File size: 4,151 Bytes
406e494 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 | """Kaggle notebook body: director -> TTS in one run, with two interpreters.
The clash: Ming pins transformers 4.46.3, Qwen3.5 needs 5.x, and one
site-packages cannot hold both. So transformers 5 is installed into a private
directory and the director runs with PYTHONPATH pointing at it, while the radio
stage keeps the pinned 4.46.3.
Order matters: Ming's dependencies are installed first and never touched again.
Stages:
1. Ming environment (repo, pinned deps, weights, bundle)
2. transformers 5 into /kaggle/working/tf5
3. fine-tune the director on the 160 scripts in the bundle (skipped if the
model is already on disk)
4. director writes one programme to director/script.json
5. radio stage renders it with the 4-step student and the teacher
"""
import os
import subprocess
import sys
import time
ROOT = "/kaggle/working"
BUNDLE = f"{ROOT}/webfm-eval"
TF5 = f"{ROOT}/tf5"
def log(msg: str) -> None:
print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
def sh(cmd: str, env: dict | None = None) -> int:
log(f"$ {cmd}")
rc = subprocess.run(cmd, shell=True, env=env).returncode
if rc != 0:
log(f"failed ({rc}): {cmd}")
return rc
def stage_ming() -> None:
if not os.path.exists(f"{ROOT}/Ming-3.0-tts"):
sh(f"git clone https://github.com/inclusionAI/Ming-3.0-tts.git {ROOT}/Ming-3.0-tts")
sh(f"git -C {ROOT}/Ming-3.0-tts checkout e999c39")
sh('pip install -q "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')
if not os.path.exists(f"{ROOT}/Ming-3.0-tts-0.5b"):
sh("python3 -c \"from huggingface_hub import snapshot_download as s; "
f"s('inclusionAI/Ming-3.0-tts-0.5B', local_dir=r'{ROOT}/Ming-3.0-tts-0.5b', repo_type='model')\"")
def stage_tf5() -> None:
if os.path.exists(f"{TF5}/transformers"):
log("transformers 5 already installed")
return
# Private target: never touches the pinned environment Ming uses. Only
# transformers and hub go in; everything else (torch, accelerate) resolves
# from the main interpreter through PYTHONPATH fallback.
sh(f"pip install -q --target {TF5} 'transformers>=5' 'huggingface_hub>=1.31,<2'")
out = subprocess.run([sys.executable, "-c",
"import transformers; print(transformers.__version__)"],
capture_output=True, text=True,
env=dict(os.environ, PYTHONPATH=TF5))
log(f"tf5 transformers: {out.stdout.strip() or out.stderr[-300:]}")
def main() -> None:
log("== stage 1: ming ==")
stage_ming()
from huggingface_hub import snapshot_download
snapshot_download("idefinity/webfm-eval", repo_type="dataset", local_dir=BUNDLE)
log("== stage 2: transformers 5 in a private dir ==")
stage_tf5()
env5 = dict(os.environ, WORK_ROOT=ROOT, PYTHONPATH=TF5)
model = f"{ROOT}/director/model"
if not os.path.exists(f"{model}/config.json"):
log("== stage 3: fine-tune the director ==")
rc = subprocess.run([sys.executable, f"{BUNDLE}/kaggle/director_sft.py"],
env=env5).returncode
log(f"train rc={rc}")
if rc != 0:
raise SystemExit("director training failed")
log("== stage 4: write the programme ==")
rc = subprocess.run([sys.executable, f"{BUNDLE}/kaggle/director_probe.py"],
env=dict(env5, MODE="script",
TOPIC=os.environ.get("TOPIC", "雨夜的末班公交"),
DIRECTOR_MODEL=model)).returncode
log(f"director rc={rc}")
script = f"{ROOT}/director/script.json"
if not os.path.exists(script):
raise SystemExit("no script.json")
log("== stage 5: render ==")
rc = subprocess.run([sys.executable, f"{BUNDLE}/kaggle/radio_demo.py"],
env=dict(os.environ, WORK_ROOT=ROOT, SCRIPT_FILE=script)).returncode
log(f"radio rc={rc}")
if __name__ == "__main__":
main()
|