scripts/job_quantize.py

#!/usr/bin/env python3
"""
Fish Speech S2 Pro — Quantization Experiment (HF Job version)
Downloads model, applies quantization at all phases, generates voice clone samples.
"""
import os, sys, json, time, gc, traceback
import torch
import torch.nn as nn
import numpy as np

os.environ["TOKENIZERS_PARALLELISM"] = "false"
DEVICE = "cuda"
DTYPE = torch.bfloat16
BASE_MODEL = "fishaudio/s2-pro"
OUT = "/app/output"

print("=== Fish Speech S2 Pro Quantization Experiment ===")
print(f"PyTorch: {torch.__version__}, CUDA: {torch.cuda.is_available()}")
print(f"GPU: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'none'}")
print(f"VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB" if torch.cuda.is_available() else "")

# Install deps
print("\n[1/8] Installing dependencies...")
os.system("pip install -q einops loguru ormsgpack hydra-core omegaconf safetensors torchaudio soundfile")
os.system("pip install -q datasets")

# Clone fish-speech
if not os.path.exists("/app/fish-speech"):
    print("\n[2/8] Cloning fish-speech repo...")
    os.system("cd /app && git clone --depth 1 https://github.com/fishaudio/fish-speech.git")
else:
    print("\n[2/8] fish-speech already cloned")
sys.path.insert(0, "/app/fish-speech")

# Download model
print("\n[3/8] Downloading S2 Pro model...")
os.system(f"huggingface-cli download {BASE_MODEL} --local-dir /app/checkpoints/s2-pro")

BASE_MODEL = "/app/checkpoints/s2-pro"

# ============ QUANTIZATION CLASSES ============

class FP8Linear(nn.Module):
    def __init__(self, in_f, out_f, bias=True):
        super().__init__()
        self.in_features, self.out_features = in_f, out_f
        self.register_buffer("weight", torch.empty(out_f, in_f, dtype=torch.float8_e4m3fn))
        self.register_buffer("weight_scale", torch.empty(out_f, 1, dtype=torch.float32))
        self.has_bias = bias
        if bias: self.register_buffer("bias", torch.zeros(out_f, dtype=torch.bfloat16))
        else: self.bias = None

    @staticmethod
    def from_linear(linear):
        fp8 = FP8Linear(linear.in_features, linear.out_features, linear.bias is not None)
        w = linear.weight.data.detach().bfloat16()
        scale = w.abs().amax(dim=1, keepdim=True).clamp(min=1e-12) / 448.0
        w_q = (w / scale).round().clamp(-448, 448).to(torch.float8_e4m3fn)
        fp8.weight.data.copy_(w_q)
        fp8.weight_scale.data.copy_(scale)
        if linear.bias is not None:
            fp8.bias.data.copy_(linear.bias.data.detach().bfloat16())
        return fp8

    def forward(self, x):
        return nn.functional.linear(x, self.weight.to(torch.bfloat16) * self.weight_scale, self.bias)

class INT8Linear(nn.Module):
    def __init__(self, in_f, out_f, bias=True):
        super().__init__()
        self.in_features, self.out_features = in_f, out_f
        self.register_buffer("weight", torch.empty(out_f, in_f, dtype=torch.int8))
        self.register_buffer("weight_scale", torch.empty(out_f, 1, dtype=torch.float32))
        self.has_bias = bias
        if bias: self.register_buffer("bias", torch.zeros(out_f, dtype=torch.bfloat16))
        else: self.bias = None

    @staticmethod
    def from_linear(linear):
        q = INT8Linear(linear.in_features, linear.out_features, linear.bias is not None)
        w = linear.weight.data.detach().bfloat16()
        scale = w.abs().amax(dim=1, keepdim=True).clamp(min=1e-12) / 127.0
        q.weight.data.copy_((w / scale).round().clamp(-128, 127).to(torch.int8))
        q.weight_scale.data.copy_(scale)
        if linear.bias is not None: q.bias.data.copy_(linear.bias.data.detach().bfloat16())
        return q

    def forward(self, x):
        return nn.functional.linear(x, self.weight.to(torch.bfloat16) * self.weight_scale, self.bias)

class INT4Linear(nn.Module):
    def __init__(self, in_f, out_f, group_size=128, bias=True):
        super().__init__()
        self.in_features, self.out_features, self.group_size = in_f, out_f, group_size
        self.register_buffer("weight_q", torch.empty(out_f, in_f, dtype=torch.int8))
        self.register_buffer("weight_scale", torch.empty(out_f, (in_f + group_size - 1) // group_size, dtype=torch.float32))
        self.has_bias = bias
        if bias: self.register_buffer("bias", torch.zeros(out_f, dtype=torch.bfloat16))
        else: self.bias = None

