OOM when unwrap_model_for_generation

hlnchen commented 1 week ago

System Info

torch==2.4.0 transformers==4.43.4 trl==0.9.6 tokenizers==0.19.1 accelerate==0.32.0 peft==0.12.0 datasets==2.20.0 deepspeed==0.15.0 bitsandbytes==0.43.3 sentencepiece==0.2.0 flash-attn==2.6.3

gcc version 11.4.0 (Ubuntu 11.4.0-1ubuntu1~22.04)

Information

[ ] The official example scripts
[X] My own modified scripts

Tasks

[ ] An officially supported task in the examples folder
[X] My own task or dataset (give details below)

Reproduction

Hi TRL team,

I am hitting OOM errors when fine-tuning a Llama-3.1-70B model on my modified RL trainer. It looks like the error happens on unwrapping the model for generation (I have an on policy algorithm and each training step I will generate some sequences)

My machine has 8 H100 80GB GPUs and I used lora. But it looks like unwrap_model_for_generation will load the entire model into memory and causing OOM. Any suggestions?

[rank7]: Traceback (most recent call last):
[rank7]:   File "/export/scripts/training.py", line 243, in <module>
[rank7]:     trainer.train(resume_from_checkpoint=config.checkpoint_path)
[rank7]:   File "/export/trainer/trainer_simple_rloo.py", line 195, in train
[rank7]:     with torch.no_grad(), unwrap_model_for_generation(model, self.accelerator) as unwrapped_model:
[rank7]:   File "/opt/conda/lib/python3.12/contextlib.py", line 137, in __enter__
[rank7]:     return next(self.gen)
[rank7]:            ^^^^^^^^^^^^^^
[rank7]:   File "/export/venv/lib/python3.12/site-packages/trl/models/utils.py", line 162, in unwrap_model_for_generation
[rank7]:     with deepspeed.zero.GatheredParameters(model.parameters()):
[rank7]:   File "/export/venv/lib/python3.12/site-packages/deepspeed/runtime/zero/partition_parameters.py", line 2224, in __enter__
[rank7]:     self.params[0].all_gather(param_list=self.params)
[rank7]:   File "/export/venv/lib/python3.12/site-packages/deepspeed/runtime/zero/partition_parameters.py", line 1143, in all_gather
[rank7]:     return self._all_gather(param_list, async_op=async_op, hierarchy=hierarchy)
[rank7]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank7]:   File "/export/venv/lib/python3.12/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank7]:     ret_val = func(*args, **kwargs)
[rank7]:               ^^^^^^^^^^^^^^^^^^^^^
[rank7]:   File "/export/venv/lib/python3.12/site-packages/deepspeed/runtime/zero/partition_parameters.py", line 1511, in _all_gather
[rank7]:     self._allgather_params_coalesced(all_gather_nonquantize_list, hierarchy, quantize=False)
[rank7]:   File "/export/venv/lib/python3.12/site-packages/deepspeed/runtime/zero/partition_parameters.py", line 1799, in _allgather_params_coalesced
[rank7]:     flat_tensor = torch.empty(tensor_size, dtype=param_list[0].ds_tensor.dtype,
[rank7]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank7]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 448.00 MiB. GPU 7 has a total capacity of 79.11 GiB of which 256.56 MiB is free. Including non-PyTorch memory, this process has 0 bytes memory in use. Of the allocated memory 74.68 GiB is allocated by PyTorch, and 7.22 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Expected behavior

OOM issue resolved.

JiahuiSun commented 4 days ago

I also met the same issue. I use the official example script, dpo_online.py, to train a 75b LLM with a 75b reward model. Even with 60x8 H100 GPUs, the problem still happens. Any help please?

lewtun commented 2 days ago

Hello @hlnchen would you mind sharing a reproducible example that uses the unwrap_model_for_generation() method in a simple training loop that simulates your application?

huggingface / trl