Blame for kaggle/notebook_eval.py

"""Kaggle notebook body for the WebFM eval run. Not imported by the app.

Pulled and executed by the kernel `kkonda/webfm-ming-tts-eval`, which does
nothing but snapshot_download idefinity/webfm-eval and exec this file. The run
logic lives here so it can change without touching the kernel.

Stages, in order:

  1. install the pinned runtime, then prove `from ming_tts import MingTTS`
     imports — repairing missing modules by name, because guessing the
     transitive dependency set is how the first run died on `loguru`
  2. clone the repo at the pinned commit, download weights, pull the bundle
  3. bake prompt_latent / speaker_emb for the eight WebFM voices
  4. smoke: one voice, two sentences, teacher + 3step. Cheaper to find API
     drift here than 40 minutes into the matrix
  5. the full matrix, one voice shard per visible GPU
  6. score

Everything writes into /kaggle/working; ming_eval.py skips cells that already
have a row, so a killed session keeps its progress.
"""
import os
import re
import subprocess
import sys
import time
from pathlib import Path

ROOT = Path("/kaggle/working")
REPO = ROOT / "Ming-3.0-tts"
MODEL = ROOT / "Ming-3.0-tts-0.5b"
BUNDLE = ROOT / "webfm-eval"
PIN = "e999c39"
WEIGHTS = "inclusionAI/Ming-3.0-tts-0.5B"
BUNDLE_REPO = "idefinity/webfm-eval"
MODELS = "teacher,3step,4step"
HF_MIRROR = "https://hf-mirror.com"

# requirements.txt at PIN, minus torch/torchaudio: Kaggle ships a newer CUDA
# build and downgrading it would cost the session.
DEPS = [
    "transformers==4.46.3", "tokenizers", "accelerate==1.3.0", "huggingface_hub",
    "soundfile==0.12.1", "tqdm", "loguru", "x_transformers==2.9.2",
    "diffusers==0.33.0", "onnxruntime", "onnx", "onnx_ir", "onnxscript",
    "inflect", "einops", "packaging",
]
ALIASES = {"cv2": "opencv-python", "yaml": "pyyaml", "PIL": "pillow", "sklearn": "scikit-learn"}


def sh(cmd: str, check: bool = True, env: dict | None = None) -> int:
    print(f"\n$ {cmd}", flush=True)
    rc = subprocess.run(cmd, shell=True, text=True, env=env).returncode
    if check and rc != 0:
        raise SystemExit(f"command failed ({rc}): {cmd}")
    return rc


def pip(*pkgs: str) -> None:
    sh("pip install -q " + " ".join(f'"{p}"' for p in pkgs), check=False)


def pull(repo_id: str, local_dir: Path, repo_type: str = "model") -> None:
    if local_dir.exists() and any(local_dir.iterdir()):
        print(f"{local_dir} already populated")
        return
    code = ("from huggingface_hub import snapshot_download as s; "
            f"s('{repo_id}', local_dir=r'{local_dir}', repo_type='{repo_type}')")
    if sh(f'python3 -c "{code}"', check=False) != 0:
        print("direct download failed, retrying through hf-mirror")
        sh(f'python3 -c "{code}"', env=dict(os.environ, HF_ENDPOINT=HF_MIRROR))


def stage_repo() -> None:
    """Clone before the import probe, which imports from this directory."""
    if not REPO.exists():
        sh(f"git clone https://github.com/inclusionAI/Ming-3.0-tts.git {REPO}")
    if sh(f"git -C {REPO} checkout {PIN}", check=False) != 0:
        print(f"WARNING: commit {PIN} not found, staying on the default branch")


def stage_deps() -> None:
    sh("nvidia-smi --query-gpu=name,memory.total --format=csv,noheader")
    import torch

    print("torch", torch.__version__, "cuda", torch.cuda.is_available(),
          "devices", torch.cuda.device_count())

