"version",       "Show version",                                       {},           false,         version            },
+    {"licenses",      "Show third-party licenses",                          {"credits"},  false,         licenses           },
+    {"help",          "Show available commands",                            {},           false,         help               },
 };
 
+#undef UPDATE_HIDDEN
+
 static int version(int argc, char ** argv) {
     printf("%s\n", llama_build_info());
     return 0;

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/ggml/CMakeLists.txt

@@ -249,7 +249,7 @@ option(GGML_SYCL                            "ggml: use SYCL"
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
 option(GGML_SYCL_HOST_MEM_FALLBACK          "ggml: allow host memory fallback in SYCL reorder (requires kernel 6.8+)" ON)
-option(GGML_SYCL_SUPPORT_LEVEL_ZERO         "ggml: use Level Zero API in SYCL backend"  ON)
+option(GGML_SYCL_SUPPORT_LEVEL_ZERO_API     "ggml: use Level Zero API in SYCL backend"  ON)
 option(GGML_SYCL_DNN                        "ggml: enable oneDNN in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/ggml/src/ggml-metal/ggml-metal.metal

@@ -7557,7 +7557,9 @@ typedef decltype(kernel_concat<float>) kernel_concat_t;
 
 template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
 template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
 template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
 template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
 template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt

@@ -39,8 +39,8 @@ if (WIN32)
         set(CMAKE_CXX_COMPILER "icx")
         set(CMAKE_CXX_COMPILER_ID "IntelLLVM")
     endif()
-    # Level Zero SDK path for Windows (only when GGML_SYCL_SUPPORT_LEVEL_ZERO is enabled)
-    if(GGML_SYCL_SUPPORT_LEVEL_ZERO)
+    # Level Zero SDK path for Windows (only when GGML_SYCL_SUPPORT_LEVEL_ZERO_API is enabled)
+    if(GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
         if(DEFINED ENV{LEVEL_ZERO_V1_SDK_PATH})
             set(LEVEL_ZERO_V1_SDK_PATH $ENV{LEVEL_ZERO_V1_SDK_PATH})
             if(EXISTS "${LEVEL_ZERO_V1_SDK_PATH}")
@@ -105,8 +105,8 @@ endif()
 
 target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
 
-message(STATUS "GGML_SYCL_SUPPORT_LEVEL_ZERO ${GGML_SYCL_SUPPORT_LEVEL_ZERO}")
-if (GGML_SYCL_SUPPORT_LEVEL_ZERO)
+message(STATUS "GGML_SYCL_SUPPORT_LEVEL_ZERO_API ${GGML_SYCL_SUPPORT_LEVEL_ZERO_API}")
+if (GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
     # Link against Level Zero loader for direct device memory allocation.
     # Avoids sycl::malloc_device triggering DMA-buf/TTM system RAM staging
     # in the xe kernel driver during multi-GPU inference.
@@ -114,7 +114,7 @@ if (GGML_SYCL_SUPPORT_LEVEL_ZERO)
     find_library(ZE_LOADER_LIB ze_loader HINTS ${ONEAPI_ROOT}/lib ${LEVEL_ZERO_V1_SDK_LIB_PATH} ENV LD_LIBRARY_PATH)
     if(ZE_LOADER_LIB AND LEVEL_ZERO_INCLUDE_DIR)
         target_link_libraries(ggml-sycl PRIVATE ${ZE_LOADER_LIB})
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_SUPPORT_LEVEL_ZERO)
+        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
         message(STATUS "Level Zero loader found: ${ZE_LOADER_LIB}")
         message(STATUS "Level Zero headers found: ${LEVEL_ZERO_INCLUDE_DIR}")
     else()

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/ggml/src/ggml-sycl/common.cpp

@@ -12,7 +12,7 @@
 
 #include "common.hpp"
 #include <sycl/backend.hpp>
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
 #include <level_zero/ze_api.h>
 #endif
 
@@ -84,9 +84,9 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
   return sycl_down_blk_size;
 }
 
