scripts/train_compressor.py

"""
Train OpticalCompressor on Qwen3-VL-8B for UI-to-Code.

Architecture:
  Qwen3-VL ViT+Merger (frozen) → OpticalCompressor (trainable) → Qwen3 LLM (LoRA)

Usage:
  # Single GPU smoke test
  CUDA_VISIBLE_DEVICES=0 python scripts/train_compressor.py \
      --max_samples 100 --epochs 1 --batch_size 1

  # Full 6-GPU DDP training
  torchrun --nproc_per_node=6 scripts/train_compressor.py \
      --max_samples 50000 --epochs 5 --batch_size 1 --grad_accum 8

  # Resume from checkpoint
  torchrun --nproc_per_node=6 scripts/train_compressor.py \
      --resume checkpoints/optical/latest.pt --epochs 5

  # Mix WebSight + eval-aligned subset (ref_screenshots_websight + websight_gt_html, 100 pairs)
  CUDA_VISIBLE_DEVICES=1 python scripts/train_compressor.py \
      --mix_root data --epochs 10 --max_samples 5000

  # Mix WebSight + Design2Code eval split (484 pairs; run scripts/export_design2code_gt_html.py first)
  CUDA_VISIBLE_DEVICES=1 python scripts/train_compressor.py \
      --mix_root data --mix_images_subdir ref_screenshots --mix_gt_subdir gt_html

  # Per-epoch CLIP eval + continue training later (optimizer in checkpoint; --epochs is final count)
  CUDA_VISIBLE_DEVICES=1 python scripts/train_compressor.py ... --epochs 5 --eval_after_epoch
  CUDA_VISIBLE_DEVICES=1 python scripts/train_compressor.py ... --resume checkpoints/optical/latest.pt --epochs 10
"""

import os
os.environ["HF_ENDPOINT"] = os.environ.get("HF_ENDPOINT", "https://hf-mirror.com")
os.environ["HF_HOME"] = os.environ.get("HF_HOME", "/root/rivermind-data/huggingface")
os.environ["TOKENIZERS_PARALLELISM"] = "false"
# Suppress NCCL debug noise
os.environ.setdefault("NCCL_DEBUG", "WARN")

import argparse
import gc
import math
import subprocess
import sys
import time
from pathlib import Path

import torch
import torch.nn as nn
import torch.distributed as dist
from torch.utils.data import ConcatDataset, Dataset, DataLoader, DistributedSampler

PROJECT_ROOT = Path(__file__).parent.parent
sys.path.insert(0, str(PROJECT_ROOT))

from models.optical_compressor import OpticalCompressor

# ---------- constants ----------
UI2CODE_PROMPT = (
    "Convert this webpage screenshot to HTML code. "
    "Generate a complete, self-contained HTML file with inline CSS. "
    "Output only the code."
)
IMAGE_TOKEN_ID = 151655  # <|image_pad|> in Qwen3-VL


def log(msg, rank=0):
    """Print only on main process."""
    if int(os.environ.get("LOCAL_RANK", 0)) == rank:
        print(msg, flush=True)


def log_all(msg):
    """Print on all processes (for debugging hangs)."""
    rank = int(os.environ.get("LOCAL_RANK", 0))
    print(f"[rank{rank}] {msg}", flush=True)


def _uipress_stack_to_cpu(model):
    model.base_model.cpu()
    model.compressor.cpu()
    model.lora_modules.cpu()


def _uipress_stack_to_device(model, device):
    model.base_model.to(device)
    model.compressor.to(device)
    model.lora_modules.to(device)


def _run_subprocess_eval_and_clip(args, out_dir: Path, epoch: int) -> Path | None:
    """Free GPU, run eval_all + step_clip_batch; returns method_dir with clip_scores.json or None."""
    eval_root = (
        Path(args.eval_output_dir).resolve()
        if args.eval_output_dir
        else (PROJECT_ROOT / "results" / "clip_per_epoch" / out_dir.name)
    )
    eval_epoch_dir = eval_root / f"epoch_{epoch}"
    eval_epoch_dir.mkdir(parents=True, exist_ok=True)
    run_name = f"uipress_{args.target_tokens}"
    method_dir = eval_epoch_dir / run_name

    cmd_base = [sys.executable, str(PROJECT_ROOT / "scripts" / "eval_all.py")]
    r1 = subprocess.run(
        cmd_base
        + [
            "--method",
            "uipress",
            "--checkpoint",
            str(out_dir / "latest.pt"),
            "--max_samples",
            str(args.eval_max_samples),
            "--data_dir",
            args.eval_data_dir,
            "--output_dir",
            str(eval_epoch_dir),
            "--target_tokens",
            str(args.target_tokens),
        ],
        cwd=str(PROJECT_ROOT),
    )
    if r1.returncode != 0:
        print(f"  [eval] eval_all.py failed rc={r1.returncode}", flush=True)
        return None

