Jee Jee Li 8e92248f79 [Kernel] Retire the DSv3 router GEMM CUDA kernel (#54040) ...

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

2026-08-31 07:15:31 +00:00

..

configs

[Kernel][Perf] Tune fused_moe FP8 config for Qwen3.5 on L40S (+7%) (#53819)

2026-08-26 04:02:26 +00:00

experts

[Model] Support Qwen3.8-Flash-Next (#53896)

2026-08-31 05:57:56 +00:00

oracle

[Model] Support Qwen3.8-Flash-Next (#53896)

2026-08-31 05:57:56 +00:00

prepare_finalize

[DeepEPv2] Support MXFp8 Activation Scale Dispatch (#51398)

2026-08-26 12:26:11 +00:00

router

[Kernel] Retire the DSv3 router GEMM CUDA kernel (#54040)

2026-08-31 07:15:31 +00:00

runner

[Perf][ROCm] Dual-stream decode with hipgraphs (#52033)

2026-08-30 08:24:20 +00:00

__init__.py

[MoE] Share apply_moe_activation support metadata (#44359)

2026-08-05 12:05:47 -07:00

activation.py

[Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE (#51332)

2026-08-25 10:21:00 -04:00

all2all_utils.py

[MoE] Refine FlashInfer one-sided All2All integration (#51924)

2026-08-18 00:18:14 +00:00

b12x.py

[Kernel] Add b12x FP4 MoE backend (#52018)

2026-08-21 08:04:36 -07:00

config.py

[K3 Perf] Fuse MXFP4 top-k finalization into latent-tail, ~5% E2E latency reduction (#53152)

2026-08-21 13:29:59 -04:00

deep_gemm_utils.py

[Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE (#51332)

2026-08-25 10:21:00 -04:00

eep_reconfigure.py

[MoE Refactor] FusedMoE/MoERunner inversion refactor (#41184)

2026-06-08 10:42:58 -04:00

expert_map_manager.py

[MoE Refactor] Move expert map related code into ExpertMapManager class (#41046)

2026-05-12 09:18:27 -04:00

fused_flydsl_moe.py

[ROCm] Pass pointers to FlyDSL MoE kernels (#50378)

2026-07-30 12:49:26 -07:00

fused_moe_method_base.py

[K3 Perf] Fuse MXFP4 top-k finalization into latent-tail, ~5% E2E latency reduction (#53152)

2026-08-21 13:29:59 -04:00

fused_moe_modular_method.py

[MoE Refactor] Remove MoE legacy code (#51078)

2026-08-05 14:35:36 -07:00

fused_moe.py

[ROCm][CI] Extended Fused MoE and FP8 MoE test support (#41100)

2026-08-19 10:19:42 -05:00

hpc_moe.py

[HPC Attention Backend] hpc attention backend support bf16 kv cache with fp8 weight (#50980)

2026-08-06 15:40:14 +08:00

layer.py

[MOE] Standardize and abstract fused shared expert optimization selection (#51695)

2026-08-18 12:56:56 -04:00

modular_kernel.py

[Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE (#51332)

2026-08-25 10:21:00 -04:00

moe_align_block_size.py

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

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

moe_fused_mul_sum.py

[Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE (#51332)

2026-08-25 10:21:00 -04:00

moe_output.py

[K3 Perf] Fuse MXFP4 top-k finalization into latent-tail, ~5% E2E latency reduction (#53152)

2026-08-21 13:29:59 -04:00

moe_permute_unpermute.py

[Bug] Fix torch device issue for MOE permute (#44005)

2026-05-29 18:55:00 +00:00

routed_experts_capturer.py

[Bugfix][R3] Unwrap UniformTypeKVCacheSpecs when selecting the routed-experts KV group (#53240)

2026-08-21 19:13:53 -07:00

routed_experts.py

[Model] Support Qwen3.8-Flash-Next (#53896)

2026-08-31 05:57:56 +00:00

topk_weight_and_reduce.py

[Kernel] TD operand loads for batched MoE GEMM (moe_mmk) on XPU (#46340)

2026-07-26 08:50:07 +08:00

unquantized_fused_moe_method.py

[Kimi K3 Refactor] Add UnfinalizedMoEOutput proto following up for #53152 (#53310)

2026-08-24 11:21:11 -04:00

utils.py

[Perf][DSv4] Use native CUDA SwiGLU clamp kernel for Humming MoE (throughput +1.40%) (#53685)

2026-08-27 13:52:45 -07:00