Issues

• sm90 kernel can deadlock when PV blocks are skipped (pvthreshd)

  thu-ml/SpargeAttn#127 · Filter by author HanzhiZhang-Ulrica opened on Oct 8, 2026

• Masked tail loads leave undefined values in sparse block pooling

  thu-ml/SpargeAttn#124 · Filter by author arcusbuilds opened on Sep 22, 2026

• Build failure on modern toolchains (GCC 13+ / CUDA 13) due to redefinition errors from forced -include,cassert

  thu-ml/SpargeAttn#122 · Filter by author saccie opened on Aug 26, 2026

• Code release of SpargeAttention2

  thu-ml/SpargeAttn#119 · Filter by author billbliss3 opened on Jun 11, 2026

• Is multiple archs supported?

  thu-ml/SpargeAttn#118 · Filter by author SamTyurenkov opened on May 20, 2026

• Cpu stuck when running SpargeAttn

  thu-ml/SpargeAttn#116 · Filter by author thanhhung0112 opened on Mar 10, 2026

• Build failure with CUDA 13.0

  thu-ml/SpargeAttn#115 · Filter by author gzsiang opened on Mar 9, 2026

• Compiling on Python 3.14 fails due to redefinition errors

  thu-ml/SpargeAttn#114 · Filter by author blepping opened on Jan 18, 2026
  1

• does spas_sage2_attn_meansim_topk_cuda support Deepseek V3.1?

  thu-ml/SpargeAttn#113 · Filter by author staugust opened on Jan 15, 2026

• Random black output when using spas_sage2_attn_meansim_topk_cuda

  thu-ml/SpargeAttn#110 · Filter by author xuanmingShang opened on Jan 14, 2026
  2

• 5090 black is not support？

  thu-ml/SpargeAttn#109 · Filter by author secret-deus opened on Jan 12, 2026
  1

• First, I created a sparse attention module for using Flux 1.D which I tested. Some thoughts, questions...

  thu-ml/SpargeAttn#107 · Filter by author ukaprch opened on Dec 29, 2025

• Using the same selection algorithm, but employing non-quantized Flashinfer for computation, why are the generated results all garbled?

  thu-ml/SpargeAttn#106 · Filter by author hjdjs12 opened on Dec 27, 2025

• Possible logical issues in some code within the setup.py file.

  thu-ml/SpargeAttn#104 · Filter by author junyii37 opened on Dec 25, 2025

• windows 下 源码安装的时候 执行link 报错 如何解决？

  thu-ml/SpargeAttn#103 · Filter by author tailuchen opened on Dec 24, 2025
  7

• Does SparseAttn support NVIDIA GB10 architecture?

  thu-ml/SpargeAttn#102 · Filter by author GarfieldHuang opened on Dec 22, 2025

• Please use spas_sage2_attn_meansim_topk_cuda and block_sparse_sage2_attn_cuda APIs

  enhancement
  Feature

  thu-ml/SpargeAttn#101 · Filter by author jt-zhang opened on Dec 16, 2025

• Question about TopK

  thu-ml/SpargeAttn#99 · Filter by author jhl13 opened on Nov 24, 2025
  1

• Calling SpargeAttn Twice Corrupts CUDA Context (unspecified launch failure)

  thu-ml/SpargeAttn#97 · Filter by author sajalmaheshwari624 opened on Nov 19, 2025
  2

• Release of FLUX model dict

  thu-ml/SpargeAttn#96 · Filter by author SeanAlpaca818 opened on Nov 13, 2025

• [BUG] SPARSE ATTRS OF TUNED-MODELS WILL BE LOADED INCORRECTLY!

  thu-ml/SpargeAttn#95 · Filter by author kissint8 opened on Nov 13, 2025

• Different Block Sizes for Query and Key

  thu-ml/SpargeAttn#93 · Filter by author mikeallen39 opened on Nov 4, 2025

• Update: Using the spas_sage2_attn_meansim_topk_cuda API

  enhancement
  Feature

  thu-ml/SpargeAttn#92 · Filter by author jt-zhang opened on Nov 2, 2025
  1

• Test code on the dataset

  thu-ml/SpargeAttn#91 · Filter by author MuHeDing opened on Oct 30, 2025
  2

• kernel do not support NHD

  thu-ml/SpargeAttn#88 · Filter by author DamonsJ opened on Oct 16, 2025
  2