    cmd_clip = [
        sys.executable,
        str(PROJECT_ROOT / "scripts" / "step_clip_batch.py"),
        "--method_dir",
        str(method_dir),
        "--ref_dir",
        args.eval_ref_dir,
        "--clip_device",
        args.eval_clip_device,
    ]
    r2 = subprocess.run(cmd_clip, cwd=str(PROJECT_ROOT))
    if r2.returncode != 0:
        print(f"  [eval] step_clip_batch.py failed rc={r2.returncode}", flush=True)
        return None

    clip_path = method_dir / "clip_scores.json"
    if clip_path.exists():
        import json

        summary = json.loads(clip_path.read_text(encoding="utf-8"))
        print(
            f"  [eval] epoch {epoch} CLIP avg={summary.get('avg_clip')} "
            f"n={summary.get('n')} -> {clip_path}",
            flush=True,
        )
    return method_dir


def _torch_load_compat(path, map_location):
    try:
        return torch.load(path, map_location=map_location, weights_only=False)
    except TypeError:
        return torch.load(path, map_location=map_location)


def _rebuild_optimizer_scheduler(args, model, device, total_steps, ckpt_path: Path):
    lora_trainable = [p for p in model.lora_modules.parameters() if p.requires_grad]
    trainable_params = (
        list(model.compressor.parameters())
        + lora_trainable
    )
    optim_groups = [{"params": list(model.compressor.parameters()), "lr": args.lr_compressor}]
    if lora_trainable:
        optim_groups.append({"params": lora_trainable, "lr": args.lr_lora})
    optimizer = torch.optim.AdamW(optim_groups, weight_decay=0.01)
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
        optimizer, T_max=max(total_steps, 1), eta_min=1e-6,
    )
    if ckpt_path.exists():
        blob = _torch_load_compat(ckpt_path, map_location=device)
        if isinstance(blob.get("optimizer"), dict):
            try:
                optimizer.load_state_dict(blob["optimizer"])
            except Exception as e:
                print(f"  [warn] optimizer state not loaded: {e}", flush=True)
        if isinstance(blob.get("scheduler"), dict):
            try:
                scheduler.load_state_dict(blob["scheduler"])
            except Exception as e:
                print(f"  [warn] scheduler state not loaded: {e}", flush=True)
    return optimizer, scheduler, trainable_params


# ---------- dataset ----------
class WebSightDataset(Dataset):
    """Loads WebSight screenshot-HTML pairs from local files.

    Supports two formats:
    1. Raw files: data_dir/images/*.png + data_dir/*.html (+ optional metadata.json)
    2. HuggingFace datasets: load_from_disk format
    """

    def __init__(self, data_dir, max_samples=None):
        import json
        from PIL import Image
        data_path = Path(data_dir)

        # Check for metadata JSON first
        json_files = list(data_path.glob("*.json"))
        img_dir = data_path / "images"

        if json_files and img_dir.exists():
            # Format 1: JSON metadata + raw files
            with open(json_files[0], "r") as f:
                metadata = json.load(f)
            self.samples = []
            for item in metadata:
                img_path = data_path / "images" / f"{item['id']}.png"
                html_path = data_path / f"{item['id']}.html"
                if img_path.exists() and html_path.exists():
                    self.samples.append({
                        "image_path": str(img_path),
                        "html_path": str(html_path),
                    })
            self.mode = "files"
        elif img_dir.exists():
            # Format 1b: just images + html, no JSON
            self.samples = []
            for img_path in sorted(img_dir.glob("*.png")):
                html_path = data_path / f"{img_path.stem}.html"
                if html_path.exists():
                    self.samples.append({
                        "image_path": str(img_path),
                        "html_path": str(html_path),
                    })
            self.mode = "files"
        else:
            # Format 2: HuggingFace dataset
            from datasets import load_from_disk
            self.ds = load_from_disk(data_dir)
            if isinstance(self.ds, dict):
                self.ds = self.ds["train"]
            self.mode = "hf"

        if max_samples:
            if self.mode == "files":
                self.samples = self.samples[:max_samples]
            elif self.mode == "hf" and max_samples < len(self.ds):
                self.ds = self.ds.select(range(max_samples))

    def __len__(self):
        if self.mode == "files":
            return len(self.samples)
        return len(self.ds)

    def __getitem__(self, idx):
        from PIL import Image
        if self.mode == "files":
            s = self.samples[idx]
            image = Image.open(s["image_path"]).convert("RGB")
            html = Path(s["html_path"]).read_text(encoding="utf-8", errors="ignore")
            return {"image": image, "html": html}
        else:
            item = self.ds[idx]
            image = item["image"].convert("RGB")
            html = item["code"] if "code" in item else item.get("text", "")
            return {"image": image, "html": html}


class Design2CodeDataset(Dataset):
    """Screenshot + HTML pairs (e.g. eval subset with local GT).

