happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml-backend: re-enable graph reuse with pipeline parallelism (#20927)

This commit is contained in:
Aman Gupta 2026-03-24 20:47:00 +08:00 committed by GitHub
parent 29771a0a4c
commit 3fc6f1aed1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/llama-context.cpp
View File

@ -342,14 +342,6 @@ llama_context::llama_context(
if (cparams.pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled\n", __func__);
if (!graph_reuse_disable) {
// TODO: figure out a way to make graph reuse work with pipeline parallelism
// ref: https://github.com/ggml-org/llama.cpp/pull/20463
LLAMA_LOG_WARN("%s: graph reuse is currently not compatible with pipeline parallelism - disabling\n", __func__);
graph_reuse_disable = true;
}
}
sched_reserve();
@ -1189,6 +1181,13 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
if (!graph_reuse_disable && res->can_reuse(gparams)) {
//LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
// with pipeline parallelism, the previous graph_compute_async may still be running
// on the GPU. we must synchronize before set_inputs to avoid overwriting input tensors
// that the previous compute is still reading.
if (cparams.pipeline_parallel) {
ggml_backend_sched_synchronize(sched.get());
}
n_reused++;
} else {
res->reset();