Repository Issue Activity (beta)

#2691 Make dropout-disabled FA2 ABI stable with libtorch + cpython

#2413 FA4 support for RTX 6000 Pro Blackwell

#2635 Training Extremely Slow on Qwen3.5-35B-A3B + 8×B300 (280s/step), py-spy Shows FlashAttention-4 CUTLASS JIT Compilation During Backward

#2299 [Wheel] Pre-built flash-attn 2.8.3 for CUDA 12 + PyTorch 2.10 (Python 3.12)

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

#2677 [Cute,Bwd,Sm100] Sparse MLA backward: two findings about dPsum stride divisibility and redundant elect

#2427 Support head_dim=512 for Gemma 4 global attention layers

#975 ImportError: /home/linjl/anaconda3/envs/sd/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c105ErrorC2ENS_14SourceLocationENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

#1926 pip install flash-attn hangs on Google Colab A100 with latest environment

#1672 JVP support for FlashAttention

#2672 Any plans to support MLA bwd in FA4 on SM 100?

#879 Has anyone successfully compiled this on ARM Linux (aarch64)?

#2668 H100 benchmark: full-layer joules/token vs FlashAttention-2 (honest attn kernel comparison)

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

#2665 [Cute,FA4,Fwd,Sm100] Correctness issue when QKV is non-contiguous on SM100 for hdim 256

#2362 How to reproduce reported 1613 TFLOPS at seqlen=32k on Blackwell?

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

#2653 _flash_attn_fwd/_flash_attn_bwd: DSLRuntimeError — window_size_left/right expects Int32 but gets plain Python int

#2652 atomic_add_fp32: TypeError on nvvm.atomicrmw(res=T.f32(), ...) — 'res' kwarg removed in cutlass-dsl 4.5.2

#2651 FlashAttentionBackwardSm80: compute_softmax_scale_log2 overwrites softmax_scale with None → DSLRuntimeError

#2650 _flash_attn_bwd: UnboundLocalError for dQ_single_wg on SM_12x (arch//10 == 12)

#2580 ROCM + triton build broken

#2646 CuteDSL profiling for FA4

#2639 [Cute,Fwd,Sm100] FP8 fwd is incompatible with nvidia-cutlass-dsl>=4.5.2

#2637 Bare tarfile.extractall into Predictable Cache During Build in flash-attention