CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

mpi : fix after master merge

Browse Source
This commit is contained in:
Georgi Gerganov
2023-07-09 22:23:04 +03:00
parent 81c5ddd532
commit 0492363137
3 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
ggml-mpi.c
View File

@@ -78,7 +78,8 @@ void ggml_mpi_graph_compute(
struct ggml_mpi_context * ctx_mpi, struct ggml_mpi_context * ctx_mpi,
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_cgraph * gf, struct ggml_cgraph * gf,
int n_layers) { int n_layers,
int n_threads) {
const int mpi_rank = ctx_mpi->rank; const int mpi_rank = ctx_mpi->rank;
const int mpi_size = ctx_mpi->size; const int mpi_size = ctx_mpi->size;
@@ -194,7 +195,7 @@ void ggml_mpi_graph_compute(
//fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1); //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
} }
ggml_graph_compute(ctx, gf); ggml_graph_compute_with_ctx(ctx, gf, n_threads);
//fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank); //fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank);

3
ggml-mpi.h
View File

@@ -28,7 +28,8 @@ void ggml_mpi_graph_compute(
struct ggml_mpi_context * ctx_mpi, struct ggml_mpi_context * ctx_mpi,
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_cgraph * gf, struct ggml_cgraph * gf,
int n_layers); int n_layers,
int n_threads);
#ifdef __cplusplus #ifdef __cplusplus
} }

2
llama.cpp
View File

@@ -1657,7 +1657,7 @@ static bool llama_eval_internal(
ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads); ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
} }
#elif GGML_USE_MPI #elif GGML_USE_MPI
ggml_mpi_graph_compute(lctx.ctx_mpi, ctx0, &gf, n_layer); ggml_mpi_graph_compute(lctx.ctx_mpi, ctx0, &gf, n_layer, n_threads);
cur = gf.nodes[gf.n_nodes - 1]; cur = gf.nodes[gf.n_nodes - 1];
#else #else