mirror of
https://github.com/kvcache-ai/Mooncake.git
synced 2026-11-05 01:17:55 +00:00
[GH-ISSUE #1207] PD mode failure #411
Labels
No labels
ascend
auto-closed
bug
bug
CI/Build
Common
documentation
enhancement
feature request
good first issue
help wanted
high priority
MetaX
nightly-failure
pull-request
PyTorch Backend
question
RFC
Roadmap
stale
Store
Tests
Transfer Engine
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
kvcache-ai/Mooncake#411
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @UralmashFox on GitHub (Dec 12, 2025).
Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/1207
Originally assigned to: @alogfans on GitHub.
I try to follow sglang-integration-v1.md guide. Did everything like in the guide, tried to increase\dircrease size of page, max-total-tokens and chunked-prefill. Also tried to do some minor changes in docker image as far as I use Prefill, Decode and Proxy in different podes. For now the dockerfile is based on sglang:v0.5.1.post2-cu126 and looks like following:
and the logs of prefill are the following:
Despite "The server is fired up", the proxy drops on the requests the following:
KVTransferError(bootstrap_room=6677116944225248479): Could not fetch prefill parallel info from bootstrap_addr: sglang-prefill-svc:30000which else installations should I add? Thank you in advance!
@ykwd commented on GitHub (Dec 15, 2025):
Based on the logs, it appears that the failure occurred while registering memory with the RDMA NIC. A possible cause is a Docker configuration issue that prevents proper access to the RDMA device.
@UralmashFox commented on GitHub (Dec 15, 2025):
I’ve indeed tried quite a few things on my end - mostly through the Helm chart - such as:
I’d really appreciate any guidance on what to check next - whether it’s a host-level RDMA setting, a missing capability, or a known issue with this SGLang version.
I can also do the same attemps I mentioned above and provide you information that you think could be useful, just tell me what exactly
@stmatengss commented on GitHub (Dec 15, 2025):
Could you try to test the connectivity between prefill and decode?
@stmatengss commented on GitHub (Dec 15, 2025):
Make sure GDR is supported. REF: https://github.com/kvcache-ai/Mooncake/blob/main/docs/source/troubleshooting/troubleshooting.md#sglang-common-questions
@UralmashFox commented on GitHub (Dec 16, 2025):
currently (till the end of January) I switched to another environment with less gpu sources and much-much lighter llm. If the results below do not make sence with described logs abode, let me know, we will postpone the discussion to later
if to run from the prefill pod, the result is the following:
@UralmashFox commented on GitHub (Dec 16, 2025):
From the prefill pod I run the following:
No new logs from prefill;
Logs from proxy:
"POST /generate HTTP/1.1" 500 Internal Server ErrorLogs from decode:
How decode was run:
How prefill was run:
I use --disaggregation-bootstrap-port=30000 just because I can not use 8xxx, only 3xxxxx. Actually, I'm not sure how exactly bootstrap-port should be modified - I tried to play with extra flags in params for prefill
Additionlly from decode pod I tried to do the following request:
curl http://sglang-prefill-svc:30003/routeand get Missing inputs for bootstrap server.
Then I tried to extend the request with the following:
The prefill logs:
Based on this I make an assumption that both - prefill and bootstrap reply on decode requests.
@github-actions[bot] commented on GitHub (Jun 10, 2026):
This issue has had no activity for 90 days and will be closed in 7 days if there is no further activity. Please comment or react if it should stay open.
@github-actions[bot] commented on GitHub (Jun 18, 2026):
Closing due to 3 months of inactivity. If this is still relevant, please comment and we can reopen.