Bug: AutoModelForVision2Seq.from_config() Compatibility Issue

[Maxwell-Jia]

Description

When using FSDPCheckpointManager to save checkpoints for Vision Language Models (e.g., Qwen2.5-VL-3B-Instruct) with the hf_model save option enabled, a TypeError occurs due to incompatible parameters in the from_config() method.

Error Message

TypeError: AutoModelForVision2Seq.from_config() got an unexpected keyword argument 'torch_dtype'

Error Location

File: verl/utils/checkpoint/fsdp_checkpoint_manager.py
Line: 331

save_model = auto_model_cls.from_config(model_config, torch_dtype=torch.bfloat16)

Environment

• transformers version: 4.55.4
• Python version: 3.10
• PyTorch version: 2.7.1
• CUDA version: 12.6

Steps to Reproduce

1. Load a Vision Language Model (e.g., Qwen2.5-VL-3B-Instruct)
2. Wrap the model with FSDP
3. Create FSDPCheckpointManager with hf_model in save configuration:

   checkpoint_config = DictConfig({
       'save_contents': ['model', 'hf_model'],
       'load_contents': ['model']
   })

4. Call checkpoint_manager.save_checkpoint()

Minimal Reproduction Script

#!/usr/bin/env python3
import os
import tempfile
import torch
import torch.distributed as dist
import warnings
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy
from transformers import AutoConfig, AutoModelForVision2Seq, AutoProcessor
from verl.utils.checkpoint.fsdp_checkpoint_manager import FSDPCheckpointManager
from omegaconf import DictConfig

def setup_distributed():
    if not dist.is_initialized():
        os.environ['CUDA_VISIBLE_DEVICES'] = '0'
        os.environ['MASTER_ADDR'] = 'localhost'
        os.environ['MASTER_PORT'] = '12355'
        os.environ['RANK'] = '0'
        os.environ['WORLD_SIZE'] = '1'
        dist.init_process_group(backend='nccl' if torch.cuda.is_available() else 'gloo')

def main():
    setup_distributed()
    
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"  # or local path
    
    config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
    processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
    
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        model = AutoModelForVision2Seq.from_pretrained(
            model_name, 
            torch_dtype=torch.bfloat16,
            trust_remote_code=True,
            device_map={"": device}
        )
    
    auto_wrap_policy = size_based_auto_wrap_policy
    fsdp_model = FSDP(
        model,
        auto_wrap_policy=auto_wrap_policy,
        mixed_precision=None,
        device_id=0 if torch.cuda.is_available() else None,
        use_orig_params=True,
    )
    
    checkpoint_config = DictConfig({
        'save_contents': ['model', 'hf_model'],
        'load_contents': ['model']
    })
    
    checkpoint_manager = FSDPCheckpointManager(
        model=fsdp_model,
        optimizer=None,
        lr_scheduler=None,
        processing_class=processor,
        checkpoint_config=checkpoint_config
    )
    
    import shutil
    temp_dir = tempfile.mkdtemp()
    try:
        checkpoint_path = os.path.join(temp_dir, "test_checkpoint")
        checkpoint_manager.save_checkpoint(local_path=checkpoint_path, global_step=1)
    finally:
        if os.path.exists(temp_dir):
            shutil.rmtree(temp_dir, ignore_errors=True)
    
    if dist.is_initialized():
        dist.destroy_process_group()

if __name__ == "__main__":
    main()

Root Cause

The issue occurs at fsdp_checkpoint_manager.py:331:

save_model = auto_model_cls.from_config(model_config, torch_dtype=torch.bfloat16)

In transformers>=4.54.0, the AutoModelForVision2Seq.from_config() method does not accept the torch_dtype parameter, and AutoModelForVision2Seq is deprecated and will be removed in v5.0. We should use "AutoModelForImageTextToText instead."