Datasets:

idefinity
/

webfm-eval

Languages:
Chinese
Tags:
text-to-speech
evaluation
webfm
License:
Dataset card Files Files and versions
xet
Community
webfm-eval / kaggle /train_voices.py
zziang
release candidate: 4000 steps, seeded teacher, keep audio
4457a96 verified 4 days ago
Raw History Blame Contribute Delete
25.6 kB
"""Distil the Ming flow head on the seven mined voices.
Why this run: the existing 3/4-step students were trained on 200 utterances of a
single voice (`lingguang`). On a clean prompt at eval time they cost +5.2 pp
(v08, 4step) to +11.8 pp (3step) CER against the teacher. The hypothesis is that
the gap is voice generalisation, not step count, and the fix is trajectory data
that spans the voices we actually ship.
One variable changes this round: the data. The training recipe is copied from
ming_multistep.py unchanged, including the CFG coefficient, so the new student
is comparable to the old one. If multi-voice data does not close the gap, that
coefficient is the next thing to look at.
Stages:
  STAGE=traj   shard by VOICES env, write traj/{voice}/{i}.pt, resumable
  STAGE=train  merge shards, train the 4-step CFG-free student
  STAGE=export ONNX fp32 + int4, validate against PyTorch
  STAGE=eval   40 sentences x EVAL_VOICES x {teacher, student}
  STAGE=all    train -> export -> eval (traj is run separately, per GPU)
"""
import json
import os
import sys
import time
import traceback
from pathlib import Path

import numpy as np

ROOT = Path(os.environ.get("WORK_ROOT", "/kaggle/working"))
REPO = ROOT / "Ming-3.0-tts"
MODEL = ROOT / "Ming-3.0-tts-0.5b"
BUNDLE = ROOT / "webfm-eval"
OUT = ROOT / "train"
TRAJ = OUT / "traj"

STAGE = os.environ.get("STAGE", "all")
SEED = 20261030
STEPS = int(os.environ.get("STEPS", "4"))
TRAIN_STEPS = int(os.environ.get("TRAIN_STEPS", "1500"))
BATCH = int(os.environ.get("BATCH", "16"))
LR = float(os.environ.get("LR", "1e-4"))
VOICE_FILTER = [v for v in os.environ.get("VOICES", "").split(",") if v]
EVAL_VOICES = [v for v in os.environ.get("EVAL_VOICES", "").split(",") if v]
HOLDOUT = {v for v in os.environ.get("HOLDOUT", "").split(",") if v}
N_TEXTS = int(os.environ.get("N_TEXTS", "120"))
SMOKE = os.environ.get("SMOKE", "0") == "1"
SAVE_WAVS = os.environ.get("SAVE_WAVS", "0") == "1"

TEACHER_CFG = 2.0
CFG_MULT = 2.0          # unchanged from ming_multistep.py, for comparability
DROP = {"Miss哈尼"}


def log(msg: str) -> None:
    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)


def patch_fp32() -> None:
    for name in ("ming_tts.py", "modeling_bailingmm.py"):
        path = REPO / name
        s = path.read_text()
        s = s.replace("torch_dtype=torch.bfloat16", "torch_dtype=torch.float32")
        s = s.replace(".eval().to(torch.bfloat16).to(device)", ".eval().to(torch.float32).to(device)")
        s = s.replace("dtype=torch.bfloat16", "dtype=torch.float32")
        path.write_text(s)


def voices(include_holdout: bool = False) -> list[dict]:
    vs = json.loads((BUNDLE / "mined/manifest.json").read_text())["voices"]
    vs = [v for v in vs if v["artist"] not in DROP]
    if not include_holdout:
        vs = [v for v in vs if v["slot"] not in HOLDOUT]
    if VOICE_FILTER:
        vs = [v for v in vs if v["slot"] in VOICE_FILTER]
    return vs


def texts() -> list[str]:
    t = json.loads((BUNDLE / "train_texts.json").read_text())["texts"]
    return t[:N_TEXTS]


def stage_traj() -> None:
    import torch

    sys.path.insert(0, str(REPO))
    from ming_tts import MingTTS
    import tts_tasks

