deepspeedai / deepspeedai/DeepSpeed
cpu memory out of use when infering on 30b model
Open
Nobody has claimed this yet.
bug
inference
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
The memory of my machine is 500GB. When I want to run opt-30b from huggingface on 4 cards, in the initialization of DSPipeline, the overall cpu memory runs out of use(each subprocess consumes 130GB for model loading).
Is there a way to reduce the overall cpu memory use of all the subprocesses before tensor slicing?
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 at DSPipeline initialization and trace how each subprocess loads the opt-30b model before tensor slicing. Reproduce the four-card Hugging Face inference setup, measure per-process CPU memory, and determine whether the overall memory usage can be reduced without exhausting the 500GB machine.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100