    @staticmethod
    def from_linear(linear, group_size=128):
        in_f, out_f = linear.in_features, linear.out_features
        q = INT4Linear(in_f, out_f, group_size, linear.bias is not None)
        w = linear.weight.data.detach().bfloat16()
        n_groups = (in_f + group_size - 1) // group_size
        pad = n_groups * group_size - in_f
        if pad > 0: w = nn.functional.pad(w, (0, pad))
        w_g = w.reshape(out_f, n_groups, group_size)
        scale = w_g.abs().amax(dim=-1, keepdim=True).clamp(min=1e-10) / 7.0
        q.weight_q.data.copy_((w_g / scale).round().clamp(-7, 7).to(torch.int8).reshape(out_f, -1)[:, :in_f])
        q.weight_scale.data.copy_(scale.squeeze(-1)[:, :n_groups])
        if linear.bias is not None: q.bias.data.copy_(linear.bias.data.detach().bfloat16())
        return q

    def forward(self, x):
        s = self.weight_scale.repeat_interleave(self.group_size, dim=1)[:, :self.in_features]
        return nn.functional.linear(x, self.weight_q[:, :self.in_features].to(torch.bfloat16) * s, self.bias)

class INT3Linear(nn.Module):
    def __init__(self, in_f, out_f, group_size=128, bias=True):
        super().__init__()
        self.in_features, self.out_features, self.group_size = in_f, out_f, group_size
        self.register_buffer("weight_q", torch.empty(out_f, in_f, dtype=torch.int8))
        self.register_buffer("weight_scale", torch.empty(out_f, (in_f + group_size - 1) // group_size, dtype=torch.float32))
        self.has_bias = bias
        if bias: self.register_buffer("bias", torch.zeros(out_f, dtype=torch.bfloat16))
        else: self.bias = None

    @staticmethod
    def from_linear(linear, group_size=128):
        in_f, out_f = linear.in_features, linear.out_features
        q = INT3Linear(in_f, out_f, group_size, linear.bias is not None)
        w = linear.weight.data.detach().bfloat16()
        n_groups = (in_f + group_size - 1) // group_size
        pad = n_groups * group_size - in_f
        if pad > 0: w = nn.functional.pad(w, (0, pad))
        w_g = w.reshape(out_f, n_groups, group_size)
        scale = w_g.abs().amax(dim=-1, keepdim=True).clamp(min=1e-10) / 3.0
        q.weight_q.data.copy_((w_g / scale).round().clamp(-3, 3).to(torch.int8).reshape(out_f, -1)[:, :in_f])
        q.weight_scale.data.copy_(scale.squeeze(-1)[:, :n_groups])
        if linear.bias is not None: q.bias.data.copy_(linear.bias.data.detach().bfloat16())
        return q

    def forward(self, x):
        s = self.weight_scale.repeat_interleave(self.group_size, dim=1)[:, :self.in_features]
        return nn.functional.linear(x, self.weight_q[:, :self.in_features].to(torch.bfloat16) * s, self.bias)

class INT2Linear(nn.Module):
    def __init__(self, in_f, out_f, group_size=64, bias=True):
        super().__init__()
        self.in_features, self.out_features, self.group_size = in_f, out_f, group_size
        self.register_buffer("weight_q", torch.empty(out_f, in_f, dtype=torch.int8))
        self.register_buffer("weight_scale", torch.empty(out_f, (in_f + group_size - 1) // group_size, dtype=torch.float32))
        self.has_bias = bias
        if bias: self.register_buffer("bias", torch.zeros(out_f, dtype=torch.bfloat16))
        else: self.bias = None

    @staticmethod
    def from_linear(linear, group_size=64):
        in_f, out_f = linear.in_features, linear.out_features
        q = INT2Linear(in_f, out_f, group_size, linear.bias is not None)
        w = linear.weight.data.detach().bfloat16()
        n_groups = (in_f + group_size - 1) // group_size
        pad = n_groups * group_size - in_f
        if pad > 0: w = nn.functional.pad(w, (0, pad))
        w_g = w.reshape(out_f, n_groups, group_size)
        scale = w_g.abs().amax(dim=-1, keepdim=True).clamp(min=1e-10) / 1.0
        q.weight_q.data.copy_((w_g / scale).round().clamp(-1, 1).to(torch.int8).reshape(out_f, -1)[:, :in_f])
        q.weight_scale.data.copy_(scale.squeeze(-1)[:, :n_groups])
        if linear.bias is not None: q.bias.data.copy_(linear.bias.data.detach().bfloat16())
        return q

    def forward(self, x):
        s = self.weight_scale.repeat_interleave(self.group_size, dim=1)[:, :self.in_features]
        return nn.functional.linear(x, self.weight_q[:, :self.in_features].to(torch.bfloat16) * s, self.bias)


