mirror of
https://github.com/kvcache-ai/Mooncake.git
synced 2026-11-05 01:17:55 +00:00
[GH-ISSUE #602] [Usage]: How can i use tcp protocol in sglang RTX4090 #201
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#201
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 @ZhenshengWu on GitHub (Jul 8, 2025).
Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/602
Describe your usage question
8 * 4090
sglang-0.4.8-post1
mooncake-transfer-engine 0.3.4.post2
prefill
decode
error_info
MoonCake test Case has same error
target:
initiator:
Mooncake test case would be ok if set use_vram to false
Before submitting a new issue...
@alogfans commented on GitHub (Jul 10, 2025):
Currently TCP transport requires that the memory region should be DRAM only. We will fix soon to support VRAM buffer by adding extra copies.
Note: We strongly recommend to use PD disaggregation under the RDMA network for best performance.
@ZhenshengWu commented on GitHub (Jul 10, 2025):
Thank you very much, and I look forward to your excellent work!
@YangZeyu95 commented on GitHub (Jul 11, 2025):
Hi @alogfans, thx for your work. And I have a question, when I run P/D with sglang in a single node. GPU 0,1,2,3 for prefill and 4,5,6,7 for decode. Why still need a RDMA device to communication because it's on one node.
@alogfans commented on GitHub (Jul 14, 2025):
In the current implementation, we require RDMA devices to communicate. However, in the next major revision (expected to be released in Q3), we will allow to use NVLink instead.
@CSEEduanyu commented on GitHub (Jul 22, 2025):
Why use RDMA without crossing machines? Isn't NVLink better?