[Kernel] Support Flashinfer trtllm fused MoE non gated FP8 & NVFP4#33506

Merged

vllm-bot merged 15 commits intovllm-project:mainfrom

amitz-nv:support-fi-fused-moe-non-gated-fp8-nvfp4

Feb 12, 2026

Commits on Feb 12, 2026

Add support for FI fused moe non-gated FP8 and NVFP4
amitz-nv
committed
Fix test_flashinfer.py
amitz-nv
committed
Update support filters of FlashInfer for FP8 and FP4 - support non-gated MoE and rel2_no_mul activation, support DeepSeek routing in FP8 per-tensor, fix prepare_static_weights_for_trtllm_fp4_moe
amitz-nv
committed
Fixes after rebase
amitz-nv
committed
Use intermediate size correctly in flashinfer_fp4_moe.py
amitz-nv
committed
Handle silu & gelu (without '_no_mul' suffix) correctly in FI flow
amitz-nv
committed
Make is_gated_activation and activation_str_to_enum_value more robust
amitz-nv
committed
Fix tests/kernels/moe/test_flashinfer.py
amitz-nv
committed
Remove unnecessary scale .max() in test_flashinfer.py
amitz-nv
committed
Import ActivationType directly from FlashInfer, revert changes in vllm/utils/flashinfer.py
amitz-nv
committed
Fix activation type passed (passing int instead of enum class)
amitz-nv
committed
With Flashinfer fused MoE NVFP4 - use min_alignment=128 only for non-gated, otherwise use min_alignment=16
amitz-nv
committed
Remove check of layer.weight_block_size in flashinfer_fp4_moe.py
amitz-nv
committed
Remove default value of activation_type
amitz-nv
committed
Fix formatting
amitz-nv
committed