    TRAJ.mkdir(parents=True, exist_ok=True)
    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
    mm = tts.model
    device = next(mm.parameters()).device
    dtype = next(mm.parameters()).dtype
    task_prompt = tts_tasks.build_tts_prompt("Chinese")
    torch.set_grad_enabled(False)

    total = 0
    for v in voices():
        vdir = TRAJ / v["slot"]
        vdir.mkdir(parents=True, exist_ok=True)
        prompt_latent, spk = tts.prepare_prompt(str(BUNDLE / v["wav"]), use_spk_emb=True)
        prompt_latent = prompt_latent.to(device=device, dtype=dtype)
        spk = [x.to(device=device, dtype=dtype) for x in spk]

        for i, text in enumerate(texts()):
            dst = vdir / f"{i:04d}.pt"
            if dst.exists():
                continue
            try:
                torch.manual_seed(SEED + i)
                input_ids, inputs_embeds = mm.prepare_input_embed(
                    prompt=task_prompt, text=text, spk_emb=spk,
                    prompt_latent=prompt_latent, prompt_text=v["prompt_text"],
                )
                attention_mask = torch.ones(input_ids.shape, device=device)
                position_ids = (attention_mask.cumsum(-1) - 1).masked_fill_(attention_mask == 0, 1)
                past = None
                hist = torch.zeros(1, mm.history_patch_size, mm.latent_dim,
                                   device=device, dtype=dtype)
                start = mm.history_patch_size - prompt_latent.size(1)
                if start < 0:
                    hist[:] = prompt_latent[:, -start:, :]
                else:
                    hist[:, start:, :] = prompt_latent

                budget = (512 - inputs_embeds.shape[1]) // mm.patch_size
                n = max(1, min(200, max(16, int(len(text) * 2.6)), budget))
                zs, hs, ns, ls = [], [], [], []
                for _ in range(n):
                    with torch.no_grad():
                        out = mm.model(
                            attention_mask=attention_mask, position_ids=position_ids,
                            inputs_embeds=inputs_embeds, audio_mask=None, image_mask=None,
                            output_hidden_states=True, return_dict=True, use_cache=True,
                            past_key_values=past,
                        )
                    past = out.past_key_values
                    z = out.hidden_states[-1][:, -1:, :]
                    noise = torch.randn(1, 1, mm.latent_dim, device=device, dtype=dtype)
                    sde = torch.randn(10, 1, 1, mm.latent_dim, device=device, dtype=dtype)
                    with torch.no_grad():
                        latent = mm.flowloss.cfm.sample(
                            noise=noise, c=z, latent_history=hist, sde_noise=sde,
                            cfg_scale=TEACHER_CFG, sway_sampling_coef=-1.0, use_epss=True,
                            patch_size=1, sigma=0.25, temperature=0.0,
                        )
                    zs.append(z[0, 0].detach().float().cpu())
                    hs.append(hist[0].detach().float().cpu())
                    ns.append(noise[0, 0].detach().float().cpu())
                    ls.append(latent[0, 0].detach().float().cpu())
                    inputs_embeds = mm.linear_proj_audio(latent)
                    position_ids = position_ids[:, -1:] + 1
                    attention_mask = torch.ones(inputs_embeds.shape[0], 1, device=device)
                    hist[:, :-mm.patch_size, :] = hist[:, mm.patch_size:, :].clone()
                    hist[:, -mm.patch_size:, :] = latent

