deepspeedai / deepspeedai/DeepSpeed
[REQUEST] Enable both CPU and NVMe for optimizer
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 43.1k
- Forks
- 5k
- Avg merge
- 4d 15h
- Merged PRs (30d)
- 112
Description
Describe the solution you'd like
Dear developers, I am trying to test the optimizer offloading function with deepspeed. I notice that when nvme offloading is enabled, all optimizers will go to NVMe, which misses the chances to put some in CPU. Although there is a configurable variable "ratio" in "offload_optimizer", it will result in some optimizers on GPU and some on CPU, leading an error when specifying "device" as "nvme".
An example expectation is:
0.05 optimizer states on GPU
0.55 optimizer states on CPU
0.4 optimizer states on NVMe
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 by tracing the offload_optimizer configuration and the optimizer offloading path in DeepSpeed. Compare the current handling of the "ratio" and "device" settings with the requested allocation across GPU, CPU, and NVMe. Done means a supported configuration can place optimizer states on all three devices with the stated proportions without an error.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 25/100