trl: Error with Multi-GPU peft Reward Training

There is an issue when you combine all four:

peft quantization
gradient checkpointing
multi-gpu ddp
two gradients on the same parameters (as you have in the loss function for Reward Trainer)

This is reproducible if you correctly enable gradient checkpointing in examples/multi-adapter-rl as shown in PR #479 and then run in a multi-gpu setup

accelerate launch --multi_gpu reward_modeling.py --gradient_checkpointing True

you will receive the error

RuntimeError: Expected to mark a variable ready only once. This error is caused by one of the following reasons: 1) Use of a module parameter outside the forward function. Please make sure model parameters are not shared across multiple concurrent forward-backward passes. or try to use _set_static_graph() as a workaround if this module graph does not change during training loop.2) Reused parameters in multiple reentrant backward passes. For example, if you use multiple checkpoint functions to wrap the same part of your model, it would result in the same set of parameters been used by different reentrant backward passes multiple times, and hence marking a variable ready multiple times. DDP does not support such use cases in default. You can try to use _set_static_graph() as a workaround if your module graph does not change over iterations. Parameter at index 127 has been marked as ready twice. This means that multiple autograd engine hooks have fired for this particular parameter during this iteration. You can set the environment variable TORCH_DISTRIBUTED_DEBUG to either INFO or DETAIL to print parameter names for further debugging.

With TORCH_DISTRIBUTED_DEBUG=DETAIL, we find the affected parameter is a LoRA parameter. It is not related to https://github.com/pytorch/pytorch/issues/60844 because find_unused_parameters is set to False.

This is likely a problem between peft and accelerate/ddp but I’m putting the issue here because it affects RewardTrainer and quantization + multi gpu + gradient checkpointing are a common combination

About this issue

Original URL
State: closed
Created a year ago
Reactions: 4
Comments: 15 (7 by maintainers)

Most upvoted comments

Yes indeed! if you use latest releases from transformers, trl and peft, simply pass gradient_checkpointing_kwargs={"use_reentrant":False} and it should be resolved

younesbelkada on Nov 10, 2023

I tested this and it is resolved by #912, thank you @younesbelkada !

Sorry I didn’t do this myself earlier. I tried doing the use_rentrant trick you mentioned by it didn’t work. I must have missed something in my tests. Thanks for doing it!

mnoukhov on Nov 10, 2023

I am using huggyllama/llama-7b

EDIT: I misunderstood “unused parameters”. As long as the frozen parameters are used in the backward computation, they are fine. So if the model has two peft adapters, then there could be unused parameters, but not if there’s just one.

After looking into this, I am mostly convinced that DDP does not work with gradient checkpointing when there are unused parameters in the forward computation or two forward passes. This means you should not use gradient checkpointing with peft and DDP, ~regardless of the gradient passes but~ as it will explicitly fail with two forward passes on the same parameters.

This is a note from the Pytorch docs

DistributedDataParallel currently offers limited support for gradient checkpointing with torch.utils.checkpoint(). DDP will work as expected when there are no unused parameters in the model and each layer is checkpointed at most once (make sure you are not passing find_unused_parameters=True to DDP). We currently do not support the case where a layer is checkpointed multiple times, or when there unused parameters in the checkpointed model.

two peft adapters can create unused parameters and two forward passes (as in the reward trainer) will cause a layer to be checkpointed twice.

I think a next step is to check whether DDP training with peft is worse with gradient checkpointing even without two forward passes. If so, peft should probably add a warning or error when a user tries to combine these things until DDP supports the use case. Or maybe switch to DataParallel although it feels like it will be deprecated.

mnoukhov on Oct 13, 2023

I’m getting the same error with accelerate launch --multi_gpu reward_modeling.py --load_in_8bit --gradient_checkpointing True with llama2-7b, however if disabling either gradient_checkpointing or load_in_8bit, I got OOM.

mollynatsu on Sep 13, 2023