emo4_train_wrapper.sh

#!/bin/bash
set -e
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
echo '== SETUP =='
apt-get update -qq && apt-get install -y -qq git build-essential ffmpeg >/dev/null 2>&1
git clone -q https://github.com/vibevoice-community/VibeVoice /VibeVoice-classic
git -C /VibeVoice-classic checkout -q d5895e7
git clone -q https://github.com/vibevoice-community/VibeVoice /vv-main
git clone -q https://github.com/voicepowered-ai/VibeVoice-finetuning /vvft
cd /vvft && pip install -q -e . 2>&1 | tail -1
pip uninstall -y -q transformers 2>&1 | tail -1
pip install -q 'setuptools<81' 'torch==2.6.0' 'transformers==4.51.3' 'peft==0.15.2' 'diffusers==0.29.0' 'safetensors==0.5.3' 'datasets==3.6.0' 'resampy==0.4.3' 'librosa==0.11.0' 'numpy<2' 'resemble-perth==1.0.1' 'omegaconf==2.3.0' 'conformer==0.3.2' 's3tokenizer' 'soundfile' 'jiwer' 'resemblyzer' 2>&1 | tail -1
echo '== PREFLIGHT =='
for b in git gcc ffmpeg; do command -v $b >/dev/null || { echo "PREFLIGHT FAIL: missing binary $b"; exit 2; }; done
python - <<'PF'
import importlib
from huggingface_hub import HfApi
for m, v in [("torch", "2.6.0"), ("transformers", "4.51.3"), ("peft", "0.15.2"), ("diffusers", "0.29.0"), ("resemblyzer", None), ("webrtcvad", None)]:
    mod = importlib.import_module(m)
    if v: assert mod.__version__.split('+')[0] == v, f"{m} {mod.__version__} != {v}"
api = HfApi()
files = api.list_repo_files("ACloudCenter/vibevoice-lora-joint", repo_type="model")
for need in ("emo3/lora/adapter_config.json", "emo3/lora/adapter_model.safetensors", "emo3/lora/diffusion_head_full.bin"):
    assert need in files, f"{need} missing (emo3 warm-start source)"
assert api.repo_exists("ACloudCenter/fast-vibevoice-notes", repo_type="dataset"), "notes missing"
print("PREFLIGHT OK", flush=True)
PF
echo '== DAILYTALK PREP (500 convs) =='
python -c "import urllib.request; urllib.request.urlretrieve('https://huggingface.co/datasets/ACloudCenter/fast-vibevoice-notes/resolve/main/prep_dailytalk.py', '/prep_dailytalk.py'); s=open('/prep_dailytalk.py').read(); assert 'PREP DONE' in s; print('PREP FETCH OK')"
python /prep_dailytalk.py /data 500
echo '== EMO4 DATA PREP (clean-filtered) =='
python -c "import urllib.request; urllib.request.urlretrieve('https://huggingface.co/datasets/ACloudCenter/fast-vibevoice-notes/resolve/main/emo4_data_prep.py', '/emo4_data_prep.py'); s=open('/emo4_data_prep.py').read(); assert 'EMO4 PREP DONE' in s and 'FLOOR_MAX' in s; print('EMO4 FETCH OK')"
python /emo4_data_prep.py
echo '== FETCH EMO3 (warm-start init) =='
python - <<'CK'
import os
from huggingface_hub import HfApi, hf_hub_download
api = HfApi()
sub = "emo3/lora"
for f in [f for f in api.list_repo_files("ACloudCenter/vibevoice-lora-joint", repo_type="model") if f.startswith(sub + "/")]:
    local = hf_hub_download("ACloudCenter/vibevoice-lora-joint", f, repo_type="model")
    with open(local, "rb") as fh:
        data = fh.read()
    op = os.path.join("/emo3-init", f[len(sub) + 1:])
    os.makedirs(os.path.dirname(op), exist_ok=True)
    with open(op, "wb") as fh:
        fh.write(data)
    assert os.path.isfile(op) and os.path.getsize(op) == len(data)
print("EMO3 FETCH OK:", sorted(os.listdir("/emo3-init")), flush=True)
CK
echo '== WARMSTART PATCH (from emo3) =='
python - <<'WS'
p = "/vvft/src/finetune_vibevoice_lora.py"
s = open(p).read()
a1 = "        model.model.language_model = get_peft_model(model.model.language_model, lora_cfg)"
assert a1 in s, "lm anchor missing"
s = s.replace(a1, a1 + """
