Update custom model files, README, and requirements

alignment.py

@@ -2,7 +2,6 @@
 
 from __future__ import annotations
 
-import itertools
 from collections.abc import Sequence
 from typing import TYPE_CHECKING, TypedDict
 
@@ -127,7 +126,9 @@ class QwenForcedAligner:
         if not todo:  # nothing to time: don't load the model
             return results
         model, processor = cls.get_instance()
-        for batch in itertools.batched(todo, cls.BATCH_SIZE):
+        # Slices, not itertools.batched (3.12+): the Hub copy runs on Python 3.11.
+        for start in range(0, len(todo), cls.BATCH_SIZE):
+            batch = todo[start : start + cls.BATCH_SIZE]
             features, word_lists = processor.prepare_forced_aligner_inputs(
                 [audio for _, audio, _ in batch],
                 [" ".join(kept) for _, _, kept in batch],

asr_processing.py

@@ -1,7 +1,7 @@
 """Processor that turns raw audio (and optional text) into model inputs."""
 
 from collections.abc import Mapping, Sequence
-from typing import TYPE_CHECKING, Any, ClassVar, cast, overload
+from typing import TYPE_CHECKING, Any, ClassVar, TypeVar, cast, overload
 
 import numpy as np
 import numpy.typing as npt
@@ -111,12 +111,15 @@ def left_pad_prompt_rows(
     return input_ids, attention_mask
 
 
+ScalarT = TypeVar("ScalarT", bound=np.generic)
+
+
 @overload
-def prepend_lead_in[ScalarT: np.generic](
+def prepend_lead_in(
     audio: npt.NDArray[ScalarT], sampling_rate: int, seconds: float | None
 ) -> npt.NDArray[ScalarT]: ...
 @overload
-def prepend_lead_in[ScalarT: np.generic](
+def prepend_lead_in(
     audio: list[npt.NDArray[ScalarT]], sampling_rate: int, seconds: float | None
 ) -> list[npt.NDArray[ScalarT]]: ...
 @overload

requirements.txt

@@ -5,11 +5,16 @@
 # Speaker diarization (return_speakers=True) uses Nemotron-3-Diarization, which is in
 # transformers main after 5.17 -- pinned to the commit the pipeline was verified on.
 transformers @ git+https://github.com/huggingface/transformers@91ea1192a00e0ebb684400434375fbd2345c8449
-# Pinned floor, not just "torch": the Inference Endpoints image ships an older
-# torch that satisfies a bare requirement, and transformers disables PyTorch
-# below 2.5 ("Could not import module 'WhisperForConditionalGeneration'").
-torch>=2.5
-torchaudio>=2.5
+# The Inference Endpoints image ships an older torch, torchvision and torchaudio.
+# Pin all three as one matching set (torch 2.8, as in poetry.lock): a bare
+# "torch" keeps the image's version, which this transformers rejects (< 2.5),
+# and upgrading torch alone leaves the image's torchvision built for the old
+# one, so transformers' torchvision import fails ("operator torchvision::nms
+# does not exist", surfacing as "Could not import module
+# 'WhisperForConditionalGeneration'").
+torch==2.8.0
+torchvision==0.23.0
+torchaudio==2.8.0
 peft
 safetensors
 numpy