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
 static bool ggml_sycl_use_level_zero_device_alloc(sycl::queue &q) {
-    return g_ggml_sycl_enable_level_zero &&
+    return g_ggml_sycl_use_level_zero_api &&
         q.get_device().is_gpu() &&
         q.get_backend() == sycl::backend::ext_oneapi_level_zero;
 }
@@ -95,7 +95,7 @@ static bool ggml_sycl_use_level_zero_device_alloc(sycl::queue &q) {
 // Use Level Zero zeMemAllocDevice to avoid sycl::malloc_device triggering
 // DMA-buf/TTM system RAM staging in the xe kernel driver during multi-GPU inference.
 void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
     if (ggml_sycl_use_level_zero_device_alloc(q)) {
         void *ptr = nullptr;
         auto ze_ctx = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(q.get_context());
@@ -127,7 +127,7 @@ void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
 
 void ggml_sycl_free_device(void *ptr, sycl::queue &q) {
     if (!ptr) return;
-#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
     if (ggml_sycl_use_level_zero_device_alloc(q)) {
         auto ze_ctx = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(q.get_context());
         zeMemFree(ze_ctx, ptr);

{llama_cpp_bin-9692.0.0 → llama_cpp_bin-9702.0.0}/llama.cpp/ggml/src/ggml-sycl/common.hpp

@@ -324,7 +324,7 @@ struct ggml_tensor_extra_gpu {
   optimize_feature optimized_feature;
 };
 
-extern int g_ggml_sycl_enable_level_zero;
+extern int g_ggml_sycl_use_level_zero_api;
 void * ggml_sycl_malloc_device(size_t size, sycl::queue &q);
 void ggml_sycl_free_device(void *ptr, sycl::queue &q);
 

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d-dw.cpp

@@ -0,0 +1,158 @@
+#include "conv2d-dw.hpp"
+
+struct conv2d_dw_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct conv2d_dw_kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+static inline conv2d_dw_kernel_bounds dw_calculate_kernel_bounds(int out_x, int out_y,
+                                                                  const conv2d_dw_params & p) {
+    conv2d_dw_kernel_bounds bounds;
+    bounds.y_min = sycl::max(0, (p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.y_max = sycl::min(p.kernel_h,
+                             (p.in_h + p.padding_y - out_y * p.stride_y + p.dilation_y - 1) / p.dilation_y);
+    bounds.x_min = sycl::max(0, (p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    bounds.x_max = sycl::min(p.kernel_w,
+                             (p.in_w + p.padding_x - out_x * p.stride_x + p.dilation_x - 1) / p.dilation_x);
+    return bounds;
+}
+
+static inline int dw_calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+// whcn layout: input/output stored as [N, C, H, W]
+struct dw_whcn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + c * p.in_w * p.in_h + y * p.in_w + x;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return c * p.kernel_h * p.kernel_w + ky * p.kernel_w + kx;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + c * p.out_w * p.out_h + y * p.out_w + x;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        out_x  = global_idx % p.out_w;
+        out_y  = (global_idx / p.out_w) % p.out_h;
+        c      = (global_idx / (p.out_w * p.out_h)) % p.channels;
+        n      = global_idx / (p.out_w * p.out_h * p.channels);
+    }
+};
+
+// cwhn layout: input/output stored as [N, H, W, C]
+struct dw_cwhn_layout {
+    static int input_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.in_w * p.in_h) + (y * p.in_w + x) * p.channels + c;
+    }
+    static int kernel_index(int c, int ky, int kx, const conv2d_dw_params & p) {
+        return (ky * p.kernel_w + kx) * p.channels + c;
+    }
+    static int output_index(int n, int c, int y, int x, const conv2d_dw_params & p) {
+        return n * (p.channels * p.out_w * p.out_h) + y * (p.out_w * p.channels) + x * p.channels + c;
+    }
+    static void unpack_indices(int global_idx, const conv2d_dw_params & p,
+                               int & n, int & c, int & out_y, int & out_x) {
+        c      = global_idx % p.channels;
+        out_x  = (global_idx / p.channels) % p.out_w;
+        out_y  = (global_idx / (p.channels * p.out_w)) % p.out_h;
+        n      = global_idx / (p.channels * p.out_w * p.out_h);
+    }
+};
+
+template <typename Layout>
+static void conv2d_dw_kernel(const float * input, const float * kernel, float * output,
+                             const conv2d_dw_params p, const sycl::nd_item<3> & item_ct1) {
+    const int global_idx     = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int total_elements = p.batches * p.channels * p.out_h * p.out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    int n, c, out_y, out_x;
+    Layout::unpack_indices(global_idx, p, n, c, out_y, out_x);
+
+    float acc = 0.0f;
+    const conv2d_dw_kernel_bounds bounds = dw_calculate_kernel_bounds(out_x, out_y, p);
+
+    for (int ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+        const int in_y = dw_calculate_input_coord(out_y, ky, p.stride_y, p.dilation_y, p.padding_y);
+        for (int kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+            const int in_x = dw_calculate_input_coord(out_x, kx, p.stride_x, p.dilation_x, p.padding_x);
+            acc += input[Layout::input_index(n, c, in_y, in_x, p)] *
+                   kernel[Layout::kernel_index(c, ky, kx, p)];
+        }
+    }
+
+    output[Layout::output_index(n, c, out_y, out_x, p)] = acc;
+}
+
+template <typename Layout>
+static void conv2d_dw_sycl(const float * x_d, const float * w_d, float * y_d,
+                            const conv2d_dw_params p, const queue_ptr & stream) {
+    const int total      = p.batches * p.channels * p.out_h * p.out_w;
+    const int num_blocks = (total + SYCL_CONV2D_DW_BLOCK_SIZE - 1) / SYCL_CONV2D_DW_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_DW_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_dw_kernel<Layout>(x_d, w_d, y_d, p, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv2d_dw(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    const conv2d_dw_params params = { in_w, in_h, out_w, out_h, kernel_w, kernel_h,
+                                      stride_x, stride_y, padding_x, padding_y,
+                                      dilation_x, dilation_y, channels, batches };
+
+    const queue_ptr stream = ctx.stream();
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_sycl<dw_whcn_layout>(x_d, w_d, y_d, params, stream);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_sycl<dw_cwhn_layout>(x_d, w_d, y_d, params, stream);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv2d_dw");
+    }
+}

