Skip to content

context : fix worst-case reserve outputs #12545

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Mar 25, 2025
+21 −4
Merged

context : fix worst-case reserve outputs #12545

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 21 additions & 4 deletions src/llama-context.cpp
View file
Original file line number Diff line number Diff line change
@@ -294,10 +294,7 @@ llama_context::llama_context(
// TODO: something cleaner
const auto n_outputs_save = n_outputs;

// max number of outputs
n_outputs = n_tokens;

LLAMA_LOG_DEBUG("%s: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);

int n_splits_pp = -1;
int n_nodes_pp = -1;
@@ -313,8 +310,15 @@ llama_context::llama_context(
// reserve pp graph first so that buffers are only allocated once
{
llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

// max number of outputs
n_outputs = ubatch_pp.n_tokens;

LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);

auto * gf = graph_init();
graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);

if (!ggml_backend_sched_reserve(sched.get(), gf)) {
throw std::runtime_error("failed to allocate compute pp buffers");
}
@@ -326,20 +330,33 @@ llama_context::llama_context(
// reserve with tg graph to get the number of splits and nodes
{
llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

n_outputs = ubatch_tg.n_tokens;

LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_tg.n_tokens, ubatch_tg.n_seqs);

auto * gf = graph_init();
graph_build(ctx_compute.get(), gf, ubatch_tg, LLM_GRAPH_TYPE_DEFAULT);

if (!ggml_backend_sched_reserve(sched.get(), gf)) {
throw std::runtime_error("failed to allocate compute tg buffers");
}

n_splits_tg = ggml_backend_sched_get_n_splits(sched.get());
n_nodes_tg = ggml_graph_n_nodes(gf);
}

// reserve again with pp graph to avoid ggml-alloc reallocations during inference
{
llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

n_outputs = ubatch_pp.n_tokens;

LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);

auto * gf = graph_init();
graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);

if (!ggml_backend_sched_reserve(sched.get(), gf)) {
throw std::runtime_error("failed to allocate compute pp buffers");
}