# ============ HELPERS ============

def apply_quant(model, qcls, target="slow_ar", **kw):
    count = 0
    skip = ['embed', 'norm']
    for name, mod in list(model.named_modules()):
        if not isinstance(mod, nn.Linear): continue
        if any(s in name for s in skip): continue
        if target == "slow_ar" and "fast_" in name: continue
        parts = name.split(".")
        parent = model
        for p in parts[:-1]: parent = getattr(parent, p)
        try:
            setattr(parent, parts[-1], qcls.from_linear(mod, **kw))
            count += 1
        except: pass
    return model, count

def model_size_mb(m):
    t = sum(p.numel() * p.element_size() for p in m.parameters())
    t += sum(b.numel() * b.element_size() for b in m.buffers())
    return t / (1024*1024)

def generate_sample(model, codec, text, out_path):
    """Generate TTS sample"""
    import soundfile as sf
    from fish_speech.models.text2semantic.inference import generate, decode_one_token_ar
    from fish_speech.content_sequence import TextPart
    from fish_speech.conversation import Conversation, Message

    try:
        conv = Conversation()
        conv.add_message(Message(role="user", parts=[TextPart(text="")]))
        conv.add_message(Message(role="assistant", parts=[TextPart(text=text)]))
        prompt = conv.encode_for_inference(model.config)
        cd = 1 + model.config.num_codebooks
        am = torch.zeros(1, cd, prompt.shape[-1], dtype=torch.bool, device=DEVICE)
        ap = torch.zeros(1, cd, prompt.shape[-1], dtype=torch.long, device=DEVICE)

        if not getattr(model, '_cache_done', False):
            model.setup_caches(1, model.config.max_seq_len, dtype=DTYPE)
            model._cache_done = True

        with torch.autocast(device_type="cuda", dtype=DTYPE):
            result = generate(model=model, prompt=prompt, max_new_tokens=512,
                audio_masks=am, audio_parts=ap, temperature=0.7, top_p=0.7, top_k=30,
                decode_one_token=decode_one_token_ar)

        codes = result[0:1, :, :].unsqueeze(0)
        with torch.autocast(device_type="cuda", dtype=DTYPE):
            audio = codec.decode(codes.to(DEVICE))
        audio_np = audio.squeeze().cpu().float().numpy()
        sr = getattr(codec, 'sample_rate', 44100)
        sf.write(out_path, audio_np, sr)
        dur = len(audio_np) / sr
        print(f"  Saved {out_path} ({dur:.1f}s)")
        return True, dur
    except Exception as e:
        print(f"  Sample gen failed: {e}")
        traceback.print_exc()
        return False, 0

def generate_clone(model, codec, text, ref_path, ref_text, out_path):
    """Voice clone sample"""
    import torchaudio, soundfile as sf
    from fish_speech.models.text2semantic.inference import generate, decode_one_token_ar
    from fish_speech.content_sequence import TextPart, VQPart
    from fish_speech.conversation import Conversation, Message

    try:
        wav, sr = torchaudio.load(ref_path)
        if wav.shape[0] > 1: wav = wav.mean(dim=0, keepdim=True)
        if sr != 44100: wav = torchaudio.functional.resample(wav, sr, 44100)
        wav = wav.to(DEVICE)
        with torch.autocast(device_type="cuda", dtype=DTYPE):
            enc = codec.encode(wav.unsqueeze(0))
            ptokens = (enc[0] if isinstance(enc, tuple) else enc).cpu().numpy()

        conv = Conversation()
        conv.add_message(Message(role="user", parts=[VQPart(codes=ptokens), TextPart(text=ref_text)]))
        conv.add_message(Message(role="assistant", parts=[TextPart(text=text)]))
        prompt = conv.encode_for_inference(model.config)
        cd = 1 + model.config.num_codebooks
        am = torch.zeros(1, cd, prompt.shape[-1], dtype=torch.bool, device=DEVICE)
        ap = torch.zeros(1, cd, prompt.shape[-1], dtype=torch.long, device=DEVICE)

        if not getattr(model, '_cache_done', False):
            model.setup_caches(1, model.config.max_seq_len, dtype=DTYPE)
            model._cache_done = True

        with torch.autocast(device_type="cuda", dtype=DTYPE):
            result = generate(model=model, prompt=prompt, max_new_tokens=512,
                audio_masks=am, audio_parts=ap, temperature=0.7, top_p=0.7, top_k=30,
                decode_one_token=decode_one_token_ar)
        codes = result[0:1, :, :].unsqueeze(0)
        with torch.autocast(device_type="cuda", dtype=DTYPE):
            audio = codec.decode(codes.to(DEVICE))
        audio_np = audio.squeeze().cpu().float().numpy()
        sr = getattr(codec, 'sample_rate', 44100)
        sf.write(out_path, audio_np, sr)
        dur = len(audio_np) / sr
        print(f"  Clone saved {out_path} ({dur:.1f}s)")
        return True, dur
    except Exception as e:
        print(f"  Clone failed: {e}")
        traceback.print_exc()
        return False, 0


# ============ PHASE RUNNER ============

def run_phase(pid, qcls, target, codec, ref_path, ref_text, test_text, clone_text, **kw):
    from fish_speech.models.text2semantic.inference import init_model
    from safetensors.torch import save_file

    phase_dir = f"{OUT}/{pid}"
    os.makedirs(phase_dir, exist_ok=True)
    os.makedirs(f"{OUT}/samples", exist_ok=True)

    print(f"\n{'='*60}")
    print(f"  {pid.upper()}: {qcls.__name__} ({target})")
    print(f"{'='*60}")

    model, _ = init_model(BASE_MODEL, DEVICE, DTYPE, compile=False)
    orig = model_size_mb(model)

    t0 = time.time()
    model, nl = apply_quant(model, qcls, target=target, **kw)
    model = model.to(DEVICE)
    t_q = time.time() - t0
    qs = model_size_mb(model)
    ratio = orig / qs if qs > 0 else 0
    print(f"  {orig:.0f} -> {qs:.0f} MB ({ratio:.2f}x, {nl} layers, {t_q:.1f}s)")

    sp = f"{phase_dir}/model.safetensors"
    save_file(model.state_dict(), sp)
    disk = os.path.getsize(sp) / (1024*1024)
    print(f"  Disk: {disk:.0f} MB")

    tts_ok, tts_d = generate_sample(model, codec, test_text, f"{OUT}/samples/{pid}_tts.wav")
    clone_ok, clone_d = False, 0
    if ref_path and os.path.exists(ref_path):
        clone_ok, clone_d = generate_clone(model, codec, clone_text, ref_path, ref_text, f"{OUT}/samples/{pid}_clone.wav")

    del model; gc.collect(); torch.cuda.empty_cache()

    r = {"phase": pid, "method": qcls.__name__, "target": target,
         "orig_mb": round(orig), "quant_mb": round(qs), "disk_mb": round(disk),
         "ratio": round(ratio, 3), "layers": nl, "time_s": round(t_q,1),
         "tts_ok": tts_ok, "tts_d": round(tts_d,1),
         "clone_ok": clone_ok, "clone_d": round(clone_d,1)}
    with open(f"{phase_dir}/results.json","w") as f: json.dump(r,f,indent=2)
    return r

# ============ MAIN ============

TEST_TEXT = "The quick brown fox jumps over the lazy dog. Artificial intelligence is transforming how we communicate."
CLONE_TEXT = "Hello everyone, welcome to this special presentation. Today we explore the fascinating world of neural text to speech and voice cloning technology."
REF_TEXT = "This is a reference voice recording."
CELEB_TEXT = "Good morning. I want to tell you something about the universe. Every atom in your body came from a star that exploded. We are all made of star stuff."

PHASES = {
    "1a": (FP8Linear, "slow_ar", {}),
    "1b": (INT4Linear, "slow_ar", {"group_size": 128}),
    "2a": (INT4Linear, "all", {"group_size": 128}),
    "2b": (INT8Linear, "slow_ar", {}),
    "2c": (INT3Linear, "slow_ar", {"group_size": 128}),
    "3a": (INT2Linear, "slow_ar", {"group_size": 64}),
    "3b": (INT2Linear, "all", {"group_size": 64}),
}

def main():
    os.makedirs(f"{OUT}/samples", exist_ok=True)
    from fish_speech.models.text2semantic.inference import init_model
    from fish_speech.models.dac.inference import load_codec_model

    print("\n[4/8] Loading base model...")
    model_base, _ = init_model(BASE_MODEL, DEVICE, DTYPE, compile=False)
    orig = model_size_mb(model_base)
    print(f"  Base model: {orig:.0f} MB ({sum(p.numel() for p in model_base.parameters())/1e9:.2f}B params)")

    print("\n[5/8] Loading codec...")
    codec = load_codec_model(f"{BASE_MODEL}/codec.pth", DEVICE, DTYPE)

