From 154e0d75fccf1784fe9ff6fd76a630b66563da3d Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sat, 27 Jul 2024 07:55:01 +0200 Subject: Merge mainline llama.cpp (#3) * Merging mainline - WIP * Merging mainline - WIP AVX2 and CUDA appear to work. CUDA performance seems slightly (~1-2%) lower as it is so often the case with llama.cpp/ggml after some "improvements" have been made. * Merging mainline - fix Metal * Remove check --------- Co-authored-by: Iwan Kawrakow --- ggml/include/ggml-sycl.h | 42 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) create mode 100644 ggml/include/ggml-sycl.h (limited to 'ggml/include/ggml-sycl.h') diff --git a/ggml/include/ggml-sycl.h b/ggml/include/ggml-sycl.h new file mode 100644 index 00000000..43ab1519 --- /dev/null +++ b/ggml/include/ggml-sycl.h @@ -0,0 +1,42 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#define GGML_SYCL_NAME "SYCL" +#define GGML_SYCL_MAX_DEVICES 48 + +#ifdef __cplusplus +extern "C" { +#endif + +// backend API +GGML_API ggml_backend_t ggml_backend_sycl_init(int device); + +// devide buffer +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device); + +// split tensor buffer that splits matrices by rows across multiple devices +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); + +// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); + +GGML_API void ggml_backend_sycl_print_sycl_devices(void); +GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len); +GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size); +GGML_API GGML_CALL int ggml_backend_sycl_get_device_count(); +GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); + +// SYCL doesn't support registering host memory, keep here for reference +// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size); +// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer); +#ifdef __cplusplus +} +#endif -- cgit v1.2.3