bootstrap_models.py

Rama master · réplica de lectura del repositorio original.

#!/usr/bin/env python3
"""Pre-stage the offline Whisper model Hearthwave needs.

Hearthwave is designed for fully offline operation — the server never attempts
to fetch a model at runtime. Run this script once on a machine with internet
access; the resulting Models/ directory is portable and can be copied to
air-gapped target machines alongside the source tree.

Usage:
    python bootstrap_models.py                       # default: small.en
    python bootstrap_models.py --model base.en       # smaller, faster
    python bootstrap_models.py --model medium.en     # higher accuracy

    # Two-tier RX pipeline — stage the streaming model and the final-pass
    # model in one run. large-v3-turbo is what the install scripts stage by
    # default and what whisper_model_final="auto" prefers:
    python bootstrap_models.py --model small.en large-v3-turbo

    # GPU final pass (HF transformers format, staged at Models/STT/<name>-hf):
    python bootstrap_models.py --final-model large-v3-turbo --final-backend gpu
"""
import argparse
import os
import sys

WHISPER_REPOS = {
    "tiny.en":   "Systran/faster-whisper-tiny.en",
    "base.en":   "Systran/faster-whisper-base.en",
    "small.en":  "Systran/faster-whisper-small.en",
    "medium.en": "Systran/faster-whisper-medium.en",
    "large-v3":  "Systran/faster-whisper-large-v3",
    # Distilled large: ~large-v2 accuracy at a fraction of the compute —
    # the recommended whisper_model_final for the two-pass RX pipeline.
    "distil-large-v3": "Systran/faster-distil-whisper-large-v3",
    # Community CT2 conversion of openai/whisper-large-v3-turbo (no official
    # Systran repo exists). Verified 2026-07-01: downloads, loads in
    # faster-whisper 1.2.1, transcribes. Final-pass only — too slow for the
    # 2 s streaming slices.
    "large-v3-turbo": "deepdml/faster-whisper-large-v3-turbo-ct2",
}

# HF transformers-format repos for the GPU final pass (distinct from the CT2
# faster-whisper repos above). Staged under Models/STT/<name>-hf/.
HF_FINAL_REPOS = {
    "distil-large-v3": "distil-whisper/distil-large-v3",
    "large-v3": "openai/whisper-large-v3",
    "large-v3-turbo": "openai/whisper-large-v3-turbo",
}


def final_target(name: str, backend: str):
    """Return (repo_id, local_dir) for a final-pass model and backend.
    backend 'cpu' -> CT2 faster-whisper repo, dir Models/STT/<name>.
    backend 'gpu' -> HF transformers repo, dir Models/STT/<name>-hf."""
    if backend == "gpu":
        return HF_FINAL_REPOS[name], os.path.join("Models", "STT", name + "-hf")
    return WHISPER_REPOS[name], os.path.join("Models", "STT", name)


def main():
    parser = argparse.ArgumentParser(
        description=__doc__,
        formatter_class=argparse.RawDescriptionHelpFormatter,
    )
    parser.add_argument(
        "--model",
        default=["small.en"],
        nargs="+",
        choices=sorted(WHISPER_REPOS),
        metavar="MODEL",
        help="faster-whisper variant(s) to fetch (default: small.en). Pass "
             "more than one to stage the two-tier pipeline's streaming and "
             "final-pass models together.",
    )
    parser.add_argument("--final-model", metavar="NAME", default=None,
                        help="Stage a final-pass model in the format for --final-backend.")
    parser.add_argument("--final-backend", choices=("cpu", "gpu"), default="cpu",
                        help="Format for --final-model: cpu (CT2) or gpu (HF transformers).")
    args = parser.parse_args()

    try:
        from huggingface_hub import snapshot_download
    except ImportError:
        print(
            "huggingface_hub is required. Install it with:\n"
            "    pip install -r backend/requirements.txt",
            file=sys.stderr,
        )
        return 1

    # Preserve order but drop duplicates (e.g. --model small.en small.en).
    models = list(dict.fromkeys(args.model))
    for model in models:
        target = os.path.join("Models", "STT", model)
        os.makedirs(target, exist_ok=True)
        repo_id = WHISPER_REPOS[model]
        print(f"Whisper: downloading {repo_id} -> {target}")
        snapshot_download(repo_id=repo_id, local_dir=target)
        print(f"Whisper: done. Loaded at runtime from {target}/")

    if args.final_model:
        repo_id, target = final_target(args.final_model, args.final_backend)
        os.makedirs(target, exist_ok=True)
        print(f"Final ({args.final_backend}): downloading {repo_id} -> {target}")
        snapshot_download(repo_id=repo_id, local_dir=target)
        print(f"Final: done. Loaded at runtime from {target}/")
    return 0


if __name__ == "__main__":
    sys.exit(main())