    pip(*DEPS)
    probe = ("import sys; sys.path.insert(0, r'%s'); from ming_tts import MingTTS; "
             "print('import ok')" % REPO)
    for attempt in range(6):
        r = subprocess.run([sys.executable, "-c", probe], capture_output=True, text=True)
        if r.returncode == 0:
            print("ming_tts imports cleanly")
            return
        m = re.search(r"No module named '([^']+)'", r.stderr)
        if not m:
            print(r.stderr[-3000:])
            raise SystemExit("ming_tts import failed for a reason pip cannot repair")
        missing = m.group(1).split(".")[0]
        pkg = ALIASES.get(missing, missing)
        print(f"[{attempt+1}] missing '{missing}' -> installing '{pkg}'")
        pip(pkg)
    raise SystemExit("could not satisfy ming_tts imports")


def stage_prepare() -> None:
    pull(WEIGHTS, MODEL)
    pull(BUNDLE_REPO, BUNDLE, repo_type="dataset")


def stage_bake() -> None:
    if (ROOT / "baked/voices_baked.npz").exists():
        print("voices already baked, skipping")
        return
    sh(f"cd {ROOT} && EVAL_DATA={BUNDLE} python3 {BUNDLE}/kaggle/bake_voices.py")


def run_eval(tag: str, voices: list[str], models: str, limit: int, gpu: int = 0) -> subprocess.Popen:
    env = dict(os.environ, CUDA_VISIBLE_DEVICES=str(gpu), EVAL_DATA=str(BUNDLE),
               VOICES=",".join(voices), MODELS=models, LIMIT=str(limit))
    log = open(ROOT / f"eval_{tag}.log", "w")
    return subprocess.Popen([sys.executable, str(BUNDLE / "kaggle/ming_eval.py")],
                            env=env, stdout=log, stderr=subprocess.STDOUT, cwd=str(ROOT))


def stage_smoke() -> None:
    """One voice, two sentences. Catches API drift before the long run."""
    if (ROOT / "smoke.ok").exists():
        return
    p = run_eval("smoke", ["v01"], "teacher,3step", 2)
    p.wait()
    rows = ROOT / "ming_eval/results.jsonl"
    n = sum(1 for _ in rows.open()) if rows.exists() else 0
    if n < 4:
        print((ROOT / "eval_smoke.log").read_text()[-4000:])
        raise SystemExit(f"smoke produced {n} rows, expected at least 4")
    (ROOT / "smoke.ok").write_text("ok")
    print(f"smoke ok ({n} rows)")


def stage_eval() -> None:
    import torch

    n_gpu = max(1, torch.cuda.device_count())
    voices = [f"v{i:02d}" for i in range(1, 9)]
    shards = [voices[i::n_gpu] for i in range(n_gpu)]
    print(f"{n_gpu} GPU(s), shards: {shards}")

    procs = [(gpu, run_eval(f"gpu{gpu}", shard, MODELS, 40, gpu))
             for gpu, shard in enumerate(shards)]
    last = 0.0
    while any(p.poll() is None for _, p in procs):
        time.sleep(60)
        rows = ROOT / "ming_eval/results.jsonl"
        n = sum(1 for _ in rows.open()) if rows.exists() else 0
        if n != last:
            print(f"[{time.strftime('%H:%M')}] rows={n}", flush=True)
            last = n
    for gpu, p in procs:
        print(f"gpu{gpu} exit={p.returncode}")


def stage_score() -> None:
    sh(f"python3 {BUNDLE}/eval/score.py aggregate "
       f"--results {ROOT}/ming_eval/results.jsonl --baseline teacher --by-bucket "
       f"--out {ROOT}/ming_eval/summary.json", check=False)


if __name__ == "__main__":
    stage_repo()
    stage_deps()
    stage_prepare()
    stage_bake()
    stage_smoke()
    stage_eval()
    stage_score()
    print("\nartifacts: baked/, ming_eval/results.jsonl, ming_eval/summary.json, ming_eval/wav/")