Issues

• [RFC] Triton kernels for DeepSeek Sparse Attention (DSA)

  fla-org/flash-linear-attention#1691 · Filter by author flyingdoubleG opened on Nov 7, 2026

• [Bug] CP state kernels (chunk_cp_fwd_h / chunk_cp_bwd_dh) spill heavily at K=256: 50-70x slower than K=128, ~15 min autotune on A100

  fla-org/flash-linear-attention#1683 · Filter by author NuojCheng opened on Nov 7, 2026

• [RFC] Move functional module operators into fla.ops

  fla-org/flash-linear-attention#1672 · Filter by author yzhangcs opened on Nov 6, 2026

• [Bug] Missing dy rearrange in causal_conv1d_bwd when x.ndim > 3

  fla-org/flash-linear-attention#1660 · Filter by author shatianming5 opened on Nov 4, 2026

• chunk_scaled_dot_kkt_fwd: autotune configs with BK < K and num_warps >= 4 give wrong, non-deterministic results on H20 (sm90) + triton 3.6.0

  fla-org/flash-linear-attention#1654 · Filter by author pkusp opened on Nov 3, 2026

• [RFC] Upstream ReWY backward optimizations for Delta-Rule attention

  fla-org/flash-linear-attention#1653 · Filter by author ZhangYi2026 opened on Nov 3, 2026

• [Bug] GKA FP32 with K > 64 doesn't fit in 99 KB shared memory

  bug
  fla-org/flash-linear-attention#1638 · Filter by author andrewliu2001 opened on Oct 31, 2026

• [Bug] MoM synthetic convolution padding corrupts cached short continuations

  bug
  fla-org/flash-linear-attention#1611 · Filter by author Ningke-Li opened on Oct 29, 2026
  1

• [Bug] chunk_gated_delta_rule: wrong bfloat16 gradients in the early layers of Qwen3.5-4B on A100 (fla 0.5.2)

  bug
  fla-org/flash-linear-attention#1598 · Filter by author Peilin-FF opened on Oct 27, 2026
  4

• [Bug] chunk_gated_delta_product: wrong outputs with g=None and num_householder >= 2 (GB200/sm_100)

  bug
  fla-org/flash-linear-attention#1594 · Filter by author YYuJLiu opened on Oct 26, 2026
  1

• [RFC] Set bounded-gate KDA A_log gradient tolerance to 0.004

  fla-org/flash-linear-attention#1586 · Filter by author yzhangcs opened on Oct 25, 2026

• [Feature Request] Add Sparse Delta rule op and Sparse Delta Memory (SDM) layer

  enhancement
  fla-org/flash-linear-attention#1560 · Filter by author lcabannes opened on Oct 23, 2026

• [RFC] Causal additive Laplace attention with feature-wise forgetting

  fla-org/flash-linear-attention#1556 · Filter by author AvivBick opened on Oct 22, 2026
  2

• [RFC] GDN2 intermediate precision for strict BF16 context-parallel accuracy

  fla-org/flash-linear-attention#1550 · Filter by author alfredclwong opened on Oct 21, 2026

• [RFC] AscendC backend for FLA: Scope, roadmap, and performance comparison

  ascend-npuenhancement
  fla-org/flash-linear-attention#1531 · Filter by author chenjingxiang96 opened on Oct 19, 2026
  4

• [CI] nvidia-h100-smoke keeps timing out at 120 minutes as the dependent-test fan-out grows

  fla-org/flash-linear-attention#1522 · Filter by author TecJesh opened on Oct 19, 2026
  1

  

• [RFC] Token-aligned forward for gated DeltaProduct inference (nh >= 2, K <= 64, V <= 128)

  fla-org/flash-linear-attention#1379 · Filter by author gregghy opened on Oct 12, 2026

• [RFC] Add Triadic Gated DeltaNet (Triadic Linear Attention)

  fla-org/flash-linear-attention#1370 · Filter by author ischencheng opened on Oct 11, 2026
  2

• [Feature Request] [RFC] A shared, extensible interface for residual connections

  enhancement
  fla-org/flash-linear-attention#1332 · Filter by author feiyunzhang opened on Oct 6, 2026

• [Perf] Gluon attention backend with large head dimensions

  stale
  fla-org/flash-linear-attention#1322 · Filter by author yzhangcs opened on Oct 3, 2026
  3

• [RFC] Add Graph Capture Support for FLA Prefill and Training Paths on Ascend NPUs

  ascend-npuenhancement
  fla-org/flash-linear-attention#1233 · Filter by author FightingZhen opened on Sep 8, 2026
  1

• [RFC] Graph capture support for prefill and training: GDN/KDA chunk and convolution paths

  enhancement
  fla-org/flash-linear-attention#1155 · Filter by author zhiyuan1i opened on Aug 20, 2026
  4

• [RFC] Roadmap: NPU support for flash-linear-attention

  ascend-npuenhancement
  fla-org/flash-linear-attention#942 · Filter by author zheliuyu opened on Jun 9, 2026
  13

• [Feature Request] Introducing an input buffer for final_state in fused_recurrent_gated_delta_rule_fwd

  enhancement
  fla-org/flash-linear-attention#872 · Filter by author ir2718 opened on Apr 30, 2026
  4

• [Feature Request] Implement tensor inverse

  enhancement
  fla-org/flash-linear-attention#680 · Filter by author sustcsonglin opened on Dec 10, 2025