    Layout A — legacy ``data_dir``:
      ``testset_final/*.png`` or ``ref_screenshots/*.png`` + ``gt_html/*.html``
      (falls back to ``websight_gt_html`` if ``gt_html`` is missing).

    Layout B — explicit ``images_dir`` + ``gt_dir`` (recommended for
    ``ref_screenshots_websight`` + ``websight_gt_html``).
    """

    def __init__(
        self,
        data_dir=None,
        max_samples=None,
        *,
        images_dir=None,
        gt_dir=None,
        require_html=True,
    ):
        self.samples = []

        if images_dir is not None and gt_dir is not None:
            img_root = Path(images_dir)
            gt_root = Path(gt_dir)
            if not img_root.is_dir():
                raise FileNotFoundError(f"images_dir not found: {img_root}")
            if not gt_root.is_dir():
                raise FileNotFoundError(f"gt_dir not found: {gt_root}")
            for png in sorted(img_root.glob("*.png")):
                gt_path = gt_root / f"{png.stem}.html"
                if require_html and not gt_path.exists():
                    continue
                self.samples.append({"image_path": png, "html_path": gt_path if gt_path.exists() else None})
                if max_samples and len(self.samples) >= max_samples:
                    break
            return

        if data_dir is None:
            raise ValueError("Design2CodeDataset: pass data_dir or (images_dir, gt_dir)")

        data_path = Path(data_dir)
        img_dir = data_path / "testset_final"
        if not img_dir.exists():
            img_dir = data_path / "ref_screenshots"

        gt_dir_resolved = None
        for name in ("gt_html", "websight_gt_html"):
            d = data_path / name
            if d.is_dir():
                gt_dir_resolved = d
                break

        if img_dir.exists():
            for png in sorted(img_dir.glob("*.png")):
                gt_path = (gt_dir_resolved / f"{png.stem}.html") if gt_dir_resolved else None
                if require_html and (gt_path is None or not gt_path.exists()):
                    continue
                self.samples.append({"image_path": png, "html_path": gt_path})
                if max_samples and len(self.samples) >= max_samples:
                    break

    def __len__(self):
        return len(self.samples)

    def __getitem__(self, idx):
        from PIL import Image
        s = self.samples[idx]
        image = Image.open(s["image_path"]).convert("RGB")
        html = ""
        if s["html_path"] and s["html_path"].exists():
            html = s["html_path"].read_text(encoding="utf-8", errors="ignore")
        return {"image": image, "html": html}


# ---------- LoRA ----------
class LoRALinear(nn.Module):
    """Simple LoRA wrapper for a frozen Linear layer."""

    def __init__(self, base_layer, r=16, alpha=32):
        super().__init__()
        self.base = base_layer
        self.r = r
        self.scale = alpha / r
        in_f = base_layer.in_features
        out_f = base_layer.out_features
        self.lora_a = nn.Linear(in_f, r, bias=False)
        self.lora_b = nn.Linear(r, out_f, bias=False)
        nn.init.kaiming_uniform_(self.lora_a.weight, a=math.sqrt(5))
        nn.init.zeros_(self.lora_b.weight)
        dtype = base_layer.weight.dtype
        self.lora_a = self.lora_a.to(dtype)
        self.lora_b = self.lora_b.to(dtype)

    @property
    def in_features(self):
        return self.base.in_features

    @property
    def out_features(self):
        return self.base.out_features

    @property
    def weight(self):
        return self.base.weight

    def forward(self, x):
        base_out = self.base(x)
        lora_out = self.lora_b(self.lora_a(x)) * self.scale
        return base_out + lora_out


# ---------- model wrapper ----------
class CompressedQwen3VL(nn.Module):
    """
    Qwen3-VL with OpticalCompressor inserted between ViT+Merger and LLM.

