Xin Yang 53d7f1f601

[Kernel] Use pre-allocated output buffer for triton kernel fused_experts (#29219) ...

Signed-off-by: Xin Yang <xyangx@amazon.com>

2025-11-26 10:21:00 +08:00

..

configs

chore: add RTX_PRO_6000 GLM4.6-FP8 kernel tuning (#29240)

2025-11-22 08:42:48 -08:00

__init__.py

[Kernels] Split up fused_moe/layer.py, isolate more modular kernel code (#28064)

2025-11-11 07:29:02 -07:00

all2all_utils.py

[Feat][Perf] Enable deepep-low-latency with round-robin expert placement. (#28449)

2025-11-19 13:46:24 +01:00

batched_deep_gemm_moe.py

[Performance][DeepGEMM] Estimate expected_m (#28694)

2025-11-15 13:52:14 +08:00

batched_triton_or_deep_gemm_moe.py

…

config.py

Handle triton kernel import exception (#29062)

2025-11-22 10:07:50 +00:00

cpu_fused_moe.py

[fix][cpu] Use a SwigluOAI impl which supports interleaved gate-up wei (#29273)

2025-11-25 15:11:11 +08:00

cutlass_moe.py

…

deep_gemm_moe.py

[BugFix] Fix DeepGEMM over-allocating workspace (#28254)

2025-11-10 17:01:17 -05:00

deep_gemm_utils.py

…

deepep_ht_prepare_finalize.py

[Bug] Fix DBO IMA issue for DeepEPHT (#27666)

2025-10-29 16:28:27 -04:00

deepep_ll_prepare_finalize.py

[MoE] Nvfp4 Masked Gemm: Add flashinfer grouped_gemm_nt_masked (#25990)

2025-11-19 13:29:06 -08:00

flashinfer_cutedsl_moe.py

[MoE] Nvfp4 Masked Gemm: Add flashinfer grouped_gemm_nt_masked (#25990)

2025-11-19 13:29:06 -08:00

flashinfer_cutlass_moe.py

Feature: Support Relu2 in FusedMoE fp8 cutlass path (#27261)

2025-11-16 13:39:44 -05:00

flashinfer_cutlass_prepare_finalize.py

[Kernels] Enable FlashInfer FP8 Blockscale on SM90 (for TEP DSR1) (#27134)

2025-11-14 08:02:44 -08:00

flashinfer_trtllm_moe.py

[Performance] Support FP8 flashinfer TRTLLM MOE on Qwen3 and Qwen-3next (#27492)

2025-11-10 12:34:57 -05:00

fused_batched_moe.py

…

fused_marlin_moe.py

Support DeepEP for Kimi-k2-thinking through enabling gemm selection for compressed-tensor marlin wna16 (#28574)

2025-11-12 21:40:45 -08:00

fused_moe_method_base.py

[MoE][Refactor] Make select_experts a non-static method (#29067)

2025-11-24 13:38:04 -05:00

fused_moe_modular_method.py

[MoE][Refactor] Make select_experts a non-static method (#29067)

2025-11-24 13:38:04 -05:00

fused_moe.py

[Log] Optimize startup log (#28948)

2025-11-21 08:46:20 -08:00

gpt_oss_triton_kernels_moe.py

[Kernel] Use pre-allocated output buffer for triton kernel fused_experts (#29219)

2025-11-26 10:21:00 +08:00

layer.py

[Bugfix] Fix GPT-OSS AR+NORM fusion (#28841)

2025-11-25 07:59:40 +00:00

modular_kernel.py

[BugFix] Make sure to allocate worst case MoE workspace during profile run in the DP + EP case (#27426)

2025-11-21 11:41:52 -08:00

moe_align_block_size.py

…

moe_pallas.py

…

moe_permute_unpermute.py

…

moe_torch_iterative.py

…

pplx_prepare_finalize.py

…

prepare_finalize.py

[Feature] Shared Experts Overlap with FI deepgemm swap kernel, 2.2% throughput improvement and 3.6% TTFT improvement (#28879)

2025-11-20 18:41:49 -07:00

rocm_aiter_fused_moe.py

[RFC][ROCm][AITER] Keep all AITER kernels in _aiter_ops class like _custom_ops and _ipex_ops (#24490)

2025-11-10 08:20:53 -08:00

routing_simulator.py

[Bugfix] Fix MoE Routing Simulation (#28002)

2025-11-03 22:26:49 +00:00

shared_fused_moe.py

[Bugfix] Disable shared expert overlap if Marlin MoE is used (#28410)

2025-11-11 23:16:12 +00:00

topk_weight_and_reduce.py

…

triton_deep_gemm_moe.py

…

trtllm_moe.py

[BugFix][Performance] Restore flashinfer autotuning for all scenarios (#27904)

2025-11-04 15:56:21 +08:00

unquantized_fused_moe_method.py

[MoE][Refactor] Make select_experts a non-static method (#29067)

2025-11-24 13:38:04 -05:00

utils.py

…