deepspeedai / deepspeedai/DeepSpeed
ZERO Stage 3 with NVME Offload CPU OOM
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
When using ZERO Stage 3 with NVME Offload to train a 40B parameter model, I see CPU OOM errors during the partitioning of optimizer states, which causes a crash before training can begin.
Reproduction is here: https://gist.github.com/stephenrawls/ed4d2e57d4d365ff4e5ec41fbedbb6ed
Details:
The CPU OOM error appears to happen while inside of _swap_out_unpinned_tensors(), which is called from swap_out_optimizer_state() function here: https://github.com/microsoft/DeepSpeed/blob/0d4a54a04d658db40a120bc10c6f1f1a4478f6f1/deepspeed/runtime/swap_tensor/partitioned_optimizer_swapper.py#L120-L164
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the linked reproduction gist and inspect deepspeed/runtime/swap_tensor/partitioned_optimizer_swapper.py, especially _swap_out_unpinned_tensors() as called by swap_out_optimizer_state(). Reproduce the ZERO Stage 3 NVME Offload run and investigate the CPU OOM during optimizer-state partitioning; done means the run no longer crashes before training begins.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100