    Forward flow:
      1. image → model.visual() → [N, 3584] visual embeds  (frozen)
      2. [N, 3584] → OpticalCompressor → [256, 3584]        (trainable)
      3. Build inputs_embeds with 256 compressed image tokens
      4. Forward through LLM with LoRA                       (LoRA trainable)
      5. Loss on HTML output tokens
    """

    def __init__(
        self,
        model_id,
        target_tokens=256,
        lora_r=16,
        lora_alpha=32,
        max_html_tokens=2048,
        enable_lora=True,
    ):
        super().__init__()
        from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

        self.base_model = Qwen3VLForConditionalGeneration.from_pretrained(
            model_id, trust_remote_code=True, torch_dtype=torch.bfloat16,
        )
        self.processor = AutoProcessor.from_pretrained(
            model_id, trust_remote_code=True,
        )

        llm_hidden = self.base_model.config.text_config.hidden_size  # 3584 for 8B

        # Freeze everything
        for param in self.base_model.parameters():
            param.requires_grad = False

        # Add compressor
        self.compressor = OpticalCompressor(
            hidden_dim=llm_hidden,
            target_tokens=target_tokens,
        ).to(torch.bfloat16)

        # Monkey-patch get_image_features so it auto-compresses visual tokens.
        self._patch_vision(compressor=self.compressor, target_tokens=target_tokens)

        # Add LoRA to LLM decoder (can be disabled for ablation).
        self._add_lora(lora_r, lora_alpha, enable_lora=enable_lora)

        self.target_tokens = target_tokens
        self.max_html_tokens = max_html_tokens

    def _add_lora(self, r, alpha, enable_lora=True):
        """Manually inject LoRA into q_proj, v_proj of each LLM layer."""
        self.lora_modules = nn.ModuleDict()
        if (not enable_lora) or r <= 0:
            return
        lm = self.base_model.model
        if hasattr(lm, "language_model"):
            layers = lm.language_model.layers
        elif hasattr(lm, "layers"):
            layers = lm.layers
        else:
            print("WARNING: Cannot find LLM layers, skipping LoRA")
            return

        for i, layer in enumerate(layers):
            attn = layer.self_attn
            for proj_name in ["q_proj", "v_proj"]:
                orig = getattr(attn, proj_name)
                lora_key = f"layer{i}_{proj_name}"
                self.lora_modules[lora_key] = LoRALinear(
                    orig, r=r, alpha=alpha,
                )
                setattr(attn, proj_name, self.lora_modules[lora_key])

    def _patch_vision(self, compressor, target_tokens):
        """Monkey-patch get_image_features to auto-compress visual tokens.

        The compressed tokens PER IMAGE = target_tokens (not T*H*W/spatial_merge_size^2).
        We intercept get_image_features so Qwen3-VL's M-RoPE / placeholder_mask /
        masked_scatter all work unchanged — only the actual embedding values change.
        """
        import functools

        orig_get_img = self.base_model.model.get_image_features

        @functools.wraps(orig_get_img)
        def patched_get_image_features(pixel_values, image_grid_thw=None, **kwargs):
            import torch
            from transformers.models.qwen3_vl.modeling_qwen3_vl import (
                BaseModelOutputWithDeepstackFeatures,
            )
            vision_output = orig_get_img(pixel_values, image_grid_thw, **kwargs)
            pooler = vision_output.pooler_output
            if not isinstance(pooler, (list, tuple)):
                flat = pooler
            else:
                flat = torch.cat(pooler, dim=0)

            # Compute LLM-side grid after spatial merge
            vis = self.base_model.model.visual
            sms = vis.spatial_merge_size
            grid_llm = image_grid_thw.clone()
            grid_llm[:, 1] = grid_llm[:, 1] // sms
            grid_llm[:, 2] = grid_llm[:, 2] // sms
            num_images = image_grid_thw.shape[0]
            compressed_parts = []
            offset = 0
            for i in range(num_images):
                t, h, w = grid_llm[i].tolist()
                n_orig = t * h * w
                orig_slice = flat[offset: offset + n_orig]
                offset += n_orig
                comp_part, _ = compressor(orig_slice.unsqueeze(0), grid_llm[i:i+1])
                compressed_parts.append(comp_part.squeeze(0))

            compressed_flat = torch.cat(compressed_parts, dim=0)
            split_sizes = [target_tokens] * num_images

            # Reconstruct pooler_output as list (expected by Qwen3VLModel.forward)
            vision_output.pooler_output = torch.split(compressed_flat, split_sizes)
            # Disable deepstack processing to avoid dimension mismatch
            vision_output.deepstack_features = []

            return vision_output

        self.base_model.model.get_image_features = patched_get_image_features

    def prepare_inputs(self, images, htmls):
        """Prepare model inputs for a batch of image-html pairs."""
        batch_messages = []
        for img in images:
            batch_messages.append([{"role": "user", "content": [
                {"type": "image", "image": img},
                {"type": "text", "text": UI2CODE_PROMPT},
            ]}])

        texts = [
            self.processor.apply_chat_template(
                msg, tokenize=False, add_generation_prompt=True,
            )
            for msg in batch_messages
        ]

        inputs = self.processor(
            text=texts, images=images, return_tensors="pt", padding=True,
        )
        return inputs

    def forward(self, images, htmls, device):
        """Full training forward pass. Returns loss scalar.

        The _patch_vision monkey-patch auto-compresses visual tokens inside
        get_image_features, so Qwen3-VL's forward handles M-RoPE, placeholder_mask,
        and masked_scatter automatically. We only supply the HTML teacher targets.
        """
        inputs = self.prepare_inputs(images, htmls)
        pixel_values = inputs["pixel_values"].to(device, torch.bfloat16)
        image_grid_thw = inputs["image_grid_thw"].to(device)
        input_ids = inputs["input_ids"].to(device)

        # Build HTML target (teacher forcing)
        new_input_ids, labels = self._rebuild_sequence(input_ids, htmls, device)

        # Attention mask
        pad_token_id = self.processor.tokenizer.pad_token_id or 0
        attention_mask = (new_input_ids != pad_token_id).long()

        # Forward through Qwen3-VL.
        # input_ids: provides #image tokens for M-RoPE / placeholder_mask.
        # pixel_values + image_grid_thw: trigger patched get_image_features
        #   which returns compressed embeddings → auto masked-scattered.
        # output_hidden_states=False: disables deepstack_features mismatch.
        outputs = self.base_model(
            input_ids=new_input_ids,
            attention_mask=attention_mask,
            pixel_values=pixel_values,
            image_grid_thw=image_grid_thw,
            output_hidden_states=False,
            labels=labels,
        )
        return outputs.loss

    def _rebuild_sequence(self, orig_input_ids, htmls, device):
        """Replace variable-length image tokens with fixed target_tokens placeholders."""
        tokenizer = self.processor.tokenizer
        batch_new_ids = []
        batch_labels = []
        max_len = 0

        for b in range(orig_input_ids.shape[0]):
            ids = orig_input_ids[b].tolist()

            img_positions = [i for i, t in enumerate(ids) if t == IMAGE_TOKEN_ID]

            if img_positions:
                before = ids[:img_positions[0]]
                after = ids[img_positions[-1] + 1:]
            else:
                before = ids
                after = []

            new_ids = before + [IMAGE_TOKEN_ID] * self.target_tokens + after

            html_tokens = tokenizer.encode(
                htmls[b],
                add_special_tokens=False,
                max_length=self.max_html_tokens,
                truncation=True,
            )
            html_tokens.append(tokenizer.eos_token_id)

            full_ids = new_ids + html_tokens
            labels = [-100] * len(new_ids) + html_tokens

            batch_new_ids.append(full_ids)
            batch_labels.append(labels)
            max_len = max(max_len, len(full_ids))

        pad_id = tokenizer.pad_token_id or 0
        for i in range(len(batch_new_ids)):
            pad_len = max_len - len(batch_new_ids[i])
            batch_new_ids[i] += [pad_id] * pad_len
            batch_labels[i] += [-100] * pad_len

        new_input_ids = torch.tensor(batch_new_ids, device=device)
        labels = torch.tensor(batch_labels, device=device)
        return new_input_ids, labels

    def _scatter_visual(self, text_embeds, input_ids, compressed, grid_thw):
        """Replace image placeholder embeddings with compressed visual tokens."""
        inputs_embeds = text_embeds.clone()
        offset = 0
        for b in range(input_ids.shape[0]):
            img_mask = input_ids[b] == IMAGE_TOKEN_ID
            n_img = img_mask.sum().item()
            if n_img > 0 and offset < compressed.shape[0]:
                n_take = min(n_img, compressed.shape[0] - offset)
                img_positions = img_mask.nonzero(as_tuple=True)[0][:n_take]
                inputs_embeds[b, img_positions] = compressed[offset:offset + n_take]
                offset += n_take
        return inputs_embeds

    def _build_position_ids(self, input_ids, grid_thw, device):
        """
        Build 3D position IDs compatible with Qwen3-VL's M-RoPE.
        Vectorized — no Python loops over individual tokens.
        """
        batch_size, seq_len = input_ids.shape
        position_ids = torch.zeros(batch_size, 3, seq_len, dtype=torch.long, device=device)

        for b in range(batch_size):
            ids = input_ids[b]
            img_mask = ids == IMAGE_TOKEN_ID
            text_mask = ~img_mask

