vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into gg/flash-attn

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-28 21:53:51 +02:00
commit 0ad44baf33
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
36 changed files with 85118 additions and 89 deletions
Download patch file Download diff file Expand all files Collapse all files

65
ggml.c
View file

@ -248,6 +248,10 @@ inline static void * ggml_aligned_malloc(size_t size) {
#include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h"
#elif defined(GGML_USE_VULKAN)
#include "ggml-vulkan.h"
#elif defined(GGML_USE_SYCL)
#include "ggml-sycl.h"
#endif
// floating point type used to accumulate sums
@ -2355,6 +2359,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
ggml_init_cublas();
#elif defined(GGML_USE_CLBLAST)
ggml_cl_init();
#elif defined(GGML_USE_VULKAN)
ggml_vk_init();
#elif defined(GGML_USE_SYCL)
ggml_init_sycl();
#endif
ggml_setup_op_has_task_pass();
@ -8130,7 +8138,7 @@ static void ggml_compute_forward_mul_f32(
const int ith = params->ith;
const int nth = params->nth;
#ifdef GGML_USE_CLBLAST
#if defined(GGML_USE_CLBLAST)
if (src1->backend == GGML_BACKEND_GPU) {
// TODO: OpenCL kernel support full broadcast
GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
@ -10085,7 +10093,7 @@ static void ggml_compute_forward_mul_mat(
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
if (ggml_compute_forward_mul_mat_use_blas(dst)) {
const int64_t ne_plane = ne01*ne00;
const int64_t desired_wsize = ne13*ne12*ne_plane*sizeof(float);
const size_t desired_wsize = ne13*ne12*ne_plane*sizeof(float);
UNUSED(desired_wsize);
if (params->type == GGML_TASK_INIT) {
@ -15010,8 +15018,26 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
}
GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
#elif defined(GGML_USE_VULKAN)
const bool skip_cpu = ggml_vk_compute_forward(params, tensor);
#ifdef GGML_VULKAN_CHECK_RESULTS
if (skip_cpu) {
ggml_vk_check_results_1(params, tensor);
}
#endif
if (skip_cpu) {
return;
}
GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
#endif // GGML_USE_CUBLAS
#ifdef GGML_USE_SYCL
bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
if (skip_cpu) {
return;
}
#endif // GGML_USE_SYCL
switch (tensor->op) {
case GGML_OP_DUP:
{
@ -17418,6 +17444,17 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
}
}
#ifdef GGML_USE_VULKAN
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_vk_preallocate_buffers_graph(cgraph->nodes[i]);
}
ggml_vk_preallocate_buffers();
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_vk_build_graph(cgraph->nodes[i], i == cgraph->n_nodes - 1);
}
#endif
const int n_threads = cplan->n_threads;
struct ggml_compute_state_shared state_shared = {
@ -17469,6 +17506,10 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
}
}
#ifdef GGML_USE_VULKAN
ggml_vk_graph_cleanup();
#endif
// performance stats (graph)
{
int64_t perf_cycles_cur = ggml_perf_cycles() - perf_start_cycles;
@ -20603,7 +20644,7 @@ int ggml_cpu_has_wasm_simd(void) {
}
int ggml_cpu_has_blas(void) {
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_VULKAN) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_SYCL)
return 1;
#else
return 0;
@ -20626,8 +20667,24 @@ int ggml_cpu_has_clblast(void) {
#endif
}
int ggml_cpu_has_vulkan(void) {
#if defined(GGML_USE_VULKAN)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_sycl(void) {
#if defined(GGML_USE_SYCL)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_gpublas(void) {
return ggml_cpu_has_cublas() || ggml_cpu_has_clblast();
return ggml_cpu_has_cublas() || ggml_cpu_has_clblast() || ggml_cpu_has_vulkan() || ggml_cpu_has_sycl();
}
int ggml_cpu_has_sse3(void) {