-
Notifications
You must be signed in to change notification settings - Fork 224
EXAMPLES: Modify nixl_ep to use UCX channels #1122
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
|
👋 Hi 2dm! Thank you for contributing to ai-dynamo/nixl. Your PR reviewers will review your contribution then trigger the CI to test your changes. 🚀 |
|
Do we have some perf comparison between multiple workers single channel and single worker multiple channels? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This PR does not reduce UCX backend's num_workers to 1
- @eranrs this is why you probably did not see a huge benefit with it. The PR needs to be fixed & re-tested for control path perf
Kernel assertion for 0 combined tokens: Update internode_ll.cu (ai-dynamo#374) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> num_sms calculation simplification: Fix: avoid floating point exception (ai-dynamo#379) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> Increase max topk to 11 support topk10 in low latency kernel From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> Remove fences, optimized nvlink trasnfer Canonicalize TMA usages (ai-dynamo#410) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> Speed up local copy loop macro: Speed up dispatch send by refining loop unrolling (ai-dynamo#385) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> Topk dtype (not only int64) Make dtype of topk_idx configurable (ai-dynamo#422) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35> Assertion and reinterpret_cast to static cast Fix OOB (ai-dynamo#454) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35>i support hidden-dim 3072 (ai-dynamo#458) From <https://github.com/deepseek-ai/DeepEP/commits/main/?before=92fe2deaec24bc92ebd9de276daa6ca9ed602ed4+35>
|
/ok to test |
@brminich, there was an error processing your request: See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/ |
|
/build |
|
/ok to test 0b866f3 |
|
Merged in #1175 |
What?
This PR introduce updates to nixl-ep example: