[GH-ISSUE #484] [Green Context MoE] Introduce GC MoE #202

Open
opened 2026-10-26 21:14:41 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @Bruce-Lee-LY on GitHub (Oct 26, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/484

GC MoE is a CuTe DSL implementation of FP8/FP4 mixture-of-experts inference with CUDA Green Contexts. It overlaps expert-parallel token dispatch with Tensor Core computation, then combines the expert outputs through a three-role CUDA Graph. The operator supports routed experts and optional fused shared experts (SE), with FP8 activations, packed FP4 routed weights, FP8 shared weights and BF16 outputs.

In the decode scenario, the performance of gc moe is approximately 9% faster than that of mega moe on average.

gc-moe

Originally created by @Bruce-Lee-LY on GitHub (Oct 26, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/484 GC MoE is a CuTe DSL implementation of FP8/FP4 mixture-of-experts inference with CUDA Green Contexts. It overlaps expert-parallel token dispatch with Tensor Core computation, then combines the expert outputs through a three-role CUDA Graph. The operator supports routed experts and optional fused shared experts (SE), with FP8 activations, packed FP4 routed weights, FP8 shared weights and BF16 outputs. In the decode scenario, the performance of gc moe is approximately 9% faster than that of mega moe on average. [gc-moe](https://github.com/Bruce-Lee-LY/gc-moe)
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#202
No description provided.