whisper/convert.py outputs `model.safetensors` but `mlx_whisper.load_models` expects `weights.safetensors`

The pip-installed mlx-whisper package looks for weights.safetensors or
weights.npz, but the upstream whisper/convert.py script in this repo
writes the converted weights to model.safetensors. After running the
documented conversion command, mlx_whisper.transcribe(path_or_hf_repo=...)
fails to find the weights and crashes.

Repro

python3 whisper/convert.py \
    --torch-name-or-path nectec/Pathumma-whisper-th-large-v3 \
    --mlx-path /tmp/pathumma-mlx \
    --dtype float16

ls /tmp/pathumma-mlx
# config.json
# model.safetensors    <-- written by convert.py

import mlx_whisper
mlx_whisper.transcribe("audio.wav", path_or_hf_repo="/tmp/pathumma-mlx")
# FileNotFoundError / load failure: weights.safetensors not found

Source

• whisper/convert.py (this repo) writes model.safetensors:
  https://github.com/ml-explore/mlx-examples/blob/main/whisper/convert.py#L385
  mx.save_safetensors(str(mlx_path / "model.safetensors"), weights)

• mlx_whisper/load_models.py (PyPI mlx-whisper 0.4.3) expects weights.safetensors:

  wf = model_path / "weights.safetensors"
  if not wf.exists():
      wf = model_path / "weights.npz"
  weights = mx.load(str(wf))
  

Workaround

Rename after conversion:

mv /tmp/pathumma-mlx/model.safetensors /tmp/pathumma-mlx/weights.safetensors

Possible fix

Either:

• convert.py writes to weights.safetensors, or
• load_models.py also accepts model.safetensors

The latter seems more backwards-compatible since pre-converted models on the Hub
(e.g. mlx-community/whisper-large-v3-mlx) currently ship weights.npz.

Versions: mlx-whisper==0.4.3, mlx==0.31.2, mlx-examples HEAD e52c128.

Start with whisper/convert.py at the mx.save_safetensors call and compare its output name with the loading behavior described for mlx_whisper/load_models.py. Reproduce the documented conversion and transcription flow, then verify that the converted model is accepted without the manual rename workaround.