Online Dynamic Batching for Accelerate

Train and evaluate a public multimodal fine-tuning example with Online Dynamic Batching and Hugging Face Accelerate.

This repository is a runnable integration example. It is not a reproduction package for the paper's experimental numbers; throughput and quality metrics can differ with hardware, storage, dataset composition, model checkpoints, and software versions.

Prerequisites

• A Python environment with PyTorch and NVIDIA GPU support.
• A local or Hugging Face-accessible Qwen3-VL-2B-Instruct checkpoint, provided through ODB_MM_MIX_MODEL when you do not want to use the default model id.
• Network access to GitHub and the public data/model sources, or equivalent local mirrors.
• Enough disk space for the generated public TMDB data, checkpoints, validation outputs, and MMMU-MC benchmark outputs.

Run ODB

Use a Python environment with PyTorch/GPU support, then run:

export ODB_MM_MIX_MODEL=/path/to/Qwen3-VL-2B-Instruct
./run.sh all-odb

This installs the example dependencies, builds the public data, trains the Accelerate ODB path, and runs validation loss plus MMMU-MC evaluation.

By default training uses one process. Set ODB_MM_MIX_NUM_PROCESSES=8 for an 8-GPU Accelerate run. The default training run uses a small public subset so the example finishes quickly; set ODB_MM_MIX_TRAIN_SIZE=0 to use the full public training split.

Tested Workflow

The tested workflow uses online-dynamic-batching>=0.1.2, Qwen3-VL-2B-Instruct, the public MM-Mix TMDB recipe, and the LLaMA-Factory-compatible validation split (val_size=0.05, split_seed=42). It covers:

• ./run.sh all-odb: data build, ODB training, validation loss, and MMMU-MC.
• ./run.sh train-standard plus ./run.sh eval-standard: fixed-batch baseline training and evaluation.

The records under results/ are example run records. They are useful for checking that the example behaves sensibly, but they should not be read as paper-number reproduction results.

The final full-epoch ODB path was also checked for clean process teardown after training and checkpoint save. For stable benchmark runs, pre-cache MMMU locally and run evaluation with HF_DATASETS_OFFLINE=1, HF_HUB_OFFLINE=1, and TRANSFORMERS_OFFLINE=1.

Run Step By Step

# Install ODB and the helper dependencies for this example.
./run.sh install

# Download/build the public multimodal TMDB training data.
./run.sh data

# Train the ODB run and save the final checkpoint for evaluation.
ODB_MM_MIX_SAVE_FINAL_MODEL=1 ./run.sh train-odb

# Compute validation loss and MMMU-MC for the ODB checkpoint.
./run.sh eval-odb

The default paths are:

• Public data: data/mm-mix-tmdb
• Dataset builder checkout: .deps/build-mm-mix-dataset
• Checkpoints and eval outputs: outputs/accelerate-real

Run Standard

After ./run.sh install and ./run.sh data, run the fixed-batch baseline:

ODB_MM_MIX_SAVE_FINAL_MODEL=1 ./run.sh train-standard
./run.sh eval-standard

Common Options

# Use 8 GPUs through Accelerate.
ODB_MM_MIX_NUM_PROCESSES=8 ./run.sh train-odb

# Pick a different launcher port when running multiple jobs on one machine.
ODB_MM_MIX_MASTER_PORT=29682 ./run.sh train-odb

# Use the full public training split.
ODB_MM_MIX_TRAIN_SIZE=0 ./run.sh train-odb

# Save a checkpoint for validation loss and benchmark evaluation.
ODB_MM_MIX_SAVE_FINAL_MODEL=1 ./run.sh train-odb

# Tune the image cap for larger or smaller visual inputs.
ODB_MM_MIX_IMAGE_MAX_PIXELS=589824 ./run.sh train-odb

# Evaluate a custom checkpoint.
ODB_ACCEL_EVAL_CHECKPOINT=/path/to/checkpoint ./run.sh eval-valloss
ODB_ACCEL_EVAL_CHECKPOINT=/path/to/checkpoint ./run.sh benchmark

Outputs

Default model directories:

Target	Directory
ODB	outputs/accelerate-real/odb
Standard	outputs/accelerate-real/standard

Validation-loss outputs are written under the evaluated checkpoint directory as eval_out_accelerate_valloss.

MMMU-MC outputs are written under the evaluated checkpoint directory as mmmu_mc_likelihood_accelerate and include:

• mmmu_mc_likelihood_results.json
• predictions.jsonl
• excluded.jsonl
• score_audit.json

Commands

Command	Purpose
./run.sh install	Install Python dependencies for this example.
./run.sh data	Build the public TMDB data.
./run.sh train-odb	Train with ODB.
./run.sh eval-odb	Evaluate the ODB checkpoint.
./run.sh train-standard	Train the fixed-batch baseline.
./run.sh eval-standard	Evaluate the Standard checkpoint.
./run.sh eval-valloss	Evaluate validation loss for a saved checkpoint.
./run.sh benchmark	Run the built-in MMMU-MC benchmark.
./run.sh all-odb	Run the complete ODB path.

Manual Training Launcher

Use scripts/train_accelerate.py when you need to override individual knobs without editing run.sh:

accelerate launch --num_processes 8 scripts/train_accelerate.py \
  --loader odb \
  --data data/mm-mix-tmdb \
  --model "$ODB_MM_MIX_MODEL" \
  --token-budget 12288 \
  --buffer-size 1024 \
  --loss-scaling exact \
  --join \
  --deepspeed-config configs/ds_z2.json \
  --train-size 128 \
  --max-steps 0

Set --train-size 0 for the full public training split. Use --loader standard for the fixed-batch baseline.

Integration Notes

This example uses a native Accelerate training loop. The data path is:

public MM-Mix TMDB
  -> lazy Dataset returns one processed tensor sample per __getitem__
  -> ODB groups real post-processor lengths
  -> tensor-only collator pads/stacks each ODB group
  -> Accelerate training loop runs model.forward / loss / optimizer

For a model-specific integration, keep the same boundary: processor output must be a single-sample tensor dictionary before ODB grouping.

For the package API contract behind this example, see the Accelerate integration guide.

The default image cap is chosen for stable out-of-the-box execution. Tune ODB_MM_MIX_IMAGE_MAX_PIXELS when you want to allow larger images.

Related Examples

• Shared dataset builder: build-mm-mix-dataset
• LLaMA-Factory example: odb-example-llamafactory
• HF Trainer example: odb-example-hf-trainer
• Lightning example: odb-example-lightning