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

fixes

Browse source
This commit is contained in:
Evan Miller 2023-07-04 00:09:02 -04:00
parent d05ca74dd8
commit 668ba5fe0b
3 changed files with 14 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

4
Makefile
View file

@ -150,8 +150,8 @@ ifndef LLAMA_NO_ACCELERATE
endif # LLAMA_NO_ACCELERATE
ifdef LLAMA_MPI
CFLAGS += -DGGML_USE_MPI
CXXFLAGS += -DGGML_USE_MPI
CFLAGS += -DGGML_USE_MPI -Wno-cast-qual -Wno-int-to-void-pointer-cast -Wno-void-pointer-to-int-cast
CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
endif # LLAMA_MPI
ifdef LLAMA_OPENBLAS

2
examples/simple/simple.cpp
View file

@ -173,6 +173,8 @@ int main(int argc, char ** argv)
llama_free( ctx );
llama_free_model( model );
llama_finalize_backend();
return 0;
}

8
llama.cpp
View file

@ -1581,17 +1581,17 @@ static bool llama_eval_internal(
// used at the end to optionally extract the embeddings
struct ggml_tensor * embeddings = NULL;
#ifdef GGML_USE_MPI
if (lctx.mpi_size > 1) {
cur = ggml_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
ggml_set_name(cur, "send");
#endif
}
if (lctx.mpi_rank == 0) {
#ifdef GGML_USE_MPI
if (lctx.mpi_size > 1) {
cur = ggml_recv_tensor(ctx0, cur,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
lctx.mpi_size-1);
ggml_set_name(cur, "recv");
#endif
}
// norm
{
cur = ggml_rms_norm(ctx0, cur);