Repository navigation
[BugFix][PD]: make example proxy usable with P2pNcclConnector - #26628
Conversation
Signed-off-by: PAN <1162953505@qq.com>
Signed-off-by: PAN <1162953505@qq.com>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run You ask your reviewers to trigger select CI tests on top of Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add If you have any questions, please reach out to us on Slack at https://slack.vllm.ai. 🚀 |
|
Documentation preview: https://vllm--26628.org.readthedocs.build/en/26628/ |
There was a problem hiding this comment.
Code Review
This pull request fixes runtime errors in the disaggregated prefill example when using P2pNcclConnector. The changes correctly format the request_id to include KV routing information, which is essential for the connector. The example proxy server logic is updated to perform the prefill and decode steps, and the example shell script is updated with the necessary configurations.
My review focuses on the proxy server implementation. I've identified a critical issue regarding incorrect and inefficient aiohttp session management that will cause runtime errors and performance degradation. I've provided suggestions to fix this by using a single, shared ClientSession for the application's lifetime. The rest of the changes look good and effectively address the bug described.
chaunceyjiang
left a comment
There was a problem hiding this comment.
examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_proxy_p2p_nccl_xpyd.py
As far as I know, P2pNcclConnector should use this proxy.
https://docs.vllm.ai/en/latest/examples/online_serving/disaggregated_prefill.html But in this example, P2pNcclConnector will use |
|
@chaunceyjiang Hey, my friend. Actually, you just need to run the original script to identify this issue, and running the current script will fix this error. Should we further merge this PR, or do you have other plans? |
|
cc @KuntaiDu |
|
Thanks @youkaichao for reminder. We should probably update the example so that it uses The code change of current PR is a bit large, but if @Abatom is too busy to reply we can merge this PR to fix examples first. |
|
@KuntaiDu I’ll take a look later—thanks! |
|
@KuntaiDu @pandalee99 LGTM! |
|
@Abatom thank you for reviewing! Could you please approve the PR on your side? I can then approve on my side and merge. |
|
@KuntaiDu Approve! |
…roject#26628) Signed-off-by: PAN <1162953505@qq.com> Signed-off-by: Runkai Tao <rt572@physics.rutgers.edu>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
…roject#26628) Signed-off-by: PAN <1162953505@qq.com>
Purpose
IN https://docs.vllm.ai/en/latest/examples/online_serving/disaggregated_prefill.html
such as:
Root cause
Fix
___prefill_addr_<prefill_kv_addr>___decode_addr_<decode_kv_addr>_<uuid>X-Request-Id. This format is compatible with P2pNcclConnector.parse_request_id.max_tokens=1(ormax_completion_tokens=1) and wait for it to finish, with the injected request_id.X-KV-Targetto the decode host:port as a compatibility header.notes
X-Request-Idheader to both prefill and decode requests.X-KV-Targetfor compatibility.Test Plan
examples/online_serving/disaggregated_prefill.sh
Test Result
Essential Elements of an Effective PR Description Checklist
relate :https://docs.vllm.ai/en/latest/examples/online_serving/disaggregated_prefill.html
supported_models.mdandexamplesfor a new model.