Skip to content

fix for op regression due to is_model_splitted #259

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

Open
cavusmustafa wants to merge 1 commit into
base: dev_backend_openvino
Choose a base branch
from
Open

fix for op regression due to is_model_splitted #259

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
19 changes: 16 additions & 3 deletions ggml/src/ggml-openvino/utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -170,7 +170,22 @@ enum ggml_status ov_graph_compute_dynamic(ggml_cgraph * cgraph, std::shared_ptr<
const auto & stateful = r_ctx->stateful;
static auto is_static = false;

bool model_is_splitted = is_model_splitted(cgraph);
static const bool cache_disabled = ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_CACHE");

// is_model_splitted is O(n_nodes^2) plus a create_weight_nodes scan and takes ~20 ms
// on a Llama-1B decode graph. It is called once per graph_compute invocation but the
// graph shape is identical across all decode steps, so memoize by graph_key: compute
// graph_key first (a few hundred us), and if the same key is already in decoder_cache
// we know the graph is not splitted (only not-splitted graphs get inserted there).
graph_key key(cgraph);
bool key_seen = false;
if (!cache_disabled) {
std::lock_guard<std::mutex> map_lock(r_ctx->ctx_mutex);
key_seen = r_ctx->decoder_cache.find(key) != r_ctx->decoder_cache.end();
}

bool model_is_splitted = key_seen ? false : is_model_splitted(cgraph);

if (is_naive(cgraph)) {
if (!model_is_splitted) {
return naive_compute(cgraph, core, device, config);
Expand All @@ -185,8 +200,6 @@ enum ggml_status ov_graph_compute_dynamic(ggml_cgraph * cgraph, std::shared_ptr<
ComputeParams c_params;
std::tie(m_params, c_params) = GgmlOvDecoder::compute_llm_params(cgraph, is_static);

graph_key key(cgraph);
static const bool cache_disabled = ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_CACHE");
const bool cache_enabled = !model_is_splitted && !cache_disabled;
bool cache_hit = false;

Expand Down
Loading