Michael Goin 22a97e6613 [Perf] Improve default triton fused moe configs (#34846) ...

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>

2026-02-23 16:01:28 -08:00

..

configs

[Perf] fused_moe: add int4_w4a16 benchmark support and tuning config (#34130)

2026-02-13 00:14:27 -08:00

oracle

[Model Bash]: Improve FP8 Oracle for Config Specific Kernel Selection (#34260)

2026-02-19 21:29:08 -08:00

router

[Bugfix][MoE Kernel] Fix incorrect routing selection for models without expert groups (e.g., MiniMax-M2.1) (#34673)

2026-02-18 13:03:24 -08:00

runner

[ROCm][Bugfix]: Only save unpadded sizes for shared_experts in MoERunner to fix rmsnorm pad fusion (#34636)

2026-02-20 19:56:16 -08:00

__init__.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

activation.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

all2all_utils.py

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

2026-01-27 01:28:02 +00:00

batched_deep_gemm_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

config.py

[Bugfix][MoE Kernel] Fix incorrect routing selection for models without expert groups (e.g., MiniMax-M2.1) (#34673)

2026-02-18 13:03:24 -08:00

cpu_fused_moe.py

[Bugfix] fix activation in cpu_fused_moe_torch call (#34696)

2026-02-17 23:39:46 -08:00

cutlass_moe.py

[Model] Add NVFP4 quantization support for Step3.5-Flash (#34478)

2026-02-22 12:30:46 -07:00

deep_gemm_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

deep_gemm_utils.py

[Perf] Optimize deepgemm experts initialization, 3.9% TTFT improvement (#30494)

2025-12-11 17:28:34 -05:00

deepep_ht_prepare_finalize.py

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

2026-01-27 01:28:02 +00:00

deepep_ll_prepare_finalize.py

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

2026-01-27 01:28:02 +00:00

fallback.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

flashinfer_a2a_prepare_finalize.py

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

2026-01-27 01:28:02 +00:00

flashinfer_cutedsl_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

flashinfer_cutlass_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

flashinfer_trtllm_moe.py

[UX] More descriptive reasons in is_supported_config for MoE (#34908)

2026-02-19 15:20:52 -08:00

fused_batched_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

fused_marlin_moe.py

[Model] Add NVFP4 quantization support for Step3.5-Flash (#34478)

2026-02-22 12:30:46 -07:00

fused_moe_method_base.py

[MoE Refactor] Introduce MoERunner abstraction and move execution logic from FusedMoE to DefaultMoERunner (#32344)

2026-02-10 19:51:07 -05:00

fused_moe_modular_method.py

[Bugfix] Remove assert that's no longer valid (#34443)

2026-02-12 18:18:15 -08:00

fused_moe.py

[Perf] Improve default triton fused moe configs (#34846)

2026-02-23 16:01:28 -08:00

gpt_oss_triton_kernels_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

layer.py

[Bugfix] Add method to swap quant_method on FusedMoE to fix LoRA issues (#34453)

2026-02-15 20:10:50 -08:00

modular_kernel.py

[UX] More descriptive reasons in is_supported_config for MoE (#34908)

2026-02-19 15:20:52 -08:00

moe_align_block_size.py

[Kernel][MoE] optimize moe_align_block_size (#29642)

2025-12-07 01:58:47 -08:00

moe_permute_unpermute.py

[Refactor] Remove align block size logic in moe_permute (#33449)

2026-02-06 10:57:06 -08:00

mori_prepare_finalize.py

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

2026-01-27 01:28:02 +00:00

pplx_prepare_finalize.py

[Docs] Reduce time spent generating API docs (#34255)

2026-02-11 02:56:02 -08:00

prepare_finalize.py

[Docs] Reduce time spent generating API docs (#34255)

2026-02-11 02:56:02 -08:00

rocm_aiter_fused_moe.py

[Bugfix] Remove assert causing hipErrorStreamCaptureUnsupported (#34455)

2026-02-18 18:54:54 +00:00

routed_experts_capturer.py

[Bugfix] Refactor to support DP parallel in R3 (#32306)

2026-01-16 15:13:58 +08:00

shared_fused_moe.py

[MoE Refactor] Introduce MoERunner abstraction and move execution logic from FusedMoE to DefaultMoERunner (#32344)

2026-02-10 19:51:07 -05:00

topk_weight_and_reduce.py

Update Optional[x] -> x | None and Union[x, y] to x | y (#26633)

2025-10-12 09:51:31 -07:00

triton_cutlass_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

triton_deep_gemm_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

trtllm_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

unquantized_fused_moe_method.py

[MoE Refactor] Introduce MoERunner abstraction and move execution logic from FusedMoE to DefaultMoERunner (#32344)

2026-02-10 19:51:07 -05:00

utils.py

[Refactor] Remove dead private func _fp8_perm and _extract_mask_for_item (#35068)

2026-02-23 05:05:20 -08:00

xpu_fused_moe.py

[Refactor] Replace activation: str with MoEActivation enum (#33843)

2026-02-11 17:29:32 -08:00

zero_expert_fused_moe.py

[MoE Refactor][5/N] Isolate zero expert to LongCatFlash (#28891)

2025-12-20 18:22:04 +00:00