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d-dw.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_DW_HPP
+#define GGML_SYCL_CONV2D_DW_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_DW_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d_dw(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_DW_HPP

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d-transpose.cpp

@@ -0,0 +1,125 @@
+#include "conv2d-transpose.hpp"
+#include "convert.hpp"
+
+template <typename kernel_t>
+static void conv2d_transpose_kernel(const float * input, const kernel_t * kernel, float * output,
+                                    const int in_w, const int in_h,
+                                    const int out_w, const int out_h,
+                                    const int kernel_w, const int kernel_h,
+                                    const int stride,
+                                    const int c_in, const int c_out, const int batches,
+                                    const sycl::nd_item<3> & item_ct1) {
+    const int global_idx     = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    const int total_elements = out_w * out_h * c_out * batches;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int out_x = global_idx % out_w;
+    const int out_y = (global_idx / out_w) % out_h;
+    const int c_idx = (global_idx / (out_w * out_h)) % c_out;
+    const int n_idx = global_idx / (out_w * out_h * c_out);
+
+    float acc = 0.0f;
+
+    for (int c_in_idx = 0; c_in_idx < c_in; ++c_in_idx) {
+        for (int kh = 0; kh < kernel_h; ++kh) {
+            int in_y = out_y - kh;
+            if (in_y < 0 || in_y % stride) {
+                continue;
+            }
+            in_y /= stride;
+            if (in_y >= in_h) {
+                continue;
+            }
+
+            for (int kw = 0; kw < kernel_w; ++kw) {
+                int in_x = out_x - kw;
+                if (in_x < 0 || in_x % stride) {
+                    continue;
+                }
+                in_x /= stride;
+                if (in_x >= in_w) {
+                    continue;
+                }
+
+                const int input_idx  = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + in_w * in_y + in_x;
+                const int kernel_idx = (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx +
+                                       kernel_w * kh + kw;
+
+                acc += input[input_idx] * ggml_sycl_cast<float>(kernel[kernel_idx]);
+            }
+        }
+    }
+
+    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + out_w * out_y + out_x] = acc;
+}
+
+template <typename kernel_t>
+static void conv2d_transpose_sycl(const float * input_d, const kernel_t * kernel_d, float * output_d,
+                                   const int in_w, const int in_h,
+                                   const int out_w, const int out_h,
+                                   const int kernel_w, const int kernel_h,
+                                   const int stride,
+                                   const int c_in, const int c_out, const int batches,
+                                   const queue_ptr & stream) {
+    const int total      = out_w * out_h * c_out * batches;
+    const int num_blocks = (total + SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_transpose_kernel<kernel_t>(input_d, kernel_d, output_d,
+                                             in_w, in_h, out_w, out_h, kernel_w, kernel_h,
+                                             stride, c_in, c_out, batches, item_ct1);
+        });
+}
+
+// input:  (W, H, C_in, N)
+// kernel: (W, H, C_out, C_in)
+// output: (W, H, C_out, N)
+void ggml_sycl_op_conv2d_transpose(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(input));
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const float * input_d  = (const float *) input->data;
+    float *       output_d = (float *) dst->data;
+    const void *  kernel_d = kernel->data;
+
+    const int input_w      = input->ne[0];
+    const int input_h      = input->ne[1];
+    const int channels_in  = input->ne[2];
+    const int batches      = input->ne[3];
+    const int output_w     = dst->ne[0];
+    const int output_h     = dst->ne[1];
+    const int channels_out = kernel->ne[2];
+    const int kernel_w     = kernel->ne[0];
+    const int kernel_h     = kernel->ne[1];
+    const int stride       = dst->op_params[0];
+
+    GGML_ASSERT(channels_in == kernel->ne[3]);
+    GGML_ASSERT(stride > 0);
+
+    const queue_ptr stream = ctx.stream();
+
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_transpose_sycl<sycl::half>(input_d, (const sycl::half *) kernel_d, output_d,
+                                          input_w, input_h, output_w, output_h, kernel_w, kernel_h,
+                                          stride, channels_in, channels_out, batches, stream);
+    } else {
+        conv2d_transpose_sycl<float>(input_d, (const float *) kernel_d, output_d,
+                                     input_w, input_h, output_w, output_h, kernel_w, kernel_h,
+                                     stride, channels_in, channels_out, batches, stream);
+    }
+}