                torch.save({"text": text, "voice": v["slot"],
                            "z": torch.stack(zs), "history": torch.stack(hs),
                            "noise": torch.stack(ns), "latent": torch.stack(ls)}, dst)
                total += n
                if i % 10 == 0:
                    log(f"{v['slot']} {i+1}/{len(texts())} frames={n} total={total}")
            except Exception:
                log(f"{v['slot']} {i} failed\n{traceback.format_exc()}")
            if SMOKE and i >= 4:
                break
    log(f"traj done: {total} frames")


def load_dataset():
    import torch

    cache = BUNDLE / "traj.zip"
    if not any(TRAJ.glob("*/*.pt")) and cache.exists():
        import zipfile
        log(f"unpacking {cache} ({cache.stat().st_size // 1024 // 1024} MB)")
        with zipfile.ZipFile(cache) as z:
            z.extractall(ROOT / "train")

    zs, hs, ns, ls = [], [], [], []
    for path in sorted(TRAJ.glob("*/*.pt")):
        row = torch.load(path, map_location="cpu", weights_only=False)
        zs.append(row["z"]); hs.append(row["history"])
        ns.append(row["noise"]); ls.append(row["latent"])
    if not zs:
        raise RuntimeError("no trajectories")
    z, h, n, l = (torch.cat(x).float() for x in (zs, hs, ns, ls))
    log(f"dataset: {z.shape[0]} frames from {len(zs)} files")
    return z, h, n, l


def train_student(teacher, data, n_steps, device):
    import copy
    import torch
    import torch.nn.functional as F

    z, h, n, target = [x.to(device) for x in data]
    count = z.shape[0]
    student = copy.deepcopy(teacher).to(device).train()
    for p in student.parameters():
        p.requires_grad_(True)
    opt = torch.optim.AdamW(student.parameters(), lr=LR, weight_decay=0.01)
    times = torch.linspace(0, 1, n_steps + 1).tolist()
    losses = []
    for step in range(TRAIN_STEPS):
        idx = torch.randint(0, count, (min(BATCH, count),), device=device)
        bz, bh = z[idx].unsqueeze(1), h[idx]
        bn, bt = n[idx].unsqueeze(1), target[idx].unsqueeze(1)
        y = bn
        loss = 0.0
        for i in range(n_steps):
            ti, dt = times[i], times[i + 1] - times[i]
            with torch.no_grad():
                pred_cfg = teacher.forward_with_cfg(
                    x=y, t=torch.tensor(float(ti), device=device), c=bz,
                    latent_history=bh, patch_size=1,
                )
                pred, null = torch.chunk(pred_cfg, 2, dim=0)
                tv = pred + (pred - null) * CFG_MULT
            sv = student(
                x=y, t=torch.full((y.shape[0],), float(ti), device=device),
                c=bz, latent_history=bh, mask=None,
            )[:, -1:, :]
            loss = loss + F.mse_loss(sv, tv) * dt
            y = (y + sv * dt).detach()
        loss = loss + 0.5 * F.mse_loss(y, bt)
        opt.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(student.parameters(), 1.0)
        opt.step()
        losses.append(float(loss.detach().cpu()))
        if step % 100 == 0 or step == TRAIN_STEPS - 1:
            log(f"train {step+1}/{TRAIN_STEPS} loss={losses[-1]:.5f}")
    student.eval()
    return student, losses


class _unused:
    pass


def export_onnx(student, path: Path):
    import copy
    import torch

    class Graph(torch.nn.Module):
        def __init__(self, dit):
            super().__init__()
            self.dit = dit

        def forward(self, x, t, c, history):
            return self.dit(x=x, t=t, c=c, latent_history=history, mask=None)[:, -1:, :]

