scripts/train_continue.py

import argparse
import json
import math
import os
import random

import numpy as np  # noqa: E402
import torch
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline
from diffusers.models.attention import Attention
from diffusers.models.attention_processor import AttnProcessor
from huggingface_hub import HfApi, login
from PIL import Image
from safetensors.torch import save_file

RUN_DIR = os.environ.get("RUN_DIR", "/content")
DATA_DIR = os.environ.get("DATA_DIR", "").strip() or os.path.join(RUN_DIR, ".img")


def get_token() -> str:
    tok = os.environ.get("HF_TOKEN", "").strip()
    if tok:
        return tok
    for p in (
        os.path.join(RUN_DIR, "king_hf_token.txt"),
        "/content/king_hf_token.txt",
        "/content/hf_token.txt",
    ):
        if os.path.isfile(p):
            tok = open(p, encoding="utf-8-sig").read().strip()
            if tok:
                return tok
    try:
        from google.colab import userdata

        tok = str(userdata.get("HF_TOKEN") or "").strip()
        if tok:
            return tok
    except Exception:
        pass
    try:
        from kaggle_secrets import UserSecretsClient

        tok = str(UserSecretsClient().get_secret("HF_TOKEN") or "").strip()
        if tok:
            return tok
    except Exception:
        pass
    raise RuntimeError(
        f"HF_TOKEN not found (env, {RUN_DIR}/king_hf_token.txt, Colab or Kaggle secret 'HF_TOKEN')"
    )


def build_parser():
    ap = argparse.ArgumentParser()
    ap.add_argument("--dataset_repo", default="RASHID778/king2-image-dataset")
    ap.add_argument("--lora_repo", default="RASHID778/king2-image")
    ap.add_argument("--lora_weight_file", default="pytorch_lora_weights.safetensors")
    ap.add_argument("--output_repo", default="RASHID778/king2-image")
    ap.add_argument("--base_model", default="stabilityai/stable-diffusion-xl-base-1.0")
    ap.add_argument("--vae", default="madebyollin/sdxl-vae-fp16-fix")
    ap.add_argument("--resolution", type=int, default=512)
    ap.add_argument("--grad_accum", type=int, default=4)
    ap.add_argument("--max_train_steps", type=int, default=2000)
    ap.add_argument("--checkpointing_steps", type=int, default=500)
    ap.add_argument("--lr", type=float, default=1e-4)
    ap.add_argument("--warmup_steps", type=int, default=100)
    ap.add_argument("--rank", type=int, default=16)
    ap.add_argument("--alpha", type=int, default=16)
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--max_grad_norm", type=float, default=1.0)
    ap.add_argument("--loss_cap", type=float, default=100.0)
    ap.add_argument("--pred_cap", type=float, default=1000.0)
    ap.add_argument("--keep_checkpoints", action="store_true")
    return ap


def download_dataset(repo_id: str) -> str:
    import shutil
    import tarfile

    from huggingface_hub import hf_hub_download

    data_dir = DATA_DIR
    if os.path.isdir(os.path.join(data_dir, "indoor")) or os.path.isdir(
        os.path.join(data_dir, "Indoor")
    ):
        return data_dir
    os.makedirs(data_dir, exist_ok=True)
    arc = hf_hub_download(repo_id=repo_id, filename="images/images.tar.gz", repo_type="dataset")
    print(f"[dataset] archive size {os.path.getsize(arc)}", flush=True)
    with tarfile.open(arc, "r:gz") as t:
        t.extractall(data_dir)
    meta = hf_hub_download(repo_id=repo_id, filename="metadata.jsonl", repo_type="dataset")
    shutil.copy(meta, os.path.join(data_dir, "metadata.jsonl"))
    print(f"[dataset] extracted to {data_dir}", flush=True)
    return data_dir


def load_rows(data_dir: str) -> tuple[list[dict], dict]:
    meta = os.path.join(data_dir, "metadata.jsonl")
    rows = []
    actual = {}
    for sub in ("indoor", "outdoor"):
        subdir = os.path.join(data_dir, sub)
        if not os.path.isdir(subdir):
            subdir = os.path.join(data_dir, sub.capitalize())
        if os.path.isdir(subdir):
            actual[sub] = subdir
    for line in open(meta, encoding="utf-8"):
        row = json.loads(line)
        if row.get("text"):
            rows.append(row)
    print(f"[dataset] {len(rows)} captioned rows", flush=True)
    return rows, actual


def _lora_module(name: str) -> str:
    base = name
    for token in ("lora_A", "lora_B", "lora_dense", "lora"):
        idx = base.find("." + token)
        if idx != -1:
            base = base[:idx]
            break
    return base


def _is_lora_weight(name: str) -> bool:
    return "lora" in name and name.endswith("weight")


def _lora_kind(name: str) -> str:
    if any(s in name for s in ("lora_dense1", "lora_A", ".lora.down", "lora.down.")):
        return "down"
    return "up"


def collect_lora_state_dict(model) -> dict:
    sd = {}
    for name, p in model.named_parameters():
        if not _is_lora_weight(name):
            continue
        kind = _lora_kind(name)
        key = f"{_lora_module(name)}.lora.{kind}.weight"
        for prefix in ("base_model.model.", "unet."):
            if key.startswith(prefix):
                key = key[len(prefix):]
        if not key.startswith("unet."):
            key = "unet." + key
        sd[key] = p.detach().to("cpu", dtype=torch.float16)
    return sd


def build_adapter_config(targets: list[str], rank: int, alpha: int, base: str) -> dict:
    return {
        "adapter_name": "king2",
        "alpha_pattern": {},
        "auto_mapping": None,
        "base_model_name_or_path": base,
        "bias": "none",
        "fan_in_fan_out": False,
        "inference_mode": True,
        "init_lora_weights": True,
        "layer_replication": None,
        "layers_pattern": None,
        "layers_to_transform": None,
        "loops": None,
        "megatron_config": None,
        "megatron_cfg": None,
        "modules_to_save": None,
        "non_lora_submodules": None,
        "peft_type": "LORA",
        "r": rank,
        "rank_pattern": {},
        "revision": None,
        "target_modules": targets,
        "task_type": None,
        "use_dora": False,
        "use_rslora": False,
    }


def latest_checkpoint_step(api: HfApi, repo_id: str) -> int:
    best = 0
    try:
        files = api.list_repo_files(repo_id)
    except Exception:
        return best
    prefix = "checkpoint-continue-"
    for f in files:
        if f.startswith(prefix) and f.endswith("/pytorch_lora_weights.safetensors"):
            try:
                step = int(f.split(prefix, 1)[1].split("/", 1)[0])
                best = max(best, step)
            except ValueError:
                pass
    return best


def main() -> None:
    args = build_parser().parse_args()

    torch.manual_seed(args.seed)
    random.seed(args.seed)
    login(token=get_token())

    device = "cuda"
    dtype = torch.float16

    print(
        f"[conf] base={args.base_model} vae={args.vae} res={args.resolution} "
        f"steps={args.max_train_steps} lr={args.lr} rank={args.rank} accum={args.grad_accum}",
        flush=True,
    )

    data_dir = download_dataset(args.dataset_repo)
    rows, actual = load_rows(data_dir)
    if not rows:
        raise SystemExit("no captioned rows in dataset - run caption step first")

    vae = AutoencoderKL.from_pretrained(args.vae)
    pipe = StableDiffusionXLPipeline.from_pretrained(
        args.base_model,
        vae=vae,
        torch_dtype=dtype,
        variant="fp16",
        use_safetensors=True,
        safety_checker=None,
        requires_safety_checker=False,
    )
    pipe.load_lora_weights(args.lora_repo, weight_name=args.lora_weight_file)
    pipe.vae.to(dtype=torch.float32)
    scheduler = DDPMScheduler.from_pretrained(args.base_model, subfolder="scheduler")
    print("[lora] loaded RASHID778/king2-image onto UNet", flush=True)

    unet = pipe.unet.to(device)
    for _m in unet.modules():
        if isinstance(_m, Attention):
            _m.upcast_attention = True
    unet.set_attn_processor(AttnProcessor())
    pipe.vae.to(device)
    pipe.text_encoder.to(device)
    pipe.text_encoder_2.to(device)

    for n, p in unet.named_parameters():
        p.requires_grad_(False)
    for n, p in unet.named_parameters():
        if _is_lora_weight(n):
            p.requires_grad_(True)

    trainable = [p for p in unet.parameters() if p.requires_grad]
    n_params = sum(p.numel() for p in trainable)
    print(f"[params] trainable params: {n_params/1e6:.2f}M", flush=True)

    lora_targets = sorted(
    {
        _lora_module(name)
        .removeprefix("unet.")
        .removeprefix("base_model.model.")
        for name, _ in unet.named_parameters()
        if _is_lora_weight(name)
    }
)
    saved_cfg = build_adapter_config(lora_targets, args.rank, args.alpha, args.base_model)

    optimizer = torch.optim.AdamW(trainable, lr=args.lr, weight_decay=1e-2, eps=1e-6)
    total_updates = args.max_train_steps

    def warmup_cosine(step: int) -> float:
        if step < args.warmup_steps:
            return step / max(1, args.warmup_steps)
        prog = (step - args.warmup_steps) / max(1, total_updates - args.warmup_steps)
        return 0.5 * (1.0 + math.cos(math.pi * min(1.0, prog)))

    lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, warmup_cosine)

    res = args.resolution

    def preprocess(pil: Image.Image) -> torch.Tensor:
        w, h = pil.size
        s = min(w, h)
        pil = pil.crop(((w - s) // 2, (h - s) // 2, (w + s) // 2, (h + s) // 2))
        pil = pil.resize((res, res), Image.BILINEAR)
        return (torch.from_numpy(np.array(pil)).float() / 127.5 - 1.0).permute(2, 0, 1)

    def encode_prompt(texts: list[str]):
        with torch.no_grad():
            out = pipe.encode_prompt(
                texts, device, num_images_per_prompt=1, do_classifier_free_guidance=False
            )
        if isinstance(out, dict):
            return out["prompt_embeds"], out["pooled_prompt_embeds"]
        if len(out) == 4:
            return out[0], out[2]
        return out[0], out[1]

    def encode_latents(pixels: torch.Tensor) -> torch.Tensor:
        with torch.no_grad():
            latents = pipe.vae.encode(pixels).latent_dist.sample()
        latents = latents * pipe.vae.config.scaling_factor
        if not torch.isfinite(latents).all():
            raise RuntimeError(
                f"[vae] non-finite latents at row {pos % len(rows)} (fn {rows[pos % len(rows)]['file_name']})"
            )
        return latents.to(dtype)

    add_time_ids = torch.tensor([res, res, 0, 0, res, res]).unsqueeze(0).to(device)

    def get_batch(idx: int):
        cap = rows[idx]["text"]
        fn = rows[idx]["file_name"]
        sub, name = fn.split("/", 1)
        img = os.path.join(actual.get(sub, os.path.join(data_dir, sub)), name)
        pil = Image.open(img).convert("RGB")
        pixels = preprocess(pil).unsqueeze(0).to(device)
        prompt_embeds, pooled = encode_prompt([cap])
        latents = encode_latents(pixels)
        return latents, prompt_embeds, pooled

    pos = 0
    running_loss = 0.0
    api = HfApi()
    global_step = latest_checkpoint_step(api, args.output_repo)
    if global_step >= total_updates:
        raise SystemExit(f"training already complete (latest checkpoint at step {global_step})")
    if global_step > 0:
        weight = f"checkpoint-continue-{global_step}/pytorch_lora_weights.safetensors"
        say = f"[resume] continuing from checkpoint-continue-{global_step} (step {global_step}/{total_updates})"
        print(say, flush=True)
        pipe.load_lora_weights(args.output_repo, weight_name=weight)
        for n, p in unet.named_parameters():
            p.requires_grad_(False)
        for n, p in unet.named_parameters():
            if _is_lora_weight(n):
                p.requires_grad_(True)
        trainable.clear()
        trainable.extend([p for p in unet.parameters() if p.requires_grad])

