Skip to content

perf: add MAGI-2 Preview S5000 fused MoE autotune configs - #109

Open
yeahdongcn wants to merge 2 commits into
mainfrom
codex/magi2-s5000-moe-config
Open

perf: add MAGI-2 Preview S5000 fused MoE autotune configs#109
yeahdongcn wants to merge 2 commits into
mainfrom
codex/magi2-s5000-moe-config

Conversation

@yeahdongcn

Copy link
Copy Markdown
Collaborator

Summary

Add shape-specific Triton fused-MoE autotune configs for MAGI-2 Preview on MTT S5000.

  • Local flattened experts: E=768 (EP4)
  • Expert intermediate width: N=1280
  • Tuned M buckets: M=21996, 24468, 45012
  • W13/up: BLOCK_M=128, BLOCK_N=128, BLOCK_K=32, GROUP_M=16, num_warps=16, num_stages=1
  • W2/down: K=32 for M=21996/24468; K=64 for M=45012

These configs are intended for the torchada Triton fused-MoE autotune loader. The vLLM-Omni MAGI-2 custom SGL path currently keeps its independently validated hand-tuned W13 K32/W2 K64 split.

Validation

The files were generated and reviewed against the MAGI-2 EP4 shape. Runtime integration in vLLM-Omni was tested separately; no claim is made that the custom Omni path consumes these files automatically.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant