Vision Fine Tuning Gemma 3 takes Impossiblily High VRam (OOM Error 8xH200)

[amanmehra89]

Reproduction

I'm trying to fientune Gemma 3 on a vision dataset, based on https://github.com/huggingface/trl/blob/main/examples/scripts/sft_vlm_gemma3.py

How ever i get out of memory issue even if i try with 2-3 training samples..
Below is the Code i used

# Copyright 2020-2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""
Train Gemma-3 on the HuggingFaceH4/llava-instruct-mix-vsft dataset (single-image).

accelerate launch \
    --config_file examples/accelerate_configs/deepspeed_zero3.yaml \
    examples/scripts/sft_vlm_gemma3.py \
    --dataset_name HuggingFaceH4/llava-instruct-mix-vsft \
    --model_name_or_path google/gemma-3-4b-it \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --output_dir gemma-3-4b-it-trl-sft-llava-instruct-mix-vsft \
    --bf16 \
    --torch_dtype bfloat16 \
    --use_peft \
    --lora_target_modules all-linear \
    --attn_implementation eager

Train Gemma-3 on the FanqingM/MMIU-Benchmark dataset (multi-image).

accelerate launch \
    --config_file /home/ubuntu/gemma-finetune/examples/accelerate_configs/deepspeed_zero3.yaml \
    /home/ubuntu/gemma-finetune/scripts/sft_vlm_gemma3.py \
    --dataset_name FanqingM/MMIU-Benchmark \
    --dataset_train_split test \
    --model_name_or_path google/gemma-3-4b-it \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --output_dir gemma-3-4b-it-trl-sft-MMIU-Benchmark \
    --bf16 \
    --torch_dtype bfloat16 \
    --use_peft \
    --lora_target_modules all-linear
    --attn_implementation eager


    
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
accelerate launch --num_processes 4 --config_file /home/ubuntu/gemma-finetune/examples/accelerate_configs/deepspeed_zero3.yaml /home/ubuntu/gemma-finetune/scripts/sft_vlm_gemma3.py --dataset_name FanqingM/MMIU-Benchmark --dataset_train_split test --model_name_or_path google/gemma-3-4b-it --per_device_train_batch_size 1 --gradient_accumulation_steps 4 --bf16 --torch_dtype bfloat16 --use_peft --lora_target_modules q_proj,v_proj,k_proj,o_proj --attn_implementation eager

    accelerate launch   --num_processes 4   --config_file /home/ubuntu/gemma-finetune/examples/accelerate_configs/deepspeed_zero3.yaml   /home/ubuntu/gemma-finetune/scripts/sft_vlm_gemma3.py   --dataset_name FanqingM/MMIU-Benchmark   --dataset_train_split test   --model_name_or_path google/gemma-3-4b-it   --per_device_train_batch_size 1   --gradient_accumulation_steps 1   --output_dir gemma-3-4b-it-trl-sft-MMIU-Benchmark   --bf16   --torch_dtype bfloat16   --use_peft   --lora_target_modules all-linear   --attn_implementation eager
"""

import io
import os
import zipfile

import torch
from datasets import DatasetDict, load_dataset
from huggingface_hub import hf_hub_download, list_repo_files
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfig
from peft import LoraConfig, PeftConfig
import json
from datasets import Dataset, DatasetDict
from accelerate import Accelerator
accelerator = Accelerator()
print(f"Running on rank: {accelerator.local_process_index}, world size: {accelerator.num_processes}")
rank = accelerator.local_process_index

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


# For multi-image example
def process_vision_info(messages: list[dict]) -> list[Image.Image]:
    image_inputs = []
    for msg in messages:
        content = msg.get("content", [])
        if not isinstance(content, list):
            content = [content]

        for element in content:
            if isinstance(element, dict) and ("image" in element or element.get("type") == "image"):
                if "image" in element:
                    image = element["image"]
                else:
                    image = element
                if image is not None:
                    image = Image.open(io.BytesIO(image["bytes"]))
                    image_inputs.append(image.convert("RGB"))
    return image_inputs


def format_data(samples: dict[str, any]) -> dict[str, list]:
    print(f"Rank {rank}: Processing samples: {samples['question']}")
    formatted_samples = {"messages": []}
    for cont in range(len(samples["question"])):
        images = []
        max_images = 10
        for img_path in samples["input_image_path"][cont][:max_images]:
            try:
                with open(img_path, "rb") as f:
                    img_bytes = f.read()
                image = Image.open(io.BytesIO(img_bytes)).convert("RGB")
                images.append({"type": "image", "image": image})
            except Exception as e:
                print(f"Error processing image {img_path}: {e}")
                continue

        formatted_samples["messages"].append(
            [
                {"role": "system", "content": [{"type": "text", "text": samples["context"][cont]}]},
                {"role": "user", "content": images + [{"type": "text", "text": samples["question"][cont]}]},
                {"role": "assistant", "content": [{"type": "text", "text": samples["output"][cont]}]},
            ]
        )
        print(formatted_samples)
    return formatted_samples


# For multi-image example
def prepare_dataset(dataset: DatasetDict, dataset_name: str, dataset_train_split: str) -> DatasetDict:
    all_files = list_repo_files(dataset_name, repo_type="dataset")
    zip_files = [f for f in all_files if f.endswith(".zip")]

    for zip_filename in zip_files:
        zip_path = hf_hub_download(repo_id=dataset_name, filename=zip_filename, repo_type="dataset")
        extract_folder = zip_filename.replace(".zip", "")
        if os.path.exists(extract_folder):
            continue
        else:
            os.makedirs(extract_folder, exist_ok=True)

            with zipfile.ZipFile(zip_path, "r") as zip_ref:
                zip_ref.extractall(extract_folder)
    dataset['test'] = dataset['test'].select(range(1))
    print(dataset)
    dataset = dataset.map(format_data, batched=True, batch_size=1)
    # dataset.save_to_disk("cached_dataset")
    return dataset

def prepare_custom_dataset(path):
    with open(path, "r") as f:
        data = json.load(f)

    # If you want to create a train/test split manually,
    # for example split 80/20:
    train_size = 0 # int(0.8 * len(data))
    train_data = data[:train_size]
    test_data = data[train_size:]

    # Create Hugging Face Dataset objects
    train_dataset = Dataset.from_list(train_data)
    test_dataset = Dataset.from_list(test_data)

    # Wrap into DatasetDict
    dataset = DatasetDict({
        #"train": train_dataset,
        "test": test_dataset,
    })
    print(dataset)
    dataset = dataset.map(format_data, batched=True, batch_size=1)
    return dataset


def 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}


    # START: Added checkpoint saving configuration
    training_args.save_strategy = "epoch"    
    training_args.save_total_limit = 3   
    # END: Added checkpoint saving configuration
  
     
 
    ################
    # Model, Tokenizer & Processor
    print("Loading model, tokenizer and 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)
    #print(quantization_config)
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,  # or False for 8-bit
        bnb_4bit_quant_type="nf4",  # or "fp4"
        bnb_4bit_use_double_quant=True,
        bnb_4bit_compute_dtype="bfloat16",  # or torch.float16
    )
    model_kwargs = dict(
        revision=model_args.model_revision,
        attn_implementation=model_args.attn_implementation,
        torch_dtype=torch_dtype,
        #device_map='auto',
        quantization_config=bnb_config,
    )
    processor = AutoProcessor.from_pretrained(
    model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, use_fast=True
    )
    processor.tokenizer.padding_side = "right"
    model = AutoModelForImageTextToText.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, **model_kwargs
    )
    print("Model, tokenizer and processor loaded successfully!")
    def collate_fn(examples):
        texts = [
            processor.apply_chat_template(example["messages"], tokenize=False, add_generation_prompt=False).strip()
            for example in examples
        ]
        if "images" in examples[0]:  # single-image
            images = [[img.convert("RGB") for img in example["images"]] for example in examples]
        else:  # multi-image
            images = [process_vision_info(example["messages"]) for example in examples]

        # Tokenize the texts and process the images
        batch = processor(
            text=texts, images=images, return_tensors="pt", padding=True
        )  # Encode texts and images into tensors

        # The labels are the input_ids, and we mask the padding tokens in the loss computation
        labels = batch["input_ids"].clone()  # Clone input IDs for labels
        # Mask image tokens
        image_token_id = [
            processor.tokenizer.convert_tokens_to_ids(processor.tokenizer.special_tokens_map["boi_token"])
        ]
        # Mask tokens for not being used in the loss computation
        labels[labels == processor.tokenizer.pad_token_id] = -100
        labels[labels == image_token_id] = -100
        labels[labels == 262144] = -100

        batch["labels"] = labels
        return batch  # Return the prepared batch

    ################
    # Dataset
    ################

    print("Loading dataset...")
    ## DEFAULT DATASET
    dataset = load_dataset(script_args.dataset_name, name=script_args.dataset_config)
    if script_args.dataset_name == "FanqingM/MMIU-Benchmark":
       dataset = prepare_dataset(dataset, script_args.dataset_name, script_args.dataset_train_split)
    