    g = Graph(copy.deepcopy(student).float()).cpu().eval()
    x = torch.randn(1, 1, 64); t = torch.tensor([0.5])
    c = torch.randn(1, 1, 896); h = torch.randn(1, 1, 64)
    with torch.inference_mode():
        ref = g(x, t, c, h)
        torch.onnx.export(g, (x, t, c, h), str(path),
                          input_names=["x", "t", "c", "history"],
                          output_names=["velocity"], opset_version=20,
                          do_constant_folding=True, dynamo=False)
    import onnxruntime as ort
    sess = ort.InferenceSession(str(path), providers=["CPUExecutionProvider"])
    got = sess.run(None, {"x": x.numpy(), "t": t.numpy(), "c": c.numpy(), "history": h.numpy()})[0]
    err = float(np.abs(ref.numpy().reshape(-1) - got.reshape(-1)).max())
    log(f"{path.name} {path.stat().st_size/1e6:.1f}MB max_abs={err:.2e}")
    return err


def quantize(src: Path, dst: Path) -> None:
    from onnxruntime.quantization.matmul_nbits_quantizer import MatMulNBitsQuantizer

    q = MatMulNBitsQuantizer(str(src), bits=4, block_size=32, is_symmetric=True,
                             accuracy_level=4, op_types_to_quantize=("MatMul",))
    q.process()
    q.model.save_model_to_file(str(dst), use_external_data_format=False)
    log(f"{dst.name} {dst.stat().st_size/1e6:.1f}MB")


def sample_latent(dit, z, hist, noise, n_steps):
    import torch

    y = noise
    for i in range(n_steps):
        dt = 1.0 / n_steps
        v = dit(x=y, t=torch.full((y.shape[0],), i / n_steps, device=y.device),
                c=z, latent_history=hist, mask=None)[:, -1:, :]
        y = y + v * dt
    return y


def ar_generate(mm, sample_fn, text, task_prompt, prompt_latent, prompt_text, spk, frames):
    import torch

    device = next(mm.parameters()).device
    dtype = next(mm.parameters()).dtype
    with torch.no_grad():
        input_ids, inputs_embeds = mm.prepare_input_embed(
            prompt=task_prompt, text=text, spk_emb=spk,
            prompt_latent=prompt_latent, prompt_text=prompt_text,
        )
    attention_mask = torch.ones(input_ids.shape, device=device)
    position_ids = (attention_mask.cumsum(-1) - 1).masked_fill_(attention_mask == 0, 1)
    past = None
    hist = torch.zeros(1, mm.history_patch_size, mm.latent_dim, device=device, dtype=dtype)
    start = mm.history_patch_size - prompt_latent.size(1)
    if start < 0:
        hist[:] = prompt_latent[:, -start:, :]
    else:
        hist[:, start:, :] = prompt_latent
    out = []
    for _ in range(frames):
        with torch.no_grad():
            hidden = mm.model(attention_mask=attention_mask, position_ids=position_ids,
                              inputs_embeds=inputs_embeds, audio_mask=None, image_mask=None,
                              output_hidden_states=True, return_dict=True, use_cache=True,
                              past_key_values=past)
        past = hidden.past_key_values
        z = hidden.hidden_states[-1][:, -1:, :]
        noise = torch.randn(1, 1, mm.latent_dim, device=device, dtype=dtype)
        # no_grad matters: the DiT params require grad, so an unguarded sample
        # builds a graph per frame that stays alive through `hist`.
        with torch.no_grad():
            latent = sample_fn(z, hist, noise).to(device=device, dtype=dtype)
        out.append(latent.detach())
        inputs_embeds = mm.linear_proj_audio(latent)
        position_ids = position_ids[:, -1:] + 1
        attention_mask = torch.ones(inputs_embeds.shape[0], 1, device=device)
        hist[:, :-mm.patch_size, :] = hist[:, mm.patch_size:, :].clone()
        hist[:, -mm.patch_size:, :] = latent
    with torch.no_grad():
        speech = mm.audio.decode(torch.cat(out, dim=1), past_key_values=None,
                                 use_cache=False, last_chunk=True)[0]
    return speech.cpu().float().reshape(-1).numpy()


def stage_eval() -> None:
    import copy
    import torch
    from transformers import pipeline

    sys.path.insert(0, str(REPO))
    from ming_tts import MingTTS
    import tts_tasks
    sys.path.insert(0, str(BUNDLE / "eval"))
    from score import load_t2s, score_row

    table = load_t2s()
    task_prompt = tts_tasks.build_tts_prompt("Chinese")
    ckpt = OUT / f"student{STEPS}.pt"
    state = {k: v.float() for k, v in torch.load(ckpt, map_location="cpu").items()}

