[train_rlhf_llama] When using vllm,Time-consuming optimization is not obvious as expectation.
- Dominant language
- Python
- Stars
- 450
- Forks
- 40
- PR merge metrics
- No merged PRs in 30d
Description
When running the train_rlhf_llama.sh script for LLaMA2-7B, the performance gain from using vLLM as the inference backend is not significant compared to Megatron (approximately a 20% improvement). According to some other reports, the expected improvement should be around 200%. How can it be promoted?
Contributor guide
No contributing guide indexed for this repository
Research direction
Start with train_rlhf_llama.sh and reproduce the LLaMA2-7B comparison between vLLM and Megatron, recording the reported roughly 20% gain. Trace the timing from that entry point to identify the relevant performance bottleneck; done means a measured explanation and improvement toward the expected result, with reproduction details documented.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, shell
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100