CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-05 09:36:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

PR comments

Browse Source
This commit is contained in:
Evan Miller
2023-07-06 20:18:41 -04:00
parent 32deabfdc8
commit 06a239343c
7 changed files with 125 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
llama.cpp
View File

@@ -19,6 +19,9 @@
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_MPI
#include "ggml-mpi.h"
#endif
#ifdef GGML_USE_K_QUANTS
#ifndef QK_K
#ifdef GGML_QKK_64
@@ -1332,10 +1335,10 @@ static bool llama_eval_internal(
if (lctx.mpi_rank > 0) {
#ifdef GGML_USE_MPI
inpL = ggml_recv_tensor(ctx0, NULL,
inpL = ggml_mpi_recv_tensor(ctx0, NULL,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
lctx.mpi_rank-1);
ggml_set_name(inpL, "recv");
ggml_set_name(inpL, "mpi_recv");
#else
GGML_ASSERT(false);
#endif
@@ -1591,15 +1594,23 @@ static bool llama_eval_internal(
struct ggml_tensor * embeddings = NULL;
if (lctx.mpi_size > 1) {
cur = ggml_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
ggml_set_name(cur, "send");
#ifdef GGML_USE_MPI
cur = ggml_mpi_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
ggml_set_name(cur, "mpi_send");
#else
GGML_ASSERT(false);
#endif
}
if (lctx.mpi_rank == 0) {
if (lctx.mpi_size > 1) {
cur = ggml_recv_tensor(ctx0, cur,
#ifdef GGML_USE_MPI
cur = ggml_mpi_recv_tensor(ctx0, cur,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
lctx.mpi_size-1);
ggml_set_name(cur, "recv");
ggml_set_name(cur, "mpi_recv");
#else
GGML_ASSERT(false);
#endif
}
// norm
{
@@ -3504,14 +3515,6 @@ int llama_n_embd(const struct llama_context * ctx) {
return ctx->model.hparams.n_embd;
}
int llama_mpi_rank(const struct llama_context * ctx) {
return ctx->mpi_rank;
}
int llama_mpi_size(const struct llama_context * ctx) {
return ctx->mpi_size;
}
int llama_get_vocab(
const struct llama_context * ctx,
const char * * strings,