Repository Issue Activity (beta)

#2811 tests/test_flash_attn.py fails collection without einops

#2307 FA4 SM120 support?

#2581 Support head_dim=512 on SM89 (Ada) for Gemma 4 global attention layers

#2799 Flash Attention install with Triton backend replaces preinstalled Triton no matter what.

#2801 [FA4] max_seqlen_q tensor causes an argument-annotation mismatch in varlen forward

#2800 [ROCm][CK][gfx1201][Windows] FlashAttention backward is 4.4–7.4x slower than PyTorch SDPA on RX 9070 XT

#2797 FA4 backward does not return against cutlass-dsl 4.6.0 and 4.6.1 on B300 (sm_103)

#2438 pack_gqa fails on SM120 (SM80 CuTe DSL path): crd2idx coordinate resolution error

#2793 Varlen version of `flash_attn_with_kvcache`

#2769 publish.yml: create trigger has no tag filter, so branch pushes fail the workflow and new matrix entries never build

#2782 CuTe DSL: replace removed quack.activation.sub_packed_f32x2 API

#2789 ops/triton/linear.py kernel_fwd computes its bounds mask and never passes it to the store

#2783 raise error cuda12.6 pytorch 2.10, vllm0.18.0 when install flash-attn==2.8.1

#2765 [CuTe, SM100] test_clc_fuzz fails after #2559 due to stale scheduler expectations

#2442 [Wheel] Pre-built flash_attn v2.8.3 for CUDA 13 + PyTorch 2.11 (Python 3.12)

#2764 [Cute, SM100] FA4 and FA4_Varlen fwd runtime difference reaches 27%

#2760 [CuTe, SM100] Deadlock in varlen + block-sparse + SplitKV forward: packed `split_idx` is not unpacked on the block-sparse paths (regression from #2559)

#1981 Benchmarking FA on 5070Ti

#2699 [FA4] FA4 performs about the same as FA2 on B300.

#2725 FlashAttention-2 on AMD Radeon RX 9060 XT / gfx1200 with ROCm 7.14

#1810 FA3 attention sinks blackwell sm120

#2750 flash_attn_func crashes with cudaErrorInvalidValue (SMEM overflow) at head_dim=256, causal, bf16 on sm_86/sm_89 (SM80 tile heuristic not head-dim-aware)

#2425 [Wheel] Pre-built flash-attn 2.8.3 for CUDA 12 + PyTorch 2.11 (Python 3.10-3.13)

#2649 FlashAttentionForwardSm120: use_tma_O incorrectly True on SM_121 → AttributeError on tma_atom_O=None

#1632 Clarification on autotune using the triton backend for amd cards