            # Text positions before image: sequential
            img_positions = img_mask.nonzero(as_tuple=True)[0]
            text_positions = text_mask.nonzero(as_tuple=True)[0]

            if len(img_positions) == 0:
                # No image tokens — just sequential
                pos = torch.arange(seq_len, device=device)
                position_ids[b, 0] = pos
                position_ids[b, 1] = pos
                position_ids[b, 2] = pos
                continue

            first_img = img_positions[0].item()
            last_img = img_positions[-1].item()

            # Text before image: sequential (same across all 3 dims)
            text_before_mask = text_mask.clone()
            text_before_mask[last_img + 1:] = False
            text_before_positions = text_before_mask.nonzero(as_tuple=True)[0]
            n_text_before = text_before_positions.shape[0]
            if n_text_before > 0:
                seq_vals = torch.arange(n_text_before, device=device)
                position_ids[b, 0, text_before_positions] = seq_vals
                position_ids[b, 1, text_before_positions] = seq_vals
                position_ids[b, 2, text_before_positions] = seq_vals

            text_offset = n_text_before

            # Image positions: 2D grid layout
            t, h, w = grid_thw[b % grid_thw.shape[0]].tolist()
            t, h, w = int(t), int(h), int(w)
            n_img = len(img_positions)

            idx = torch.arange(n_img, device=device)
            hw = h * w
            ti = idx // hw
            hi = (idx % hw) // w
            wi = idx % w

            position_ids[b, 0, img_positions] = text_offset + ti
            position_ids[b, 1, img_positions] = text_offset + hi
            position_ids[b, 2, img_positions] = text_offset + wi

            # Text after image: sequential, offset past image grid
            img_max_pos = text_offset + max(h, w)
            text_after_positions = text_mask[last_img + 1:].nonzero(as_tuple=True)[0] + last_img + 1
            if len(text_after_positions) > 0:
                after_vals = torch.arange(len(text_after_positions), device=device) + img_max_pos + 1
                position_ids[b, 0, text_after_positions] = after_vals
                position_ids[b, 1, text_after_positions] = after_vals
                position_ids[b, 2, text_after_positions] = after_vals

        return position_ids


# ---------- training loop ----------
def train(args):
    # DDP setup
    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    is_distributed = world_size > 1

    if is_distributed:
        dist.init_process_group("nccl", device_id=torch.device(f"cuda:{local_rank}"))
        torch.cuda.set_device(local_rank)
    device = torch.device(f"cuda:{local_rank}")
    is_main = local_rank == 0

    if is_main:
        print(f"=== UIPress Optical Compressor Training ===", flush=True)
        print(f"  GPUs: {world_size}, target_tokens: {args.target_tokens}", flush=True)
        if args.disable_lora:
            print("  LoRA: disabled", flush=True)
        else:
            print(f"  LoRA r={args.lora_r}, alpha={args.lora_alpha}", flush=True)
        print(f"  Batch: {args.batch_size} x {args.grad_accum} x {world_size} "
              f"= {args.batch_size * args.grad_accum * world_size}", flush=True)
        if args.eval_after_epoch and is_distributed:
            print(
                "  Warning: --eval_after_epoch is disabled under DDP (single-GPU only).",
                flush=True,
            )

    # Load model
    log_all("Loading model...")
    t0 = time.time()
    model = CompressedQwen3VL(
        model_id=args.model_id,
        target_tokens=args.target_tokens,
        lora_r=args.lora_r,
        lora_alpha=args.lora_alpha,
        max_html_tokens=args.max_html_tokens,
        enable_lora=not args.disable_lora,
    )
    model.base_model.to(device)
    model.compressor.to(device)
    log_all(f"Model loaded in {time.time() - t0:.1f}s")

    lora_trainable = [p for p in model.lora_modules.parameters() if p.requires_grad]

    # Count trainable params
    if is_main:
        comp_params = model.compressor.count_parameters()
        lora_params = sum(p.numel() for p in lora_trainable)
        print(f"  Compressor params: {comp_params['trainable']:,}", flush=True)
        print(f"  LoRA params: {lora_params:,}", flush=True)
        print(f"  Total trainable: {comp_params['trainable'] + lora_params:,}", flush=True)

    # Collect all trainable parameters
    trainable_params = (
        list(model.compressor.parameters())
        + lora_trainable
    )

    # No DDP wrapper — LoRA is injected via setattr into base_model,
    # so DDP can't track it. Instead we manually allreduce gradients.
    if is_distributed:
        log_all("Syncing ranks...")
        dist.barrier()
        log_all("Ranks synced")