    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
    mm = tts.model
    device = next(mm.parameters()).device
    teacher = mm.flowloss.cfm.model.float().eval()

    asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-small",
                        device=0, torch_dtype=torch.float16)

    def asr(y):
        out = asr_pipe({"array": y.astype(np.float32), "sampling_rate": 44100},
                       generate_kwargs={"language": "zh", "task": "transcribe"})
        return (out.get("text", "") if isinstance(out, dict) else str(out)).strip()

    sents = json.loads((BUNDLE / "eval/sentences.json").read_text())["items"]
    want = EVAL_VOICES or [v["slot"] for v in voices()[:2]]
    sel = [x for x in voices(include_holdout=True) if x["slot"] in want]
    wav_dir = OUT / "eval_wavs"
    wav_dir.mkdir(parents=True, exist_ok=True)
    rows = []

    # Pass 1: teacher only. Both DiTs resident does not fit on a 15 GB T4.
    for v in sel:
        latent, spk = tts.prepare_prompt(str(BUNDLE / v["wav"]), use_spk_emb=True)
        latent = latent.to(device=device, dtype=next(mm.parameters()).dtype)
        spk = [x.to(device=device, dtype=next(mm.parameters()).dtype) for x in spk]
        for item in sents:
            text = item["text"]
            try:
                # Seed the teacher too: unseeded teacher sampling moved the
                # baseline CER by ~2 pp between runs, which is the same size as
                # the effect we are trying to measure.
                torch.manual_seed(SEED)
                tw = tts.synthesize(task="TTS", text=text, language_id="Chinese",
                                    prompt_wav_path=str(BUNDLE / v["wav"]),
                                    prompt_text=v["prompt_text"], use_spk_emb=True,
                                    raw_text=True, max_decode_steps=max(32, min(400, int(len(text) * 3.5))))
                tw = tw.cpu().numpy().reshape(-1)
                p = wav_dir / f"{v['slot']}_{item['id']}_teacher.wav"
                torch.save(torch.from_numpy(tw), p.with_suffix(".pt"))
                if SAVE_WAVS:
                    import soundfile as sf
                    sf.write(str(p), tw, 44100)
                hyp = asr(tw)
                s = score_row({"text": text, "hyp_text": hyp}, table)
                rows.append({"voice": v["slot"], "artist": v["artist"], "id": item["id"],
                             "model": "teacher", "text": text, "cer": s["cer_norm"],
                             "cer_head": s["cer_prefix"], "hyp": hyp})
            except Exception:
                log(f"teacher {v['slot']} {item['id']} failed\n{traceback.format_exc()}")
            torch.cuda.empty_cache()
        log(f"teacher {v['slot']} done ({len(rows)} rows)")

    del teacher
    torch.cuda.empty_cache()

    # Pass 2: student only. Load it here, not alongside the teacher: two DiTs
    # plus the 1.48 B backbone do not fit in 15 GB.
    import copy as _copy

    student = _copy.deepcopy(mm.flowloss.cfm.model.float())
    student.load_state_dict(state)
    student = student.to(device).eval()