    ## BARC SAMPLE DATASET
    # dataset = prepare_custom_dataset('/home/ubuntu/barc_client_data/Training/gemma_sample_dataset_colors/Chunk_colors/gemma_dataset.json')

    def check_dataloader(trainer):
        dataloader = trainer.get_train_dataloader()
        sampler = dataloader.sampler
        print(f"Rank {accelerator.local_process_index}: Using sampler {type(sampler).__name__}")
        if isinstance(sampler, DistributedSampler):
            indices = list(sampler)
            print(f"Rank {accelerator.local_process_index}: Dataset size: {len(sampler.dataset)}, Samples per replica: {sampler.num_samples}, Replicas: {sampler.num_replicas}")
            print(f"Rank {accelerator.local_process_index}: Assigned sample indices: {indices}")
            # Log sample details for this rank
            for idx in indices:
                sample = sampler.dataset[idx]
                sample_id = sample.get('id', idx)
                question = sample['messages'][1]['content'][-1]['text'] if 'messages' in sample else 'Unknown'
                print(f"Rank {accelerator.local_process_index}: Sample index {idx}, ID: {sample_id}, Question: {question}")
    ################
    # Training
    ################
    
    print("Starting training...")
    model_args.lora_r = 2 # Reduce rank from 8 to 4
    model_args.lora_target_modules = ["q_proj", "v_proj", "k_proj", "o_proj"]
    peft_config = LoraConfig(
        r=model_args.lora_r,
        target_modules=model_args.lora_target_modules,
        lora_alpha=8,
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM"
    )


    trainer = SFTTrainer(
        model=model,
        args=training_args,
        data_collator=collate_fn,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
        processing_class=processor.tokenizer,
        peft_config=peft_config,
        eval_accumulation_steps=1
    )
    #check_dataloader(trainer)
    print("Training started...")
    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)


if __name__ == "__main__":
    main()

Is this Internal SFTTrainer or accelerate issue ?

System Info

[2025-05-22 14:44:09,766] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect)

Copy-paste the following information when reporting an issue:

• Platform: Linux-6.8.0-1021-aws-x86_64-with-glibc2.35
• Python version: 3.12.8
• TRL version: 0.17.0
• PyTorch version: 2.7.0
• CUDA device(s): NVIDIA H200, NVIDIA H200, NVIDIA H200, NVIDIA H200, NVIDIA H200, NVIDIA H200, NVIDIA H200, NVIDIA H200
• Transformers version: 4.52.2
• Accelerate version: 1.7.0
• Accelerate config:
  • compute_environment: LOCAL_MACHINE
  • distributed_type: DEEPSPEED
  • mixed_precision: bf16
  • use_cpu: False
  • debug: True
  • num_processes: 8
  • machine_rank: 0
  • num_machines: 1
  • rdzv_backend: static
  • same_network: True
  • main_training_function: main
  • enable_cpu_affinity: False
  • deepspeed_config: {'deepspeed_moe_layer_cls_names': '', 'gradient_accumulation_steps': 1, 'gradient_clipping': 8.0, 'offload_optimizer_device': 'cpu', 'offload_param_device': 'none', 'zero3_init_flag': True, 'zero3_save_16bit_model': True, 'zero_stage': 3}
  • downcast_bf16: no
  • tpu_use_cluster: False
  • tpu_use_sudo: False
  • tpu_env: []
  • dynamo_config: {'dynamo_backend': 'INDUCTOR'}
• Datasets version: 3.6.0
• HF Hub version: 0.31.4
• bitsandbytes version: 0.45.5
• DeepSpeed version: 0.16.8
• Diffusers version: not installed
• Liger-Kernel version: not installed
• LLM-Blender version: not installed
• OpenAI version: not installed
• PEFT version: 0.15.2
• vLLM version: not installed

Checklist

• I have checked that my issue isn't already filed (see open issues)
• I have included my system information
• Any code provided is minimal, complete, and reproducible (more on MREs)
• Any code provided is properly formatted in code blocks, (no screenshot, more on code blocks)
• Any traceback provided is complete