Skip to content

[Whisper] Decompose cross-attention SDPA for NPU too, matching CPU/GPU #4306

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 1 commit into
base: master
Choose a base branch
from
+5 −10
Draft

[Whisper] Decompose cross-attention SDPA for NPU too, matching CPU/GPU #4306

Show file tree
Hide file tree
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
1 change: 0 additions & 1 deletion src/cpp/src/utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,6 @@ void update_npu_config_whisper(ov::AnyMap& config,
update_config(config, {"NPUW_LLM", "YES"});
update_config(config, {"NPUW_WHISPER", "YES"});
rename_key(config, "WHISPER_EOS_TOKEN", "NPUW_WHISPER_EOS_TOKEN");
rename_key(config, "WHISPER_DECOMPOSE_SDPA", "NPUW_WHISPER_DECOMPOSE_SDPA");

update_config(config, {"NPUW_LLM_BATCH_DIM", kv_pos.batch});
update_config(config, {"NPUW_LLM_SEQ_LEN_DIM", kv_pos.seq_len});
Expand Down
10 changes: 5 additions & 5 deletions src/cpp/src/whisper/models/statefull_decoder.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,11 @@ WhisperStatefullDecoder::WhisperStatefullDecoder(const std::filesystem::path& mo

m_has_cache_position = utils::has_input(model, "cache_position");

if (m_decompose_cross_attention_spda_ops) {
ov::genai::decompose_scaled_dot_product_attention_for_whisper(model);
ov::genai::add_cross_attention_qk_scaled_scores_outputs_for_whisper(model);
}
Comment on lines +37 to +40

ov::CompiledModel compiled_model;
if (device == "NPU") {
auto kv_pos = ov::genai::utils::get_kv_axes_pos(model);
Expand All @@ -43,11 +48,6 @@ WhisperStatefullDecoder::WhisperStatefullDecoder(const std::filesystem::path& mo
utils::KVDesc kv_desc;
std::tie(compiled_model, kv_desc) = utils::compile_decoder_for_npu(model, properties, kv_pos, true);
} else {
if (m_decompose_cross_attention_spda_ops) {
ov::genai::decompose_scaled_dot_product_attention_for_whisper(model);
ov::genai::add_cross_attention_qk_scaled_scores_outputs_for_whisper(model);
}

utils::apply_slice_before_matmul_transformation(model);

compiled_model = core.compile_model(model, device, properties);
Expand Down
4 changes: 0 additions & 4 deletions src/cpp/src/whisper/pipeline.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,10 +94,6 @@ class WhisperPipeline::WhisperPipelineStatefulImpl : public WhisperPipeline::Whi
auto eos_token = m_generation_config.eos_token_id == -1 ? m_tokenizer.get_eos_token_id()
: m_generation_config.eos_token_id;
properties_copy.insert({"WHISPER_EOS_TOKEN", eos_token});

if (m_generation_config.word_timestamps) {
properties_copy.insert({"WHISPER_DECOMPOSE_SDPA", m_generation_config.word_timestamps});
}
} else {
compiled_model = core.compile_model(models_path / "openvino_encoder_model.xml", device, properties_copy);
}
Expand Down
Loading