[None][feat] Share Expert Fusion with cherry-pick #11143#15297

Open

leslie-fang25 wants to merge 19 commits into

NVIDIA:mainfrom

leslie-fang25:cherry-pick/pr11143-shared-expert-fusion

Commits on Jul 16, 2026

[None][feat] Cherry-pick #11143 to Main
leslie-fang25
committed
Add E2E testing and fix
leslie-fang25
committed
[None][feat] Enhance the check to enable this fusion path
leslie-fang25
committed
[None][test] bench_moe: shared-expert fusion support
leslie-fang25
committed
[None][test] bench_moe: route unfused shared GatedMLP FP8 GEMM via cute_dsl to avoid IMA
leslie-fang25
committed
[None][fix] Restrict fused shared-expert MoE to tileN>=32 to avoid small-tile dynB IMA
leslie-fang25
committed
[None][chore] Fix yapf formatting
leslie-fang25
committed
[None][feat] Make TRTLLM-Gen shared-expert fusion opt-in via TLLM_MOE_ENABLE_SHARED_EXPERT_FUSION
leslie-fang25
committed
[None][test] Add TRTLLM-Gen shared-expert fusion coverage to modules/moe backend tests
leslie-fang25
committed
[None][test] Drop fusion additions to deprecated thop test_moe.py (coverage moved to modules/moe)
leslie-fang25
committed
[None][test] Drop temporary in-tree fusion E2E test (kept as standalone local validation script)
leslie-fang25
committed
[None][chore] Drop unconsumed EP token-sharding plumbing for fused shared experts
leslie-fang25
committed
[None][fix] bench_moe: drop stale cuda-graph max_num_tokens special-case
leslie-fang25
committed
[None][chore] bench_moe: print traceback on build failures
leslie-fang25
committed