    # Generate reference + baseline
    ref_path = f"{OUT}/reference_celebrity.wav"
    print("\n[6/8] Generating reference audio & baseline...")
    try:
        generate_sample(model_base, codec, CELEB_TEXT, ref_path)
    except Exception as e:
        print(f"  Ref gen warning: {e}")
        ref_path = None

    try:
        generate_sample(model_base, codec, TEST_TEXT, f"{OUT}/samples/baseline_bf16_tts.wav")
    except: pass

    if ref_path:
        try:
            generate_clone(model_base, codec, CLONE_TEXT, ref_path, REF_TEXT, f"{OUT}/samples/baseline_bf16_clone.wav")
        except: pass

    del model_base; gc.collect(); torch.cuda.empty_cache()

    # Run all phases
    print("\n[7/8] Running quantization phases...")
    all_r = [{"phase": "baseline_bf16", "orig_mb": round(orig), "quant_mb": round(orig), "disk_mb": round(orig), "ratio": 1.0}]

    for pid, (qcls, target, kw) in PHASES.items():
        try:
            r = run_phase(f"phase{pid}", qcls, target, codec, ref_path, REF_TEXT, TEST_TEXT, CLONE_TEXT, **kw)
            all_r.append(r)
        except Exception as e:
            print(f"Phase {pid} FAILED: {e}")
            traceback.print_exc()
            all_r.append({"phase": f"phase{pid}", "error": str(e)})

    # Summary
    print(f"\n{'='*70}")
    print("  RESULTS SUMMARY")
    print(f"{'='*70}")
    print(f"{'Phase':<14} {'Method':<14} {'Disk MB':<10} {'Ratio':<8} {'TTS':<5} {'Clone':<5}")
    print("-"*60)
    for r in all_r:
        print(f"{r.get('phase','?'):<14} {r.get('method','bf16'):<14} {r.get('disk_mb','?'):<10} {r.get('ratio',1):<8.2f} {'OK' if r.get('tts_ok') else 'FAIL':<5} {'OK' if r.get('clone_ok') else 'FAIL':<5}")

    with open(f"{OUT}/all_results.json","w") as f: json.dump(all_r,f,indent=2)

    # Upload to Hub
    print("\n[8/8] Uploading results to HuggingFace Hub...")
    try:
        from huggingface_hub import HfApi
        api = HfApi()
        repo = "Swagcrew/fish-speech-s2-quantized"
        api.create_repo(repo_id=repo, repo_type="model", exist_ok=True, private=False)

        # Upload all results
        api.upload_file(path_or_fileobj=f"{OUT}/all_results.json", path_in_repo="all_results.json", repo_id=repo, repo_type="model")

        # Upload samples
        samples_dir = f"{OUT}/samples"
        if os.path.exists(samples_dir):
            for fn in os.listdir(samples_dir):
                if fn.endswith(".wav"):
                    api.upload_file(path_or_fileobj=os.path.join(samples_dir, fn),
                                    path_in_repo=f"samples/{fn}", repo_id=repo, repo_type="model")

        # Upload individual phase results
        for pid in PHASES:
            phase_dir = f"{OUT}/phase{pid}"
            if os.path.exists(f"{phase_dir}/results.json"):
                api.upload_file(f"{phase_dir}/results.json", f"phase{pid}/results.json", repo, repo_type="model")
            if os.path.exists(f"{phase_dir}/model.safetensors"):
                api.upload_file(f"{phase_dir}/model.safetensors", f"phase{pid}/model.safetensors", repo, repo_type="model")

        # Upload README
        readme = """# Fish Speech S2 Pro — Quantization Experiments
Multi-phase quantization with voice cloning samples.
See all_results.json for details and samples/ for audio.
"""
        api.upload_file(path_or_fileobj=readme.encode(), path_in_repo="README.md", repo_id=repo, repo_type="model")

        print(f"  Uploaded to https://huggingface.co/{repo}")
    except Exception as e:
        print(f"  Upload failed: {e}")
        traceback.print_exc()

    print("\nDONE! All phases complete.")

if __name__ == "__main__":
    main()