        if os.path.isfile("/emo3-init/adapter_config.json"):
            import importlib.util as _ilu2
            import sys as _sys2
            from pathlib import Path as _P
            _spec2 = _ilu2.spec_from_file_location("lora_loading_ws", "/vv-main/vibevoice/modular/lora_loading.py")
            _mod2 = _ilu2.module_from_spec(_spec2)
            _sys2.modules[_spec2.name] = _mod2
            _spec2.loader.exec_module(_mod2)
            _rep = _mod2._LoadReport()
            _dev = next(model.parameters()).device
            _mod2._load_language_model(model, _P("/emo3-init"), _dev, _rep)
            print(f"WARMSTART LM via peft loader: lm={_rep.language_model}", flush=True)""")
a2 = """    if getattr(model_args, "train_diffusion_head", False) and hasattr(model.model, "prediction_head"):
        for p in model.model.prediction_head.parameters():
            p.requires_grad = True"""
assert a2 in s, "head anchor missing"
s = s.replace(a2, a2 + """
        if os.path.isfile("/emo3-init/diffusion_head_full.bin"):
            _hsd = torch.load("/emo3-init/diffusion_head_full.bin", map_location="cpu")
            _hm, _hu = model.model.prediction_head.load_state_dict(_hsd, strict=False)
            print(f"WARMSTART HEAD: missing={len(_hm)} unexpected={len(_hu)}", flush=True)
        for _cname in ("acoustic_connector", "semantic_connector"):
            _cp = f"/emo3-init/{_cname}/pytorch_model.bin"
            if os.path.isfile(_cp):
                getattr(model.model, _cname).load_state_dict(torch.load(_cp, map_location="cpu"), strict=False)
                print(f"WARMSTART {_cname} OK", flush=True)""")
open(p, "w").write(s)
print("WARMSTART PATCH OK (from emo3)", flush=True)
WS
echo '== TRAIN (emo4: 1600 steps, clean data) =='
python -m src.finetune_vibevoice_lora \
  --model_name_or_path microsoft/VibeVoice-1.5B \
  --processor_name_or_path src/vibevoice/processor \
  --train_jsonl /data/emo4_train.jsonl --validation_jsonl /data/emo4_val.jsonl \
  --output_dir out_emo4 --per_device_train_batch_size 1 --gradient_accumulation_steps 16 \
  --max_steps 1600 --learning_rate 1e-5 --logging_steps 10 --save_steps 400 --save_total_limit 2 \
  --eval_strategy steps --eval_steps 400 --do_eval \
  --load_best_model_at_end --metric_for_best_model eval_loss --greater_is_better False \
  --label_names input_ids \
  --remove_unused_columns False --bf16 True --do_train --gradient_clipping \
  --gradient_checkpointing True --gradient_checkpointing_kwargs '{"use_reentrant": false}' \
  --ddpm_batch_mul 2 --diffusion_loss_weight 1.4 --train_diffusion_head True --ce_loss_weight 0.04 \
  --voice_prompt_drop_rate 0.2 \
  --lora_target_modules q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj \
  --lr_scheduler_type cosine --warmup_ratio 0.03 --max_grad_norm 0.8 --report_to none \
  2>&1 | tee /tmp/joblog.txt
echo '== PUSH =='
python - <<'PUSH'
from huggingface_hub import HfApi
import glob, os
api = HfApi()
n = 0
for p in glob.glob('out_emo4/**/*', recursive=True):
    if os.path.isfile(p):
        api.upload_file(path_or_fileobj=p, path_in_repo='emo4/' + os.path.relpath(p, 'out_emo4'), repo_id='ACloudCenter/vibevoice-lora-joint')
        n += 1
print(f'PUSH OK ({n} files)', flush=True)
PUSH
python - <<'LOG'
from huggingface_hub import HfApi
try:
    HfApi().upload_file(path_or_fileobj="/tmp/joblog.txt", path_in_repo="emo4-train-log.txt", repo_id="ACloudCenter/fast-vibevoice-notes", repo_type="dataset")
    print("LOG OK", flush=True)
except Exception as e:
    print("LOG FAIL", e, flush=True)
LOG