ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : remove MPI backend (#7395)

Browse source
This commit is contained in:
slaren 2024-05-20 01:17:03 +02:00 • committed by GitHub
parent 1ea2a0036e
commit d359f30921
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 2 additions and 425 deletions
Download patch file Download diff file Expand all files Collapse all files

48
llama.cpp
View file

@ -26,9 +26,6 @@
#ifdef GGML_USE_METAL
# include "ggml-metal.h"
#endif
#ifdef GGML_USE_MPI
# include "ggml-mpi.h"
#endif
#ifndef QK_K
# ifdef GGML_QKK_64
# define QK_K 64
@ -2270,10 +2267,6 @@ struct llama_context {
// control vectors
struct llama_control_vector cvec;
#ifdef GGML_USE_MPI
ggml_mpi_context * ctx_mpi = NULL;
#endif
};
static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
@ -6336,10 +6329,7 @@ static struct ggml_tensor * llm_build_inp_embd(
inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
} else {
#ifdef GGML_USE_MPI
GGML_ASSERT(false && "not implemented");
#endif
lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
inpL = lctx.inp_embd;
ggml_set_input(lctx.inp_embd);
}
@ -11351,11 +11341,6 @@ static void llama_graph_compute(
llama_context & lctx,
ggml_cgraph * gf,
int n_threads) {
#ifdef GGML_USE_MPI
const int64_t n_layer = lctx.model.hparams.n_layer;
ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
#endif
#ifdef GGML_USE_METAL
if (ggml_backend_is_metal(lctx.backend_metal)) {
ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads);
@ -11370,10 +11355,6 @@ static void llama_graph_compute(
ggml_backend_sched_graph_compute_async(lctx.sched, gf);
// fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
#ifdef GGML_USE_MPI
ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
#endif
}
// decode a batch of tokens by evaluating the transformer
@ -11411,12 +11392,6 @@ static int llama_decode_internal(
}
lctx.n_queued_tokens += n_tokens_all;
#ifdef GGML_USE_MPI
// TODO: needs fix after #3228
GGML_ASSERT(false && "not implemented");
//ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
#endif
auto & kv_self = lctx.kv_self;
const int64_t n_embd = hparams.n_embd;
@ -15546,10 +15521,6 @@ void llama_backend_init(void) {
struct ggml_context * ctx = ggml_init(params);
ggml_free(ctx);
}
#ifdef GGML_USE_MPI
ggml_mpi_backend_init();
#endif
}
void llama_numa_init(enum ggml_numa_strategy numa) {
@ -15559,9 +15530,6 @@ void llama_numa_init(enum ggml_numa_strategy numa) {
}
void llama_backend_free(void) {
#ifdef GGML_USE_MPI
ggml_mpi_backend_free();
#endif
ggml_quantize_free();
}
@ -15962,20 +15930,6 @@ struct llama_context * llama_new_context_with_model(
}
}
#ifdef GGML_USE_MPI
ctx->ctx_mpi = ggml_mpi_init();
if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
// Enter a blocking eval loop with dummy input, letting rank=0 drive the process
// TODO: needs fix after #3228
GGML_ASSERT(false && "not implemented");
//const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos(ctx));
//while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
llama_backend_free();
exit(1);
}
#endif
return ctx;
}