shangmingc
shangmingc
@DarkLight1337 Hi there, I found all the CI checks have passed. Is it feasible to merge?
@KuntaiDu Hello, this is the initial PR for the XpYd design as we discussed before the Chinese New Year. I think it is ready for review :)
> I just discussed with other folks from vLLM and a common feedback is that it's less ideal to inject the parameters in a lot of files. A better way...
@KuntaiDu The code footprint is minimized now. Also, I moved the implementation of `KVTransferParams` to the _**vllm/distributed/kv_transfer**_ directory. The failed ci is due to huggingface timeout, it should be irrelevant.
@KuntaiDu Any other feedback? We plan to release MooncakeStore this week and implement the first version of XpYd in the second PR.
> Let's wait for next week. @KuntaiDu Is there an agenda and scheduled meeting time?
@KuntaiDu As we discussed last Friday, we refactor this PR to support MooncakeStore-based disaggregated prefill. I think this PR does not need to involve more parties anymore since it does...
> @ShangmingCai Is there a document introduce how to install the dependencies of vllm which using by mooncake-store adapter? If not, it would be better to introduce a guide document...
@maobaolong Seems like it's related to the protobuf lib. You can open an issue in the Mooncake repo, and we will find someone to help you with the build issue.
@maobaolong Maybe try this https://github.com/vllm-project/vllm/pull/14824? @billishyahao will support deepseek once this PR gets merged.