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

llama : disable pipeline parallelism if compute buffer allocation fails (#16748)

Browse Source
This commit is contained in:
Diego Devesa
2025-10-27 13:51:28 -07:00
committed by GitHub
parent 10640e31aa
commit 5a4ff43e7d
1 changed files with 8 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/llama-context.cpp
View File

@@ -268,9 +268,7 @@ llama_context::llama_context(
if (pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
}
}
if (!hparams.vocab_only) {
llama_memory_context_ptr mctx;
if (memory) {
LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -342,9 +340,16 @@ llama_context::llama_context(
// reserve pp (prompt processing) graph first so that buffers are only allocated once
{
auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
if (!gf) {
if (pipeline_parallel) {
LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload));
gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
}
if (!gf) {
throw std::runtime_error("failed to allocate compute pp buffers");
}
}
n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_pp = ggml_graph_n_nodes(gf);