    # Pass 2: student only, reusing the teacher audio for frame count.
    for v in sel:
        latent, spk = tts.prepare_prompt(str(BUNDLE / v["wav"]), use_spk_emb=True)
        latent = latent.to(device=device, dtype=next(mm.parameters()).dtype)
        spk = [x.to(device=device, dtype=next(mm.parameters()).dtype) for x in spk]
        for item in sents:
            text = item["text"]
            try:
                tp = wav_dir / f"{v['slot']}_{item['id']}_teacher.pt"
                tw = torch.load(tp, map_location="cpu").numpy()
                frames = max(16, int(round(tw.size / 3528)))
                torch.manual_seed(SEED)
                sw = ar_generate(mm, lambda z, h, n: sample_latent(student, z, h, n, STEPS),
                                 text, task_prompt, latent, v["prompt_text"], spk, frames)
                hyp = asr(sw)
                s = score_row({"text": text, "hyp_text": hyp}, table)
                if SAVE_WAVS:
                    import soundfile as sf
                    sf.write(str(wav_dir / f"{v['slot']}_{item['id']}_student.wav"), sw, 44100)
                rows.append({"voice": v["slot"], "artist": v["artist"], "id": item["id"],
                             "model": "student", "text": text, "cer": s["cer_norm"],
                             "cer_head": s["cer_prefix"], "hyp": hyp})
            except Exception:
                log(f"student {v['slot']} {item['id']} failed\n{traceback.format_exc()}")
            torch.cuda.empty_cache()
        log(f"student {v['slot']} done ({len(rows)} rows)")

    (OUT / "eval_rows.json").write_text(json.dumps(rows, ensure_ascii=False, indent=1))
    for name in ("teacher", "student"):
        sel_rows = [r for r in rows if r["model"] == name]
        if sel_rows:
            log(f"{name:<8} n={len(sel_rows)} cer={np.mean([r['cer'] for r in sel_rows]):.4f} "
                f"head={np.mean([r['cer_head'] for r in sel_rows]):.4f}")
        for v in sorted({r["voice"] for r in sel_rows}):
            vs = [r["cer"] for r in sel_rows if r["voice"] == v]
            log(f"   {v} n={len(vs)} cer={np.mean(vs):.4f}")


def stage_eval_int4() -> None:
    """Same cells as stage_eval, sampled through the int4 ONNX graph.
    Runs as its own process so the PyTorch DiT is not resident: the ORT session
    plus the backbone is already close to a 15 GB T4.
    """
    import onnxruntime as ort
    import torch
    from transformers import pipeline

    sys.path.insert(0, str(REPO))
    from ming_tts import MingTTS
    import tts_tasks
    sys.path.insert(0, str(BUNDLE / "eval"))
    from score import load_t2s, score_row

    table = load_t2s()
    model_path = OUT / f"dit{STEPS}_int4.onnx"
    if not model_path.exists():
        raise SystemExit(f"{model_path} missing")
    sess = ort.InferenceSession(str(model_path), providers=["CUDAExecutionProvider"])
    log(f"int4 session providers: {sess.get_providers()}")

    def ort_sample(z, hist, noise):
        times = [i / STEPS for i in range(STEPS + 1)]
        y = noise.detach().cpu().numpy()
        zc, hc = z.detach().cpu().numpy(), hist.detach().cpu().numpy()
        for i in range(STEPS):
            v = sess.run(None, {"x": y, "t": np.array([times[i]], np.float32),
                                "c": zc, "history": hc})[0]
            y = y + v * (times[i + 1] - times[i])
        return torch.from_numpy(y)

    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
    mm = tts.model
    device = next(mm.parameters()).device
    dtype = next(mm.parameters()).dtype
    task_prompt = tts_tasks.build_tts_prompt("Chinese")
    asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-small",
                        device=0, torch_dtype=torch.float16)

    def asr(y):
        out = asr_pipe({"array": y.astype(np.float32), "sampling_rate": 44100},
                       generate_kwargs={"language": "zh", "task": "transcribe"})
        return (out.get("text", "") if isinstance(out, dict) else str(out)).strip()

