Skip to content

[NPUW]Fixed high memory footprint with blob cache. #36195

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
dmatveev merged 1 commit into from
Jun 3, 2026
+22 −10
Merged

[NPUW]Fixed high memory footprint with blob cache. #36195

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
32 changes: 22 additions & 10 deletions src/plugins/intel_npu/src/plugin/npuw/compiled_model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -891,16 +891,28 @@ void ov::npuw::CompiledModel::CompiledModelDesc::serialize(ov::npuw::s11n::Strea
host_gather.idx_idx & quant_unpack_gather.dst_idx & quant_unpack_gather.src_w_idx &
quant_unpack_gather.src_z_idx & quant_unpack_gather.src_s_idx & quant_unpack_gather.idx_idx & spatial;

ov::npuw::moe::serialize_compiled_state(pipeline.context, stream, submodel_ctx);
ov::npuw::attn::serialize_compiled_state(pipeline.context, stream, submodel_ctx);

if (stream.input()) {
if (ov::npuw::attn::get_compiled_dynamic(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline, pipeline.context, ov::npuw::attn::BehaviorKind::Dynamic);
} else if (ov::npuw::attn::get_compiled_pyramid(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline, pipeline.context, ov::npuw::attn::BehaviorKind::Pyramid);
} else if (ov::npuw::attn::get_compiled_hfa(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline, pipeline.context, ov::npuw::attn::BehaviorKind::HFA);
// Function calls share pipeline.context with their function body at runtime.
// There is no need to serialize the compiled moe/attn state for each call –
// doing so would re-import NPU blobs for every repeated layer (one per call),
// causing O(N_layers) memory growth on import. Only the function body
// (compiled_model is set, or replaced_by is absent) writes/reads state.
const bool is_fcall = replaced_by.has_value() && !static_cast<bool>(compiled_model);
if (!is_fcall) {
ov::npuw::moe::serialize_compiled_state(pipeline.context, stream, submodel_ctx);
ov::npuw::attn::serialize_compiled_state(pipeline.context, stream, submodel_ctx);

if (stream.input()) {
if (ov::npuw::attn::get_compiled_dynamic(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline,
pipeline.context,
ov::npuw::attn::BehaviorKind::Dynamic);
} else if (ov::npuw::attn::get_compiled_pyramid(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline,
pipeline.context,
ov::npuw::attn::BehaviorKind::Pyramid);
} else if (ov::npuw::attn::get_compiled_hfa(pipeline.context) != nullptr) {
ov::npuw::attn::attach_runtime_behavior(pipeline, pipeline.context, ov::npuw::attn::BehaviorKind::HFA);
}
}
}

Expand Down
Loading