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d-transpose.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_TRANSPOSE_HPP
+#define GGML_SYCL_CONV2D_TRANSPOSE_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d_transpose(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_TRANSPOSE_HPP

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d.cpp

@@ -0,0 +1,150 @@
+#include "conv2d.hpp"
+#include "convert.hpp"
+
+struct conv2d_params {
+    const int64_t IW, IH;
+    const int64_t OW, OH;
+    const int64_t KW, KH;
+    const int64_t ST_X, ST_Y;
+    const int64_t PD_X, PD_Y;
+    const int64_t DL_X, DL_Y;
+    const int64_t IC, OC;
+    const int64_t B;
+    const int64_t TOTAL;
+};
+
+struct conv2d_kernel_bounds {
+    int64_t y_min, y_max;
+    int64_t x_min, x_max;
+};
+
+static inline int64_t conv2d_max64(int64_t a, int64_t b) {
+    return (a > b) ? a : b;
+}
+
+static inline int64_t conv2d_min64(int64_t a, int64_t b) {
+    return (a < b) ? a : b;
+}
+
+static inline conv2d_kernel_bounds calculate_kernel_bounds(int64_t out_x, int64_t out_y, const conv2d_params & P) {
+    conv2d_kernel_bounds bounds;
+    bounds.y_min = conv2d_max64(0, (P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.y_max = conv2d_min64(P.KH, (P.IH + P.PD_Y - out_y * P.ST_Y + P.DL_Y - 1) / P.DL_Y);
+    bounds.x_min = conv2d_max64(0, (P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    bounds.x_max = conv2d_min64(P.KW, (P.IW + P.PD_X - out_x * P.ST_X + P.DL_X - 1) / P.DL_X);
+    return bounds;
+}
+
+static inline int calculate_input_coord(int64_t out_coord, int64_t kern_coord, int64_t stride,
+                                        int64_t dilation, int64_t padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+// whcn layout helpers (matching ggml tensor memory order)
+static inline int64_t whcn_input_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv2d_params & P) {
+    return n * (P.IC * P.IW * P.IH) + c * P.IW * P.IH + y * P.IW + x;
+}
+
+static inline int64_t whcn_kernel_index(int64_t c_out, int64_t c_in, int64_t ky, int64_t kx, const conv2d_params & P) {
+    return c_out * (P.IC * P.KH * P.KW) + c_in * (P.KH * P.KW) + ky * P.KW + kx;
+}
+
+static inline int64_t whcn_output_index(int64_t n, int64_t c, int64_t y, int64_t x, const conv2d_params & P) {
+    return n * (P.OC * P.OW * P.OH) + c * P.OW * P.OH + y * P.OW + x;
+}
+
+template <typename T>
+static void conv2d_kernel(const float * input, const T * kernel, float * output,
+                          const conv2d_params P, const sycl::nd_item<3> & item_ct1) {
+    const int64_t global_idx = item_ct1.get_local_id(2) +
+                               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+
+    if (global_idx >= P.TOTAL) {
+        return;
+    }
+
+    const int64_t out_x  = global_idx % P.OW;
+    const int64_t out_y  = (global_idx / P.OW) % P.OH;
+    const int64_t c_out  = (global_idx / (P.OW * P.OH)) % P.OC;
+    const int64_t n      = global_idx / (P.OW * P.OH * P.OC);
