gasoonjia/Whisper-Small-ExecuTorch-Metal

Whisper Small — ExecuTorch Export

This repository contains an ExecuTorch .pte export of the OpenAI Whisper model openai/whisper-small for:

• Metal (macOS): bf16

Artifacts

File	Description
model.pte	ExecuTorch program (main model)
whisper_preprocessor.pte	ExecuTorch program (audio preprocessor / mel spectrogram)
tokenizer.json	Tokenizer
tokenizer_config.json	Tokenizer configuration
special_tokens_map.json	Special tokens mapping

Model Details

• Base Model: openai/whisper-small
• Backend: Metal (MPS)
• Precision: bf16
• Parameters: 244M
• Languages: Multilingual (99 languages)

How to Run (ExecuTorch)

Build the Whisper runner from an ExecuTorch checkout:

EXECUTORCH_BUILD_KERNELS_TORCHAO=1 TORCHAO_BUILD_EXPERIMENTAL_MPS=1 ./install_executorch.sh
make whisper-metal

Then run inference:

Metal (macOS)

./cmake-out/examples/models/whisper/whisper_runner \
  --model_path /path/to/model.pte \
  --preprocessor_path /path/to/whisper_preprocessor.pte \
  --audio_path /path/to/audio.wav \
  --tokenizer_path /path/to/tokenizer.json

Verification

The model was verified to transcribe test audio correctly and match the original PyTorch model output.

License / Attribution

• Upstream model license: MIT (see openai/whisper-small).
• Please attribute OpenAI when using these exported artifacts.