chunk parakeet into 30s clips to prevent OOMs on long audios (#36671)

Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>
This commit is contained in:
Netanel Haber
2026-03-18 14:22:24 -07:00
committed by GitHub
parent a913b612d8
commit 6ae4c8d6fc
3 changed files with 72 additions and 64 deletions
@@ -845,7 +845,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
audios: list[npt.NDArray],
) -> tuple[list[str], dict[str, Any]]:
if len(audios) == 0:
return text, {}
return text, {"audio_num_clips": []}
assert self.audio_extractor is not None
extractor = self.audio_extractor
@@ -869,13 +869,10 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
sampling_rate=extractor.sampling_rate,
return_tensors="pt",
)
input_audio_features = audio_inputs.input_features
feature_attention_mask = audio_inputs.attention_mask
audio_feature_lengths = feature_attention_mask.sum(dim=1)
audio_inputs = {
"input_audio_features": input_audio_features,
"feature_attention_mask": feature_attention_mask,
"audio_feature_lengths": audio_feature_lengths,
"input_audio_features": audio_inputs.input_features,
"feature_attention_mask": audio_inputs.attention_mask,
"audio_num_clips": audio_inputs.audio_num_clips,
}
return text, audio_inputs