[Qwen2.5] LoRA with SFT seems to be stuck forever with DeepSpeed

[sayakpaul]

Training command:

accelerate launch --config_file=deepspeed_zero3.yaml train.py \
  --dataset_name diffusers-internal-dev/ShotDEAD-single-shard \
  --model_name_or_path $MODEL_NAME \
  --attn_implementation "sdpa" \
  --per_device_train_batch_size 4 \
  --gradient_accumulation_steps 4 \
  --output_dir $OUTPUT_DIR \
  --bf16 \
  --use_peft \
  --torch_dtype bfloat16 \
  --gradient_checkpointing

train.py

"""
Adapted from 
https://github.com/huggingface/trl/blob/822653824bf084bc6c042cf0e759f86187c92569/examples/scripts/sft_vlm.py
"""

import torch
from datasets import load_dataset
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from PIL import Image 
import io

from trl import (
    ModelConfig,
    ScriptArguments,
    SFTConfig,
    SFTTrainer,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()
    training_args.gradient_checkpointing_kwargs = dict(use_reentrant=False)
    training_args.remove_unused_columns = False
    training_args.dataset_kwargs = {"skip_prepare_dataset": True}

    ################
    # Model, Tokenizer & Processor
    ################
    torch_dtype = (
        model_args.torch_dtype if model_args.torch_dtype in ["auto", None] else getattr(torch, model_args.torch_dtype)
    )
    quantization_config = get_quantization_config(model_args)
    model_kwargs = dict(
        revision=model_args.model_revision,
        attn_implementation=model_args.attn_implementation,
        torch_dtype=torch_dtype,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
    )
    processor = AutoProcessor.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
    )

    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, **model_kwargs
    )

    ################
    # Create a data collator to encode text and image pairs
    ################
    def collator_fn(examples):
        # Get the texts and images, and apply the chat template
        texts = [
            processor.apply_chat_template(example["messages"], tokenize=False) 
            for example in examples
        ]
        images = [Image.open(io.BytesIO(example["image"])).convert("RGB") for example in examples]

        # Tokenize the texts and process the images
        batch = processor(text=texts, images=images, return_tensors="pt", padding=True)
        print(f"{batch.keys()=}")

        # The labels are the input_ids, and we mask the padding tokens in the loss computation
        labels = batch["input_ids"].clone()
        labels[labels == processor.tokenizer.pad_token_id] = -100  #
        # Ignore the image token index in the loss computation (model specific)
        image_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_token)
        labels[labels == image_token_id] = -100
        batch["labels"] = labels

        return batch

    ################
    # Dataset
    ################
    dataset = load_dataset(script_args.dataset_name, name=script_args.dataset_config)["train"]
    splits = dataset.train_test_split(0.1)
    train, val = splits["train"], splits["test"]

    def filter(example):
        can_load = False
        try:
            _ = Image.open(io.BytesIO(example["image"]))
            can_load = True
        except: 
            pass 
        return can_load

    train = train.filter(filter)
    val = val.filter(filter)
    
    ################
    # Training
    ################
    model_args.lora_modules_to_save = ["lm_head", "embed_token"]
    model_args.lora_target_modules = "all-linear"
    trainer = SFTTrainer(
        model=model,
        args=training_args,
        data_collator=collator_fn,
        train_dataset=train,
        eval_dataset=val,
        processing_class=processor.tokenizer,
        peft_config=get_peft_config(model_args),
    )

    trainer.train()

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)
        if trainer.accelerator.is_main_process:
            processor.push_to_hub(training_args.hub_model_id)

Referenced from:
https://github.com/huggingface/trl/blob/main/examples/scripts/sft_vlm.py

I am on latest trl but using the latest main installations of peft and transformers. Tested this on 8xH100s.