[Bugfix][Gemma 4] Don't read fft_length when profiling unified audio (#56721)

Signed-off-by: Abhinav Verma <verma.abhinav275@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Abhinav Verma
2026-09-15 14:57:32 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent f0fbc98c48
commit 1e47ec00d2
2 changed files with 67 additions and 2 deletions
+24 -1
View File
@@ -24,6 +24,7 @@ from PIL import Image as PILImage
from torch import nn
from transformers import AutoModel, BatchFeature
from transformers.models.gemma4 import (
Gemma4AudioFeatureExtractor,
Gemma4Config,
Gemma4Processor,
Gemma4VisionConfig,
@@ -456,6 +457,28 @@ class Gemma4ProcessingInfo(BaseProcessingInfo):
# ---------------------------------------------------------------------------
def _dummy_audio_num_samples(processor: Any) -> int:
"""Length in samples of a worst-case dummy audio item.
``fft_length`` is set in ``Gemma4AudioFeatureExtractor.__init__`` and is not
defined by ``Gemma4UnifiedAudioFeatureExtractor``, which the encoder-free
Gemma 4 variant uses. Reading it unconditionally therefore raises
``AttributeError`` during memory profiling on that variant, which is reached
whenever audio is the modality with the largest per-item token count.
For any other extractor, fall back to the audio budget the processor itself
declares -- the same pair of values used to warn about over-long audio
elsewhere in this file.
"""
feature_extractor = processor.feature_extractor
if isinstance(feature_extractor, Gemma4AudioFeatureExtractor):
return feature_extractor.fft_length
sampling_rate = feature_extractor.sampling_rate
max_duration_s = processor.audio_seq_length * processor.audio_ms_per_token / 1000.0
return int(max_duration_s * sampling_rate)
class Gemma4DummyInputsBuilder(BaseDummyInputsBuilder[Gemma4ProcessingInfo]):
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
num_images = mm_counts.get("image", 0)
@@ -508,7 +531,7 @@ class Gemma4DummyInputsBuilder(BaseDummyInputsBuilder[Gemma4ProcessingInfo]):
}
if num_audios > 0:
audio_len = processor.feature_extractor.fft_length
audio_len = _dummy_audio_num_samples(processor)
data["audio"] = self._get_dummy_audios(
length=audio_len,
num_audios=num_audios,