    # Dataset (load on all ranks — each rank reads from same disk)
    log_all("Loading dataset...")
    t0 = time.time()
    dataset = WebSightDataset(args.data_dir, max_samples=args.max_samples)
    if args.mix_root:
        mix_root = (PROJECT_ROOT / args.mix_root).resolve()
        mix = Design2CodeDataset(
            images_dir=mix_root / args.mix_images_subdir,
            gt_dir=mix_root / args.mix_gt_subdir,
            max_samples=args.mix_max_samples,
            require_html=True,
        )
        if is_main:
            print(f"  Mix split: {len(mix)} pairs from {mix_root / args.mix_images_subdir}", flush=True)
        if len(mix) > 0:
            dataset = ConcatDataset([dataset, mix])
        elif is_main:
            print(
                "  Warning: mix_root set but 0 samples with GT; train on WebSight only. "
                "Check mix_images_subdir / mix_gt_subdir.",
                flush=True,
            )
    log_all(f"Dataset loaded: {len(dataset)} samples in {time.time() - t0:.1f}s")

    sampler = DistributedSampler(dataset) if is_distributed else None
    loader = DataLoader(
        dataset, batch_size=args.batch_size, sampler=sampler,
        shuffle=(sampler is None), num_workers=0, pin_memory=True,
        collate_fn=lambda batch: batch,
    )

    # Optimizer — use trainable_params collected before DDP
    optim_groups = [{"params": list(model.compressor.parameters()), "lr": args.lr_compressor}]
    if lora_trainable:
        optim_groups.append({"params": lora_trainable, "lr": args.lr_lora})
    optimizer = torch.optim.AdamW(optim_groups, weight_decay=0.01)

    total_steps = len(loader) * args.epochs // args.grad_accum
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
        optimizer, T_max=max(total_steps, 1), eta_min=1e-6,
    )

    # Resume
    start_epoch = 0
    if args.resume and Path(args.resume).exists():
        ckpt = _torch_load_compat(args.resume, map_location=device)
        comp_state = ckpt.get("compressor", ckpt)
        new_state = {}
        for k, v in comp_state.items():
            new_state[k.replace("module.", "")] = v
        model.compressor.load_state_dict(new_state)
        if "lora" in ckpt and len(model.lora_modules) > 0:
            model.lora_modules.load_state_dict(ckpt["lora"])
        if "epoch" in ckpt:
            start_epoch = ckpt["epoch"] + 1
        if isinstance(ckpt.get("optimizer"), dict):
            try:
                optimizer.load_state_dict(ckpt["optimizer"])
            except Exception as e:
                if is_main:
                    print(f"  [warn] resume: optimizer not restored: {e}", flush=True)
        if isinstance(ckpt.get("scheduler"), dict):
            try:
                scheduler.load_state_dict(ckpt["scheduler"])
            except Exception as e:
                if is_main:
                    print(f"  [warn] resume: scheduler not restored: {e}", flush=True)
        if is_main:
            print(f"  Resumed from {args.resume}, next epoch index {start_epoch}", flush=True)

    # Output dir
    out_dir = Path(args.output_dir)
    out_dir.mkdir(parents=True, exist_ok=True)

    # Training
    if is_main:
        print(f"\n--- Training starts ---", flush=True)
        print(f"  Steps/epoch: {len(loader)}, total optimizer steps: {total_steps}", flush=True)

    for epoch in range(start_epoch, args.epochs):
        if sampler is not None:
            sampler.set_epoch(epoch)

        model.compressor.train()
        for m in model.lora_modules.values():
            m.train()

        epoch_loss = 0.0
        n_steps = 0
        optimizer.zero_grad()
        t_epoch = time.time()

        for step, batch in enumerate(loader):
            images = [s["image"] for s in batch]
            htmls = [s["html"] for s in batch]

            try:
                t_step = time.time()
                loss = model.forward(images, htmls, device)
                loss = loss / args.grad_accum
                loss.backward()

                epoch_loss += loss.item() * args.grad_accum
                n_steps += 1

                # Print first step immediately + every grad_accum steps
                if is_main and (step == 0 or (step + 1) % args.grad_accum == 0):
                    elapsed = time.time() - t_step
                    avg = epoch_loss / n_steps if n_steps > 0 else 0
                    mem = torch.cuda.max_memory_allocated() / 1024**3
                    print(f"  E{epoch} S{step+1}/{len(loader)} "
                          f"loss={avg:.4f} step_time={elapsed:.1f}s "
                          f"mem={mem:.1f}GB", flush=True)

            except RuntimeError as e:
                if "out of memory" in str(e):
                    log_all(f"OOM at step {step}, skipping")
                    torch.cuda.empty_cache()
                    optimizer.zero_grad()
                    continue
                raise