+
+    float acc = 0.0f;
+
+    const conv2d_kernel_bounds bounds = calculate_kernel_bounds(out_x, out_y, P);
+
+    for (int64_t c_in = 0; c_in < P.IC; ++c_in) {
+        for (int64_t ky = bounds.y_min; ky < bounds.y_max; ++ky) {
+            const int64_t in_y = calculate_input_coord(out_y, ky, P.ST_Y, P.DL_Y, P.PD_Y);
+            for (int64_t kx = bounds.x_min; kx < bounds.x_max; ++kx) {
+                const int64_t in_x = calculate_input_coord(out_x, kx, P.ST_X, P.DL_X, P.PD_X);
+                const float input_val  = input[whcn_input_index(n, c_in, in_y, in_x, P)];
+                const T     kernel_val = kernel[whcn_kernel_index(c_out, c_in, ky, kx, P)];
+                acc += input_val * ggml_sycl_cast<float>(kernel_val);
+            }
+        }
+    }
+
+    output[whcn_output_index(n, c_out, out_y, out_x, P)] = acc;
+}
+
+template <typename T>
+static void conv2d_sycl(const float * X_D, const T * K_D, float * Y_D,
+                        const conv2d_params P, const queue_ptr & stream) {
+    const int num_blocks = (P.TOTAL + SYCL_CONV2D_BLOCK_SIZE - 1) / SYCL_CONV2D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV2D_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv2d_kernel<T>(X_D, K_D, Y_D, P, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+    const float *       K_D    = (const float *) kernel->data;
+    const float *       X_D    = (const float *) input->data;
+    float *             Y_D    = (float *) dst->data;
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    // same number of input channels
+    GGML_ASSERT(input->ne[2] == kernel->ne[2]);
+
+    const queue_ptr stream = ctx.stream();
+
+    const int32_t * p    = (const int32_t *) dst->op_params;
+    const int       ST_X = p[0];
+    const int       ST_Y = p[1];
+    const int       PD_X = p[2];
+    const int       PD_Y = p[3];
+    const int       DL_X = p[4];
+    const int       DL_Y = p[5];
+
+    // no cwhn layout support
+    GGML_ASSERT(p[6] == 0);
+
+    const int IW = input->ne[0];
+    const int IH = input->ne[1];
+    const int OW = dst->ne[0];
+    const int OH = dst->ne[1];
+    const int KW = kernel->ne[0];
+    const int KH = kernel->ne[1];
+    const int IC = input->ne[2];
+    const int OC = kernel->ne[3];
+    const int B  = input->ne[3];
+
+    const int64_t     total  = (int64_t) B * OC * OH * OW;
+    const conv2d_params params = { IW, IH, OW, OH, KW, KH, ST_X, ST_Y, PD_X, PD_Y, DL_X, DL_Y, IC, OC, B, total };
+
+    if (kernel->type == GGML_TYPE_F16) {
+        conv2d_sycl<sycl::half>(X_D, (const sycl::half *) K_D, Y_D, params, stream);
+    } else {
+        conv2d_sycl<float>(X_D, K_D, Y_D, params, stream);
+    }
+}

llama_cpp_bin-9702.0.0/llama.cpp/ggml/src/ggml-sycl/conv2d.hpp

@@ -0,0 +1,10 @@
+#ifndef GGML_SYCL_CONV2D_HPP
+#define GGML_SYCL_CONV2D_HPP
+
+#include "common.hpp"
+
+#define SYCL_CONV2D_BLOCK_SIZE 256
+
+void ggml_sycl_op_conv2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_CONV2D_HPP