    skipped = 0

    while global_step < total_updates:
        optimizer.zero_grad(set_to_none=True)
        step_loss = 0.0
        for _micro in range(args.grad_accum):
            latents, prompt_embeds, pooled = get_batch(pos % len(rows))
            pos += 1
            noise = torch.randn_like(latents)
            timesteps = torch.randint(
                0, scheduler.config.num_train_timesteps, (1,), device=device
            ).long()
            noisy = scheduler.add_noise(latents, noise, timesteps)
            with torch.autocast(device_type="cuda", dtype=torch.float16):
                pred = unet(
                    noisy,
                    timesteps,
                    encoder_hidden_states=prompt_embeds,
                    added_cond_kwargs={
                        "text_embeds": pooled,
                        "time_ids": add_time_ids,
                    },
                ).sample
            pred = pred.float()
            loss = torch.nn.functional.mse_loss(pred, noise.float()) / args.grad_accum
            loss_val = float(loss)
            pred_max = float(pred.abs().max())
            if (
                not torch.isfinite(loss)
                or not torch.isfinite(pred).all().item()
                or loss_val > args.loss_cap
                or pred_max > args.pred_cap
            ):
                skipped += 1
                print(
                    f"[warn] unstable loss at step {global_step} micro {_micro} "
                    f"row {pos - 1} loss={loss_val:.3f} pred_max={pred_max:.3f} - skipped",
                    flush=True,
                )
                continue
            loss.backward()
            step_loss += loss.item() * args.grad_accum
            if global_step == 0 and _micro == 0:
                print(
                    f"[diag] step0 row={pos - 1} fn={rows[pos - 1]['file_name']} "
                    f"latents_finite={torch.isfinite(latents).all().item()} "
                    f"pred_finite={pred.isfinite().all().item()} loss={loss.item() * args.grad_accum:.6f}",
                    flush=True,
                )
        grad_finite = all(
            p.grad is None or torch.isfinite(p.grad).all() for p in trainable
        )
        if grad_finite:
            torch.nn.utils.clip_grad_norm_(trainable, args.max_grad_norm)
            optimizer.step()
            lr_scheduler.step()
        else:
            skipped += 1
            optimizer.zero_grad(set_to_none=True)
            print(f"[warn] non-finite grads at step {global_step} - step skipped", flush=True)
        global_step += 1
        running_loss += step_loss

        if global_step % 25 == 0 or global_step == total_updates:
            print(
                f"[step] {global_step}/{total_updates} "
                f"loss={running_loss / min(25, global_step):.4f} "
                f"lr={lr_scheduler.get_last_lr()[0]:.2e} "
                f"skipped={skipped}",
                flush=True,
            )
            running_loss = 0.0

        if args.checkpointing_steps and global_step % args.checkpointing_steps == 0:
            bad = sum(1 for p in trainable if not torch.isfinite(p).all().item())
            if bad:
                raise SystemExit(
                    f"ABORT: {bad} non-finite trainable weights at step {global_step} - refusing to checkpoint/upload"
                )
            save_dir = os.path.join(RUN_DIR, f"ckpt-{global_step}")
            os.makedirs(save_dir, exist_ok=True)
            sd = collect_lora_state_dict(unet)
            save_file(sd, os.path.join(save_dir, "pytorch_lora_weights.safetensors"))
            with open(os.path.join(save_dir, "adapter_config.json"), "w", encoding="utf-8") as fh:
                json.dump(saved_cfg, fh, indent=2)
            print(f"[ckpt] saved locally {save_dir} ({len(sd)} tensors)", flush=True)
            if args.keep_checkpoints:
                api.upload_folder(
                    repo_id=args.output_repo,
                    folder_path=save_dir,
                    path_in_repo=f"checkpoint-continue-{global_step}",
                    commit_message=f"continuation checkpoint {global_step}",
                )
                print(f"[ckpt] pushed checkpoint-continue-{global_step}", flush=True)
            sd.clear()
            del sd
            torch.cuda.empty_cache()

    final_dir = os.path.join(RUN_DIR, "king2-final")
    os.makedirs(final_dir, exist_ok=True)
    bad_final = sum(1 for p in trainable if not torch.isfinite(p).all().item())
    if bad_final:
        raise SystemExit(f"ABORT: {bad_final} non-finite trainable weights at end of training - refusing to upload")
    final_sd = collect_lora_state_dict(unet)
    save_file(final_sd, os.path.join(final_dir, "pytorch_lora_weights.safetensors"))
    with open(os.path.join(final_dir, "adapter_config.json"), "w", encoding="utf-8") as fh:
        json.dump(saved_cfg, fh, indent=2)
    print(f"[done] trained {global_step} steps; final LoRA saved locally", flush=True)

    print("[upload] pushing final weights to", args.output_repo, flush=True)
    api.upload_folder(
        repo_id=args.output_repo,
        folder_path=final_dir,
        path_in_repo=".",
        commit_message="king2-image continuous fine-tune (indoor/outdoor dataset)",
    )
    print("[uploaded] main LoRA updated: pytorch_lora_weights.safetensors", flush=True)


if __name__ == "__main__":
    main()