Fix HF Space cold-start timeout: slim deps + prebake small model

- Drop torch/torchaudio/transformers/datasets from runtime requirements;
faster-whisper uses CTranslate2, not PyTorch. Moved training-only deps
to ml/finetune-requirements.txt.
- Pre-download Systran/faster-whisper-small during image build so the
first /transcribe request doesn't trigger a multi-minute download on
the free CPU tier.
- Default MODEL_PATH to the HF model id instead of a local ../models path.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Dockerfile

@@ -9,12 +9,15 @@ RUN apt-get update \
 COPY ml/requirements.txt ./requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
-COPY ml/ ./ml/
+ENV HF_HOME=/app/.cache/huggingface
+RUN python -c "from huggingface_hub import snapshot_download; snapshot_download('Systran/faster-whisper-small')"
 
-RUN mkdir -p /app/models
+COPY ml/ ./ml/
 
 WORKDIR /app/ml
 
+ENV MODEL_PATH=Systran/faster-whisper-small
+
 EXPOSE 8000
 
 CMD ["uvicorn", "serve:app", "--host", "0.0.0.0", "--port", "8000"]

ml/finetune-requirements.txt

@@ -0,0 +1,7 @@
+-r requirements.txt
+transformers>=4.40.0
+datasets>=2.18.0
+torch>=2.2.0
+torchaudio>=2.2.0
+evaluate
+jiwer

ml/requirements.txt

@@ -1,10 +1,4 @@
 faster-whisper==1.1.0
-transformers>=4.40.0
-datasets>=2.18.0
-torch>=2.2.0
-torchaudio>=2.2.0
-evaluate
-jiwer
 fastapi
 uvicorn
 python-multipart

ml/serve.py

@@ -29,7 +29,7 @@ _MODEL_MAPPING = {
 }
 
 _DEFAULT_MODEL = "small"
-_MODEL_NAME = os.getenv("MODEL_PATH", "../models/small")
+_MODEL_NAME = os.getenv("MODEL_PATH", "Systran/faster-whisper-small")
 _DEVICE = os.getenv("WHISPER_DEVICE", "cpu")
 _COMPUTE_TYPE = os.getenv("WHISPER_COMPUTE_TYPE", "int8")
 _DEFAULT_LANGUAGE = os.getenv("WHISPER_LANGUAGE", "en")