Issues

[Feature]: SM12x NVFP4 MoE decode (1-8 tokens) at intermediate 640 is 1.1-1.5x slower than a bandwidth-bound reference kernel

needs-triage

#5561

· SSHdotCodes opened

on Sep 25, 2026

Sparse MLA (SM90): `plan()` allow upper bound seq_lens from device

needs-triage

#5559

· NickLucche opened

on Sep 25, 2026

SM120 sparse-MLA decode (DSv3.2 / GLM-NSA): chunks_per_block follows the call's total token count, so a request's output depends on the other requests in the step

needs-triage

#5553

· Bizuayeu opened

on Sep 25, 2026

Experimental tracking: CuTeDSL W4A8 MegaMoE EP16

needs-triage

#5549

· hzfan opened

on Sep 25, 2026

[Feature]: Port remaining MegaMoE improvements from TensorRT-LLM #17956 into FlashInfer moe_ep

feature request

op: moe

priority: must have (P0)

#5507

· leejnau opened

on Sep 23, 2026

[Bug ]err in compilation code about "Unknown option --compress-mode=size " when using FlashInfer version 0.6.18, which is automatically downloaded by vLLM 0.29.0.

bug

needs-triage

#5479

· griezswift opened

on Sep 23, 2026

SM107: PDL support for TRT-LLM Gen NVFP4 routed MoE

model: kimi-k3

needs-triage

priority: must have (P0)

#5473

· kpham-sgl opened

on Sep 22, 2026

[Feat] MegaMoE fused expert: FC1 LoRA delta (Mn-bias) + post activation snapshot for down-LoRA

needs-triage

#5468

· VertexC opened

on Sep 22, 2026

[Feature] TRTLLM-GEN FMHA package ships int8 (SageAttention) DiT ragged kernels for sm_100 only; missing for sm_103 and sm_107

needs-triage

#5461

· Vinnie6167 opened

on Sep 22, 2026

[Bug] CuTe DSL HCA (DeepSeek V4 sparse MLA) produces non-finite output on SM107: arch dispatch covers only sm_100/sm_103

needs-triage

#5460

· Vinnie6167 opened

on Sep 22, 2026

[Bug] PrimTS attention kernels fail on CuTe DSL 4.8: Task._run_task_body_impl() got an unexpected keyword argument 'context'

needs-triage

#5459

· Vinnie6167 opened

on Sep 22, 2026

Autotuner tuning_buckets override maps runtime M DOWN while SM100 cute-dsl tactics are only valid for M <= tuned bucket: vLLM warmup crash (Invalid MXFP8 split-K tactic)

needs-triage

#5450

· ebfio opened

on Sep 22, 2026

Navigation Menu

All issues

[Feature]: SM12x NVFP4 MoE decode (1-8 tokens) at intermediate 640 is 1.1-1.5x slower than a bandwidth-bound reference kernel

Sparse MLA (SM90): `plan()` allow upper bound seq_lens from device

SM120 sparse-MLA decode (DSv3.2 / GLM-NSA): chunks_per_block follows the call's total token count, so a request's output depends on the other requests in the step

Experimental tracking: CuTeDSL W4A8 MegaMoE EP16

[Feature]: Port remaining MegaMoE improvements from TensorRT-LLM #17956 into FlashInfer moe_ep

[Bug ]err in compilation code about "Unknown option --compress-mode=size " when using FlashInfer version 0.6.18, which is automatically downloaded by vLLM 0.29.0.

SM107: PDL support for TRT-LLM Gen NVFP4 routed MoE

[Feat] MegaMoE fused expert: FC1 LoRA delta (Mn-bias) + post activation snapshot for down-LoRA

[Feature] TRTLLM-GEN FMHA package ships int8 (SageAttention) DiT ragged kernels for sm_100 only; missing for sm_103 and sm_107

[Bug] CuTe DSL HCA (DeepSeek V4 sparse MLA) produces non-finite output on SM107: arch dispatch covers only sm_100/sm_103

[Bug] PrimTS attention kernels fail on CuTe DSL 4.8: Task._run_task_body_impl() got an unexpected keyword argument 'context'

Autotuner tuning_buckets override maps runtime M DOWN while SM100 cute-dsl tactics are only valid for M <= tuned bucket: vLLM warmup crash (Invalid MXFP8 split-K tactic)

Search results