    sents = json.loads((BUNDLE / "eval/sentences.json").read_text())["items"]
    want = EVAL_VOICES or [v["slot"] for v in voices(include_holdout=True)[:2]]
    wav_dir = OUT / "eval_wavs"
    rows = []
    for v in [x for x in voices(include_holdout=True) if x["slot"] in want]:
        latent, spk = tts.prepare_prompt(str(BUNDLE / v["wav"]), use_spk_emb=True)
        latent = latent.to(device=device, dtype=dtype)
        spk = [x.to(device=device, dtype=dtype) for x in spk]
        for item in sents:
            text = item["text"]
            try:
                tp = wav_dir / f"{v['slot']}_{item['id']}_teacher.pt"
                tw = torch.load(tp, map_location="cpu").numpy()
                frames = max(16, int(round(tw.size / 3528)))
                torch.manual_seed(SEED)
                sw = ar_generate(mm, ort_sample, text, task_prompt, latent,
                                 v["prompt_text"], spk, frames)
                hyp = asr(sw)
                s = score_row({"text": text, "hyp_text": hyp}, table)
                rows.append({"voice": v["slot"], "artist": v["artist"], "id": item["id"],
                             "model": "int4", "text": text, "cer": s["cer_norm"],
                             "cer_head": s["cer_prefix"], "hyp": hyp})
            except Exception:
                log(f"int4 {v['slot']} {item['id']} failed\n{traceback.format_exc()}")
            torch.cuda.empty_cache()
        log(f"int4 {v['slot']} done ({len(rows)} rows)")

    (OUT / "eval_rows_int4.json").write_text(json.dumps(rows, ensure_ascii=False, indent=1))
    sel_rows = rows
    if sel_rows:
        log(f"int4     n={len(sel_rows)} cer={np.mean([r['cer'] for r in sel_rows]):.4f} "
            f"head={np.mean([r['cer_head'] for r in sel_rows]):.4f}")
    for v in sorted({r["voice"] for r in sel_rows}):
        vs = [r["cer"] for r in sel_rows if r["voice"] == v]
        log(f"   {v} n={len(vs)} cer={np.mean(vs):.4f}")


def main() -> None:
    import torch

    torch.manual_seed(SEED)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(SEED)
    device = torch.device("cuda")

    if STAGE == "traj":
        patch_fp32()
        stage_traj()
        return

    if STAGE == "eval_int4":
        patch_fp32()
        stage_eval_int4()
        return

    sys.path.insert(0, str(REPO))
    patch_fp32()
    from ming_tts import MingTTS
    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
    mm = tts.model

    if STAGE in ("train", "all"):
        data = load_dataset()
        teacher = mm.flowloss.cfm.model.float().to(device).eval()
        for p in teacher.parameters():
            p.requires_grad_(False)
        student, losses = train_student(teacher, data, STEPS, device)
        torch.save({k: v.half().cpu() for k, v in student.state_dict().items()},
                   OUT / f"student{STEPS}.pt")
        (OUT / "train_loss.json").write_text(json.dumps({"first": losses[0], "last": losses[-1],
                                                         "steps": TRAIN_STEPS, "lr": LR}))
        log(f"saved student{STEPS}.pt loss {losses[0]:.4f} -> {losses[-1]:.4f}")

    if STAGE in ("export", "all"):
        student = copy_state(mm, OUT / f"student{STEPS}.pt", device) if STAGE == "export" else student
        fp32 = OUT / f"dit{STEPS}.onnx"
        export_onnx(student, fp32)
        quantize(fp32, OUT / f"dit{STEPS}_int4.onnx")

    if STAGE in ("eval", "all"):
        stage_eval()


def copy_state(mm, path, device):
    import copy
    import torch

    state = {k: v.float() for k, v in torch.load(path, map_location="cpu").items()}
    dit = copy.deepcopy(mm.flowloss.cfm.model.float())
    dit.load_state_dict(state)
    return dit.to(device).eval()


if __name__ == "__main__":
    main()