[Bugfix][Voxtral] Add fetch_audio to MistralCommonFeatureExtractor (transformers>=5.10 compat) (#44559)

Signed-off-by: Yadan Wei <weiyadan@amazon.com>
Co-authored-by: Yadan Wei <weiyadan@amazon.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Yadan-Wei
2026-06-06 07:58:09 +00:00
committed by GitHub
co-authored by Yadan Wei Claude
parent f87df1df9e
commit c9b4b184b4
3 changed files with 177 additions and 0 deletions
@@ -53,6 +53,54 @@ class MistralCommonFeatureExtractor:
def get_num_audio_tokens(self, audio_length: int) -> int:
return ceil(audio_length / (self.sampling_rate // self.frame_rate))
def fetch_audio(self, audio_url_or_urls, sampling_rate=None):
"""HF-compatible duck-typed ``fetch_audio``.
Mirrors :meth:`transformers.SequenceFeatureExtractor.fetch_audio` so
:class:`transformers.ProcessorMixin.prepare_inputs_layout` (added in
transformers 5.10) works on this duck-typed feature extractor. Older
transformers versions never invoke this method, so the addition is a
no-op there.
Accepts the same shapes as ``SequenceFeatureExtractor.fetch_audio``:
* ``np.ndarray`` / ``torch.Tensor`` — returned as-is.
* ``list[float]`` — returned as-is (a single audio sample).
* ``str`` URL or path — delegated to
:func:`transformers.audio_utils.load_audio`.
* ``list`` of any of the above — recursed element-wise.
``ProcessorMixin.prepare_inputs_layout`` always passes already-decoded
audio (numpy array or torch tensor), so the str / list-of-str branches
exist only to keep the contract identical to the upstream method.
The semantics of ``transformers.audio_utils.is_valid_audio`` differ
between transformers versions (5.9 only accepts ndarray/tensor; 5.10
also accepts ``list[float]``). We detect ``list[float]`` explicitly to
keep behavior identical across versions.
"""
from transformers.audio_utils import is_valid_audio
sampling_rate = sampling_rate if sampling_rate else self.sampling_rate
if is_valid_audio(audio_url_or_urls):
return audio_url_or_urls
if isinstance(audio_url_or_urls, (list, tuple)):
if audio_url_or_urls and isinstance(audio_url_or_urls[0], float):
# A single audio represented as ``list[float]``.
return audio_url_or_urls
return [
self.fetch_audio(x, sampling_rate=sampling_rate)
for x in audio_url_or_urls
]
if isinstance(audio_url_or_urls, str):
from transformers.audio_utils import load_audio
return load_audio(audio_url_or_urls, sampling_rate=sampling_rate)
raise TypeError(
"only a numpy array, torch tensor, str URL/path, or list of those "
f"is supported but got type={type(audio_url_or_urls)}"
)
class MistralCommonVoxtralProcessor(ProcessorMixin):
attributes = ["feature_extractor", "tokenizer"]