enable_prefix_caching LMDeploy vs vLLM
Open
- Dominant language
- Python
- Stars
- 8.1k
- Forks
- 748
- Avg merge
- 6d 2h
- Merged PRs (30d)
- 54
Description
I'm currently testing LMDeploy, and it just seems like `enable_prefix_caching` can't be zero-overhead/as performance as vLLM? with vLLM when `enable_prefix_caching` is enabled, the time to generate the same prompts, is almost free.
Is there an example of this that I might be missing for Turbomind engine? Given the same prompts? I know Turbomind Engine doesn't support prefix prompts/system prompts but this also does not work in user prompts?
# vLLM V1
Contributor guide
Assessment
This issue has not been assessed yet.