            if (step + 1) % args.grad_accum == 0:
                # Manual allreduce gradients across GPUs
                if is_distributed:
                    for p in trainable_params:
                        if p.grad is not None:
                            dist.all_reduce(p.grad, op=dist.ReduceOp.AVG)

                torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)
                optimizer.step()
                scheduler.step()
                optimizer.zero_grad()

        avg_loss = epoch_loss / max(n_steps, 1)
        epoch_time = time.time() - t_epoch
        if is_main:
            print(f"  Epoch {epoch}: avg_loss={avg_loss:.4f} time={epoch_time/60:.1f}min", flush=True)
            ckpt = {
                "compressor": model.compressor.state_dict(),
                "lora": model.lora_modules.state_dict(),
                "epoch": epoch,
                "loss": avg_loss,
                "args": vars(args),
                "optimizer": optimizer.state_dict(),
                "scheduler": scheduler.state_dict(),
            }
            torch.save(ckpt, out_dir / f"epoch{epoch}.pt")
            torch.save(ckpt, out_dir / "latest.pt")
            print(f"  Saved checkpoint: {out_dir / f'epoch{epoch}.pt'}", flush=True)

        do_eval = args.eval_after_epoch and (not is_distributed) and is_main
        if do_eval:
            if is_main:
                print(f"  Running post-epoch eval (freeing GPU)...", flush=True)
            _uipress_stack_to_cpu(model)
            gc.collect()
            torch.cuda.empty_cache()
            try:
                _run_subprocess_eval_and_clip(args, out_dir, epoch)
            finally:
                _uipress_stack_to_device(model, device)
            optimizer, scheduler, trainable_params = _rebuild_optimizer_scheduler(
                args, model, device, total_steps, out_dir / "latest.pt",
            )
            if is_main:
                print(f"  Restored optimizer/scheduler from latest.pt", flush=True)

    if is_distributed:
        dist.destroy_process_group()

    if is_main:
        print("\n=== Training complete ===", flush=True)


def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--model_id", default="Qwen/Qwen3-VL-8B-Instruct")
    p.add_argument("--data_dir", default="data/websight")
    p.add_argument("--output_dir", default="checkpoints/optical")
    p.add_argument(
        "--max_samples",
        type=int,
        default=10000,
        help="Cap WebSight training pairs (lower = much faster epochs).",
    )
    p.add_argument("--epochs", type=int, default=5)
    p.add_argument(
        "--max_html_tokens",
        type=int,
        default=2048,
        help="Teacher HTML token cap (run scripts/stats_html_token_lengths.py). Mixing Design2Code GT often needs 8192+.",
    )
    p.add_argument("--batch_size", type=int, default=1)
    p.add_argument("--grad_accum", type=int, default=8)
    p.add_argument("--lr_compressor", type=float, default=2e-4)
    p.add_argument("--lr_lora", type=float, default=2e-5)
    p.add_argument("--target_tokens", type=int, default=256)
    p.add_argument("--lora_r", type=int, default=16)
    p.add_argument("--lora_alpha", type=int, default=32)
    p.add_argument(
        "--disable_lora",
        action="store_true",
        help="Disable LoRA adapters; train compressor-only for ablation.",
    )
    p.add_argument("--resume", type=str, default=None)
    p.add_argument(
        "--mix_root",
        default=None,
        help="If set (e.g. data), concat extra (image,html) pairs under mix_images_subdir + mix_gt_subdir.",
    )
    p.add_argument(
        "--mix_images_subdir",
        default="ref_screenshots_websight",
        help="Relative to mix_root; default matches data/ref_screenshots_websight + websight_gt_html.",
    )
    p.add_argument("--mix_gt_subdir", default="websight_gt_html")
    p.add_argument(
        "--mix_max_samples",
        type=int,
        default=None,
        help="Cap mixed split size (default: all pairs that have GT).",
    )
    p.add_argument(
        "--eval_after_epoch",
        action="store_true",
        help="After each epoch: save ckpt, free GPU, run eval_all (uipress) + CLIP (single-GPU only).",
    )
    p.add_argument("--eval_max_samples", type=int, default=50)
    p.add_argument(
        "--eval_output_dir",
        default=None,
        help="Defaults to results/clip_per_epoch/<output_dir name>.",
    )
    p.add_argument("--eval_data_dir", default="data")
    p.add_argument("--eval_ref_dir", default="data/ref_screenshots")
    p.add_argument(
        "--eval_clip_device",
        default="cuda",
        choices=["cuda", "cpu"],
        help="Device for CLIP ViT in post-epoch scoring.",
    )
    return p.parse_args()


if __name__ == "__main__":
    train(parse_args())