Skip to content

[NPUW] Remove NPUW-side Whisper cross-attention SDPA decomposition #1

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

Draft
parthmah14 wants to merge 2 commits into
base: parthmah/whisper-decomposed-sdpa
Choose a base branch
from
Draft

[NPUW] Remove NPUW-side Whisper cross-attention SDPA decomposition #1

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6a6aee4
[NPUW] Remove NPUW-side Whisper cross-attention SDPA decomposition
Aug 11, 2026
ea2553d
[NPUW] Address review: restore whisper_kvcache_size-1 adjustment, tri…
Aug 13, 2026
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
1 change: 0 additions & 1 deletion src/plugins/intel_npu/src/al/include/intel_npu/config/npuw_option_defs.inc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,7 +84,6 @@ INTEL_NPU_NPUW_ANYMAP_OPT(NPUW_LLM_SHARED_LM_HEAD_CONFIG, ov::intel_npu::npuw::l
INTEL_NPU_NPUW_ANYMAP_OPT(NPUW_LLM_ADDITIONAL_SHARED_LM_HEAD_CONFIG, ov::intel_npu::npuw::llm, additional_shared_lm_head_config, "++NPUW_LLM_SHARED_HEAD_CONFIG", LLM, EXPOSED, CACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_WHISPER, bool, false, ov::intel_npu::npuw::whisper, enabled, "NPUW_WHISPER", LLM, EXPOSED, CACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_WHISPER_EOS_TOKEN, uint64_t, 50257, ov::intel_npu::npuw::whisper, whisper_eos_token, "NPUW_WHISPER_EOS_TOKEN", LLM, EXPOSED, UNCACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_WHISPER_DECOMPOSE_SDPA, bool, false, ov::intel_npu::npuw::whisper, whisper_decompose_sdpa, "NPUW_WHISPER_DECOMPOSE_SDPA", LLM, EXPOSED, CACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_EAGLE, bool, false, ov::intel_npu::npuw::eagle, enabled, "NPUW_EAGLE", LLM, EXPOSED, CACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_TEXT_EMBED, bool, false, ov::intel_npu::npuw::text_embed, enabled, "NPUW_TEXT_EMBED", LLM, EXPOSED, CACHED, ALL)
INTEL_NPU_NPUW_SIMPLE_OPT(NPUW_KOKORO, bool, false, ov::intel_npu::npuw::kokoro, enabled, "NPUW_KOKORO", KOKORO, EXPOSED, UNCACHED, ALL)
Expand Down
34 changes: 5 additions & 29 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 @@ -853,32 +853,10 @@ ov::npuw::LLMCompiledModel::LLMCompiledModel(const std::shared_ptr<ov::Model>& m
m_kvcache_desc = KVCacheDesc{whisper_max_prompt_size, whisper_kvcache_size, 0u, whisper_seq_len_dim, 1u};
whisper_lhs_seq_size =
static_cast<uint32_t>(prefill_model->input("encoder_hidden_states").get_partial_shape()[1].get_length());
auto whisper_decompose_sdpa = m_cfg.get<::intel_npu::NPUW_WHISPER_DECOMPOSE_SDPA>();
if (whisper_decompose_sdpa) {
m_kvcache_desc.max_prompt_size = whisper_kvcache_size - 1;
}

@eshiryae eshiryae Aug 13, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We still should use whisper_kvcache_size - 1 for reshape to static in case of decomposed SDPA. As when word_timestamps is enabled, the decoder may receive tokens decoded from the entire audio chunk.

@parthmah14 parthmah14 Aug 13, 2026

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good catch - this is needed whenever word_timestamps is active, not just as an artifact of the removed option. Restored it, now detected structurally via a new has_decomposed_cross_attention_sdpa() helper (pushed in ea2553d).


auto prepare_prefill_model = ov::npuw::util::PrepareWhisperPrefillModel(m_kvcache_desc.max_prompt_size,
whisper_lhs_seq_size,
whisper_decompose_sdpa);
prepare_prefill_model.run_on_model(prefill_model); // Whisper decoder model
ov::npuw::util::PrepareWhisperPrefillModel(m_kvcache_desc.max_prompt_size, whisper_lhs_seq_size)
.run_on_model(prefill_model); // Whisper decoder model
ov::npuw::util::PrepareWhisperKVCacheModel().run_on_model(kvcache_model); // Whisper decoder_with_past model

