Issues · flashinfer-ai/flashinfer · GitHub

Labels Milestones New issue

IMA detected on 25b324d (#3058) Support lse in trtllm paged attn kernels

#3114

· aleozlx opened

on Apr 18, 2026

[Feat] HiSparse sparse attention (DeepSeek V3.2 / GLM 5.1) on Blackwell

#3111

· mferrato opened

on Apr 17, 2026

[Feat] Modular BF16 MoE expert kernel — parity with FP8 / NVFP4 modular path

#3110

· mferrato opened

on Apr 17, 2026

[Feat] NVLink EP dispatch: carry per-token LoRA IDs through all-to-all

#3109

· mferrato opened

on Apr 17, 2026

[Feat] FusedMoE: per-sequence LoRA ID + per-expert LoRA weights fused into the kernel

#3108

· mferrato opened

on Apr 17, 2026

[Tracking] Multi-tenant LoRA for MoE under Expert Parallelism on Blackwell

#3107

· mferrato opened

on Apr 17, 2026

grouped_gemm_nt_masked NaN in scale propagates across experts, also affects finite values depending on problem size

#3103

· elvircrn opened

on Apr 17, 2026

[Bug]test_cute_dsl_fused_moe regression on SM121a (DGX Spark)

#3099

· cindyzxq opened

on Apr 17, 2026

[Feat] flashinfer.segmented_top_k: Segmented TopK Implementation from CCCL

#3096

· kahyunnam opened

on Apr 17, 2026

[Bug]: cuBLAS bmm_fp8 segfaults on B200 during autotuning

#3085

· ZhanqiuHu opened

on Apr 16, 2026

[Bug]tests/attention/test_modular_fmha_prefill.py failed of AttributeError: module 'cutlass.cute.arch' has no attribute 'ProxyKind'

#3071

· cindyzxq opened

on Apr 15, 2026

cutlass_fused_moe FP8 incorrect for hidden_size >= 512 (per-tensor dequant path)

#3068

· kzjeef opened

on Apr 14, 2026