other_transfer/src/physics_model/load.py

"""Load the local Qwen2.5-VL checkpoint and processor. No training logic."""

from __future__ import annotations

import logging
from pathlib import Path
from typing import Any

import torch

logger = logging.getLogger(__name__)

VISION_TOKEN_IDS: tuple[int, ...] = (151652, 151653, 151654, 151655)
DEFAULT_MODEL_DIR = Path("third-party/qwen")
PROCESSOR_COPY_NAMES = (
    "preprocessor_config.json",
    "chat_template.json",
    "tokenizer.json",
    "tokenizer_config.json",
    "vocab.json",
    "merges.txt",
    "special_tokens_map.json",
    "generation_config.json",
)


def resolve_attn_implementation(preferred: str | None) -> str:
    if not preferred or preferred == "sdpa":
        return "sdpa"
    if preferred == "flash_attention_2":
        try:
            import flash_attn  # noqa: F401

            return "flash_attention_2"
        except Exception:
            logger.warning("flash_attention_2 requested but unavailable; using sdpa")
            return "sdpa"
    return preferred


def _dtype(name: str | None) -> torch.dtype | str:
    if name in (None, "auto"):
        return "auto"
    return getattr(torch, name)


def load_processor(model_dir: str | Path, min_pixels: int, max_pixels: int):
    from transformers import AutoProcessor

    processor = AutoProcessor.from_pretrained(str(model_dir), trust_remote_code=True)
    if hasattr(processor, "image_processor") and processor.image_processor is not None:
        processor.image_processor.min_pixels = min_pixels
        processor.image_processor.max_pixels = max_pixels
    return processor


def language_lora_target_names(model) -> list[str]:
    suffixes = ("q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj")
    skip = ("visual", "vision", "merger")
    names: list[str] = []
    for name, _module in model.named_modules():
        if not name.endswith(suffixes):
            continue
        lowered = name.lower()
        if any(part in lowered for part in skip):
            continue
        names.append(name)
    return names


def freeze_vision(model) -> int:
    frozen = 0
    for name, param in model.named_parameters():
        lowered = name.lower()
        if any(part in lowered for part in ("visual", "vision", "merger")):
            param.requires_grad = False
            frozen += 1
    return frozen


def load_model(
    model_dir: str | Path,
    *,
    torch_dtype: str = "bfloat16",
    attn_implementation: str | None = "flash_attention_2",
    gradient_checkpointing: bool = True,
    use_cache: bool | None = None,
) -> Any:
    try:
        from transformers import Qwen2_5_VLForConditionalGeneration
    except ImportError:  # newer transformers fold the class into AutoModel
        from transformers import AutoModelForImageTextToText as Qwen2_5_VLForConditionalGeneration

    attn = resolve_attn_implementation(attn_implementation)
    kwargs: dict[str, Any] = {
        "trust_remote_code": True,
        "torch_dtype": _dtype(torch_dtype),
        "attn_implementation": attn,
    }
    # Transformers 5.x Qwen2.5-VL rejects use_cache on from_pretrained; set it on config.
    cache_on = (not gradient_checkpointing) if use_cache is None else bool(use_cache)
    try:
        model = Qwen2_5_VLForConditionalGeneration.from_pretrained(str(model_dir), **kwargs)
    except TypeError:
        kwargs.pop("torch_dtype", None)
        kwargs["dtype"] = _dtype(torch_dtype)
        model = Qwen2_5_VLForConditionalGeneration.from_pretrained(str(model_dir), **kwargs)
    if hasattr(model, "config"):
        model.config.use_cache = cache_on
    if gradient_checkpointing and hasattr(model, "gradient_checkpointing_enable"):
        model.gradient_checkpointing_enable()
        if hasattr(model, "config"):
            model.config.use_cache = False
    frozen = freeze_vision(model)
    logger.info("Loaded %s attn=%s frozen_vision_tensors=%s", model_dir, attn, frozen)
    return model


def copy_processor_files(model_dir: str | Path, output_dir: str | Path) -> list[str]:
    src = Path(model_dir)
    dst = Path(output_dir)
    dst.mkdir(parents=True, exist_ok=True)
    copied: list[str] = []
    for name in PROCESSOR_COPY_NAMES:
        src_file = src / name
        if src_file.is_file():
            target = dst / name
            target.write_bytes(src_file.read_bytes())
            copied.append(name)
    return copied