// FIXME: Whisper Decompose SDPA
// WA: to mock new "cross_attention_qk_scaled_scores" outputs in original model
if (whisper_decompose_sdpa) {
m_decomposed_sdpa_size = prepare_prefill_model.get_decomposed_sdpa_size();
auto& mutable_outputs = const_cast<std::vector<ov::Output<const ov::Node>>&>(this->outputs());
for (size_t idx = 0; idx < m_decomposed_sdpa_size; idx++) {
auto fake_param = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape{});
auto fake_result = std::make_shared<ov::op::v0::Result>(fake_param);
fake_result->output(0).get_tensor().add_names(
{WhisperInferRequest::whisper_layer_names::qk_scores,
WhisperInferRequest::whisper_layer_names::qk_scores_ + std::to_string(idx)});

mutable_outputs.emplace_back(fake_result->output(0));
}
}
}

LOG_DEBUG("Make prefill model with static shapes");
Expand Down Expand Up @@ -1316,8 +1294,8 @@ void ov::npuw::LLMCompiledModel::serialize(std::ostream& raw_stream, const ov::n
m_kvcache_desc.dim & m_kvcache_desc.max_generation_token_len & m_kvcache_desc.v_tensors_transposed_pre &
m_kvcache_desc.v_tensors_transposed_gen & m_prefill_chunk_size & m_use_chunk_prefill & m_max_lora_rank &
m_enable_prefix_caching & m_prefix_caching_block_size & m_prefix_caching_max_num_blocks &
m_longrope_context_limit & m_is_whisper & m_eos_token_id & m_decomposed_sdpa_size & m_is_eagle &
m_is_embedding & m_is_block_kv_cache;
m_longrope_context_limit & m_is_whisper & m_eos_token_id & m_is_eagle & m_is_embedding &
m_is_block_kv_cache;

// Write config
stream & m_cfg;
Expand Down Expand Up @@ -1537,8 +1515,7 @@ std::shared_ptr<ov::npuw::LLMCompiledModel> ov::npuw::LLMCompiledModel::deserial
compiled->m_use_chunk_prefill & compiled->m_max_lora_rank & compiled->m_enable_prefix_caching &
compiled->m_prefix_caching_block_size & compiled->m_prefix_caching_max_num_blocks &
compiled->m_longrope_context_limit & compiled->m_is_whisper & compiled->m_eos_token_id &
compiled->m_decomposed_sdpa_size & compiled->m_is_eagle & compiled->m_is_embedding &
compiled->m_is_block_kv_cache;
compiled->m_is_eagle & compiled->m_is_embedding & compiled->m_is_block_kv_cache;

// Deserialize config
stream & compiled->m_cfg;
Expand Down Expand Up @@ -1672,7 +1649,6 @@ void ov::npuw::LLMCompiledModel::implement_properties() {
BIND(npuw::llm::shared_lm_head, NPUW_LLM_SHARED_HEAD, get),
BIND(npuw::whisper::enabled, NPUW_WHISPER, get),
BIND(npuw::whisper::whisper_eos_token, NPUW_WHISPER_EOS_TOKEN, get),
BIND(npuw::whisper::whisper_decompose_sdpa, NPUW_WHISPER_DECOMPOSE_SDPA, get),
BIND(npuw::eagle::enabled, NPUW_EAGLE, get),
BIND(npuw::text_embed::enabled, NPUW_TEXT_EMBED, get)});
#undef BIND
Expand Down
1 change: 0 additions & 1 deletion src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,7 +139,6 @@ class LLMCompiledModel : public ov::npuw::ICompiledModel {

bool m_is_whisper = false;
uint64_t m_eos_token_id = 0;
size_t m_decomposed_sdpa_size = 0;

bool m_is_embedding = false;

Expand Down
2 changes: 1 addition & 1 deletion src/plugins/intel_npu/src/plugin/npuw/serialization.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,7 @@ const constexpr ov::npuw::s11n::IndicatorType NPUW_LLM_COMPILED_MODEL_INDICATOR
const constexpr ov::npuw::s11n::IndicatorType NPUW_GQA_COMPILED_MODEL_INDICATOR =
{char{0x47}, char{0x51}, char{0x41}, char{0x43}, char{0x4d}, char{0x4f}};

const constexpr char* NPUW_SERIALIZATION_VERSION = "0.28";
const constexpr char* NPUW_SERIALIZATION_VERSION = "0.29";

// Forward declaration
namespace intel_npu {
Expand Down
Loading