CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : sync with recent PRs on master

Browse Source
This commit is contained in:
Georgi Gerganov
2023-08-17 19:03:15 +03:00
parent 8ace03ad3d
commit 11bf4366c2
7 changed files with 616 additions and 672 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
llama.cpp
View File

@@ -11,7 +11,7 @@
#include "ggml.h"
#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL)
#if !defined(GGML_USE_CUBLAS)
# include "ggml-alloc.h"
# define LLAMA_USE_ALLOCATOR
#else
@@ -1895,11 +1895,11 @@ static struct ggml_cgraph * llama_build_graph(
ggml_set_name(Q, "Q");
struct ggml_tensor * K =
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd_gqa, il*n_ctx*ggml_element_size(kv_self.k)*n_embd_gqa),
n_embd_head, n_head_kv, n_past + N),
0, 2, 1, 3);
ggml_view_3d(ctx0, kv_self.k,
n_embd_head, n_past + N, n_head_kv,
ggml_element_size(kv_self.k)*n_embd_gqa,
ggml_element_size(kv_self.k)*n_embd_head,
ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
offload_func_kq(K);
ggml_set_name(K, "K");
@@ -1928,9 +1928,9 @@ static struct ggml_cgraph * llama_build_graph(
struct ggml_tensor * V =
ggml_view_3d(ctx0, kv_self.v,
n_past + N, n_embd_head, n_head_kv,
n_ctx*ggml_element_size(kv_self.v),
n_ctx*ggml_element_size(kv_self.v)*n_embd_head,
n_ctx*ggml_element_size(kv_self.v)*n_embd_gqa*il);
ggml_element_size(kv_self.v)*n_ctx,
ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
offload_func_v(V);
ggml_set_name(V, "V");
@@ -2131,11 +2131,7 @@ static bool llama_eval_internal(
#endif
#ifdef GGML_USE_METAL
if (lctx.ctx_metal && N == 1) {
// TODO: disabled until #2413 is resolved
//if (!ggml_metal_if_optimized(lctx.ctx_metal)) {
// ggml_metal_graph_find_concurrency(lctx.ctx_metal, gf);
//}
if (lctx.ctx_metal) {
ggml_metal_set_n_cb (lctx.ctx_metal, n_threads);
ggml_metal_graph_compute(lctx.ctx_metal, gf);
ggml_metal_get_tensor (lctx.ctx_metal, res);
@@ -2143,22 +2139,6 @@ static bool llama_eval_internal(
ggml_metal_get_tensor(lctx.ctx_metal, embeddings);
}
} else {
// IMPORTANT:
// Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla
// ggml_graph_compute(). It uses Apple's Accelerate CBLAS API which takes advantage of the ANE or the AMX
// coprocessor.
//
// When we implement Matrix x Matrix Metal multiplication, we can avoid this branch.
// But for now, we have focused only on Matrix x Vector Metal multiplication.
//
// TODO: avoid these syncs via shared memory (ref #1696)
//
if (lctx.ctx_metal) {
// We need to sync the GPU KV cache with the CPU KV cache
ggml_metal_get_tensor(lctx.ctx_metal, kv_self.k);
ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v);
}
ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
}
#else
@@ -4141,7 +4121,18 @@ struct llama_context * llama_new_context_with_model(
int n_past = hparams.n_ctx - n_tokens;
llama_token token = llama_token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
ggml_cgraph * gf = llama_build_graph(*ctx, &token, NULL, n_tokens, n_past);
#ifdef GGML_USE_METAL
if (params.n_gpu_layers > 0) {
ctx->ctx_metal = ggml_metal_init(1);
if (!ctx->ctx_metal) {
LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
llama_free(ctx);
return NULL;
}
ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
}
#endif
// measure memory requirements for the graph
size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
@@ -4159,6 +4150,11 @@ struct llama_context * llama_new_context_with_model(
ctx->buf_alloc.resize(alloc_size);
ctx->alloc = ggml_allocr_new(ctx->buf_alloc.data, ctx->buf_alloc.size, tensor_alignment);
#ifdef GGML_USE_METAL
if (ctx->ctx_metal) {
ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
}
#endif
}
#else
ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type) + ggml_graph_overhead());
@@ -4173,13 +4169,6 @@ struct llama_context * llama_new_context_with_model(
#ifdef GGML_USE_METAL
if (params.n_gpu_layers > 0) {
// this allocates all Metal resources and memory buffers
ctx->ctx_metal = ggml_metal_init(1);
if (!ctx->ctx_metal) {
LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
llama_free(ctx);
return NULL;
}
void * data_ptr = NULL;
size_t data_size = 0;
@@ -4208,8 +4197,7 @@ struct llama_context * llama_new_context_with_model(
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].data, ctx->buf_scratch[0].size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].data, ctx->buf_scratch[1].size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
#undef LLAMA_METAL_CHECK_BUF
}
#endif