Suggestion for Fine-Grained Batch Control e.g per_device_train_batch_size or mini_batch_size

[dechunwang]

Hello there,
First, I'd like to express my appreciation for your excellent work on this project.
While experimenting with PPO/RW using this repository, I consistently encounter Out of Memory (OOM) errors with the following configuration:

• Model: LLaMA 7B
• Tensor Parallelism (TP): 1
• Pipeline Parallelism (PP): 8
• Global Batch Size (GBS): 256
• Sequence Length: 8192

This error is unexpected given the relatively small model size and parallelism configuration.

The project currently offers an n_mbs setting, which splits data batches into n_mbs chunks. However, this approach has limitations:

1. Difficulty in determining the exact batch size after data packing
2. To set n_mbs correctly without knowing globale batch size is very hard

After reviewing the documentation, I believe a crucial feature is missing:
per_device_train_batch_size or mini_batch_size

A direct mini_batch_size setting would provide more intuitive and precise control over batch sizes across different parallelism configurations.

This setting would allow users to specify the mini-batch size for each Data Parallel (DP) rank, providing several benefits:

1. Fine-grained control in MP/PP/TP scenarios (e.g., virtual parallel)
2. Better resource utilization (e.g., a 7B parameter model shouldn't require PP > 2 if mini_batch_size = 1)

Would it be possible to consider adding this feature in a future update?

[garrett4wade]

Hi dechun, sorry for the late reply.

Your suggestion is quite reasonable. I will raise a PR later. You can also make the change if you would like to contribute to this project.

I'd like to make some comments more about your scenario. Some tips here:

• Set "ppo.gen.force_no_logits_mask=True", otherwise the mask will occupy a huge amount of GPU memory but turns out to have minor learning performance improvement.

• Use the identical parallel strategy if your resource budget is tight (i.e., allocation_mode="d1m8p1", I believe TP is more memory-efficient than PP though). This will disable parameter reallocation but still enable offloading.

• For now, calculate n_mbs based on per_device_train_batch_size=1. For PPO training, the global batch size is per_device_train_batch_size * (pp_size * 2) * dp_size * ppo_n_mbs * n_mbs. For generation or inference, it is per_device_train_batch_size * pp_size * dp_size * n_mbs. Though inconvenient, I think setting per_device_train_batch_size=1 implicitly will fix the OOM issue.

If OOM still happens, it would be helpful if you can share the log and CLI configuration. Discussions are welcomed.