[Perf] Avoid more GPU<->CPU syncs in multimodal encoders (#52369)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-08-14 10:40:14 -07:00
committed by GitHub
parent 9b0ab5dd53
commit 3e3ceb1961
9 changed files with 81 additions and 72 deletions
+5 -1
View File
@@ -244,7 +244,11 @@ class UltravoxMultiModalProcessor(BaseMultiModalProcessor[UltravoxProcessingInfo
# higher than the number of audio samples
audio_features=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
audio_token_len=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
audio_lens=MultiModalFieldConfig.flat_from_sizes("audio", num_chunks),
# Only ever used to derive the encoder attention metadata on the
# host, so keep it there.
audio_lens=MultiModalFieldConfig.flat_from_sizes(
"audio", num_chunks, keep_on_cpu=True
),
# num_chunks can convert audio_chunked to audio batch dimension
audio_num_chunks=MultiModalFieldConfig.batched("audio"),
audio_embeds=MultiModalFieldConfig.batched("audio"),