Skip to content

[NPUW] Fix multiple MatMul matching issue in NPUW LM head cutting #32475

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
AsyaPronina merged 4 commits into from
Dec 10, 2025
Merged

[NPUW] Fix multiple MatMul matching issue in NPUW LM head cutting #32475

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
fd65115
Fix multiple MatMul matching issue in NPUW LM head cutting
GuoliangShiIntel Oct 20, 2025
550db89
rt_info solution
GuoliangShiIntel Dec 3, 2025
466bc14
define hidden_output_name_key
GuoliangShiIntel Dec 3, 2025
5d7e933
Add node rt_info
GuoliangShiIntel Dec 4, 2025
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
8 changes: 8 additions & 0 deletions src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -881,6 +881,14 @@ class CutLMHead : public ov::pass::MatcherPass {
auto matched_matmul = std::static_pointer_cast<ov::op::v0::MatMul>(matched_node_matmul);
auto matched_result = std::static_pointer_cast<ov::op::v0::Result>(matched_node_result);

// Some LLMs add intermediate hidden state outputs that can interfere with LM head detection.
// Skip Result nodes that were manually added (marked with "manually_added_output" in RT_INFO).
// For example, Eagle-3 target/draft models add "last_hidden_state" output which should be skipped.
const auto& rt_info = matched_result->get_rt_info();
if (rt_info.count("manually_added_output")) {
return false;
}

// Cut point:
auto matmul_first_source = matched_matmul->input(0).get_source_output();

Expand Down
Loading