Zijie Yan comments

Results 58 comments of


                                            Zijie Yan

[ENHANCEMENT] Global Batch Load Balancing for MoE Models

Thanks! Let us take a deeper look @Victarry

fix some bug

Merged in https://github.com/NVIDIA/Megatron-LM/commit/afb755f548b48151a4408b0e9caf674b8349b589 Many thanks to the Infrastructure Center of Tencent WeChat's Technical Architecture Department(微信技术架构部-基础架构中心) for their contributions. cc @shifangx

[BUG] Token routing probability all-gather precision in token_dispatcher causes differing results between EP ranks

Thanks for reporting the issue! I suspect the discrepancy is due to the different accumulation orders of reduction during token combination. We've received feedback from other customers suggesting that reduction...

[BUG] Token routing probability all-gather precision in token_dispatcher causes differing results between EP ranks

> Hi [@yanring](https://github.com/yanring) , after manually setting the probs to fp64, I still have precision issues with EP. Do you have any suggestion on what else need to be promoted...

[Dev] feat: m4 leftover changes

LGTM, left a few comments. Will trigger merge after resolving them.

[BUG] Unnecessary initialization for router in megatron-core

Thanks for reporting and fixing it. Unfortunately, we can't directly merge this PR on Github, but we'll include the fix in the next version.

docs(megatron-fsdp): add Megatron-FSDP user guide

+ @sbhavani

[Feature] Dev add gptoss example

@yiakwy-xpu-ml-framework-team Hey, thanks for the PR. Since this is a follow-up PR, I'll temporarily make it a draft.

[Dev] Optimize TE cudagraph input memory

This PR introduces significant logic changes but no corresponding unit tests. Please consider add tests for _get_sample_arguments()

[Dev] feat(MoE): Refactor cuda_graph_scope - part2

/ok to test efb8f5dc4fe7ee0d7aec6d8a546fb82182430489