Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-05 18:44:51 +01:00
Code Issues Packages Projects Releases Wiki Activity

mpi : fix output tensor after MPI compute (still not working)

Browse Source
This commit is contained in:
Georgi Gerganov 2023-07-09 17:01:08 +03:00
parent c717c5185f
commit ef37dd14e7
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 10 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
ggml-mpi.c
View File

@ -57,7 +57,7 @@ void ggml_mpi_eval_init(
MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD); MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
} }
int ggml_graph_get_node_idx( struct ggml_cgraph * gf, const char * name) { int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
struct ggml_tensor * t = ggml_graph_get_tensor(gf, name); struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
if (t == NULL) { if (t == NULL) {
fprintf(stderr, "%s: tensor %s not found\n", __func__, name); fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
@ -141,8 +141,8 @@ void ggml_mpi_graph_compute(
const int il0 = (mpi_idx + 0) * n_per_node; const int il0 = (mpi_idx + 0) * n_per_node;
const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node); const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
char name_l0[64]; char name_l0[GGML_MAX_NAME];
char name_l1[64]; char name_l1[GGML_MAX_NAME];
snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0); snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1); snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
@ -175,6 +175,11 @@ void ggml_mpi_graph_compute(
ggml_graph_compute(ctx, gf); ggml_graph_compute(ctx, gf);
//if (mpi_rank == 0) {
// ggml_graph_print(gf);
// ggml_graph_dump_dot(gf, NULL, "llama.dot");
//}
//fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank); //fprintf(stderr, "%s: node %d: done\n", __func__, mpi_rank);
// send the output data to the next node // send the output data to the next node

2
llama.cpp
View File

@ -1637,6 +1637,8 @@ static bool llama_eval_internal(
} }
#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);
cur = gf.nodes[gf.n_nodes - 1];
#else #else
ggml_graph_compute(ctx0, &gf); ggml_graph_compute(ctx0, &gf);
#endif #endif