Dao-AILab/flash-attention

Issues

• 6 hours building on 5090 Windows

  #1560 opened 10 days ago by HDANILO
  6

• undefined symbol: _ZNK3c1011StorageImpl27throw_data_ptr_access_errorEv

  #1509 opened a month ago by luna0804
  2

• ImportError: /home/data/miniconda3/envs/museTalk/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c105ErrorC2ENS_14SourceLocationENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE

  #1518 opened a month ago by gg22mm
  2

• Noticeable diff between `flash_attn_varlen_qkvpacked_func` and `flash_attn_with_kvcache` outputs

  #1569 opened 5 days ago by NonameUntitled
  2

• FlashAttention 3 has numerical inconsistency on H100 with varlen kv cache

  #1570 opened 5 days ago by preminstrel
  2

• `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.

  #1530 opened 6 days ago by allentsouhuang
  1

• no cuda runtime while pip installing flash_attn=2.7.4.post1

  #1566 opened 6 days ago by hwang136
  0

• Compilation Error

  #1568 opened 7 days ago by hwang136
  1

• installation is too long

  #1567 opened 7 days ago by diablounbounded
  0

• flash_attetion doesn't work on windows+wsl RTX 5090

  #1563 opened 9 days ago by HDANILO
  16

• Window Attention cannot benefit from GQA

  #1564 opened 8 days ago by LKJacky
  0

• Triton version is faster in both forward and backward when head dim is 64 but slower in both when head dim is 128

  #1556 opened 11 days ago by SonicZun
  10

• Forced Flash Attention For Vision Description Models Excludes Turing GPUs

  #1557 opened 10 days ago by amadou-6e
  1

• Serialization when creating hash reduces pipeline option customizability

  #1558 opened 10 days ago by amadou-6e
  1

• FA3 support for 50 series Nvidia GPU(sm_120)

  #1549 opened 15 days ago by shahizat
  1

• Is FlashAttention-3 incompatible with Windows compilation? #1551

  #1553 opened 12 days ago by digitalusman99
  1

• How to Implement Block Sparse Attention with FlashAttention?

  #1555 opened 12 days ago by RafaDD
  0

• Is flash-attention3 unable to be compiled on Windows?

  #1551 opened 14 days ago by lestersssss
  0

• Bad `RPATH` in pre-compiled Linux wheels

  #1548 opened 16 days ago by sisp
  2

• window11install flash_attn train issue

  #1550 opened 14 days ago by partcompany
  0

• Cannot use FlashAttention-2 backend for head size 72.

  #1542 opened 19 days ago by ifyoulovexxz
  3

• FlashAttention forward support for Turing

  #1533 opened a month ago by ssiu
  4

• Wheel compilation hangs

  #1545 opened 17 days ago by KLL535
  0

• [QST] Got compilation error when compiling flash-attention-3 with CUDA 12.3

  #1544 opened 18 days ago by umiswing
  0

• Does v's head_dim have to equal q's or k's ?

  #1541 opened 20 days ago by janelu9
  1

• RTX 5090 support - built from scratch but still getting - RuntimeError: FlashAttention only supports Ampere GPUs or newer.

  #1540 opened 20 days ago by elkay
  2

• Some question about the TFLOP of FA2 backward

  #1525 opened a month ago by foreverpiano
  3

• A way to seamlessly replace `flash-attention` with something to allow for low-data inference on other platforms

  #1527 opened 22 days ago by marinegor
  2

• prebuild wheels dear lord

  #1531 opened a month ago by kunibald413
  2

• build variable FLASH_ATTENTION_TRITON_AMD_ENABLE used at runtime

  #1538 opened 25 days ago by trixirt
  0

• bug in flash_attn_varlen_func when max_seqlen == 1

  #1537 opened 25 days ago by zhenwendai
  0

• Which branch is flash attention 3 correct branch?

  #1523 opened a month ago by FurkanGozukara
  7

• Inquiry About flash_attn3 b1 usability

  #1516 opened 25 days ago by luyvlei
  3

• Flash_attn 1.x whl ?

  #1508 opened 25 days ago by darkon12
  1

• Does it support RTX 8000?

  #1513 opened 25 days ago by youde2000
  1

• Flash attention hangs when running an openchat model inside a docker container

  #1532 opened a month ago by zeionara
  1

• flash_attn_3 can not be import

  #1536 opened 25 days ago by zdxff
  0

• 我使用 src/train.py 进行训练，提示 --tokenized_path没有这个参数，用的是0.9.2的版本

  #1535 opened 25 days ago by yifanhunter
  0

• How to speed up setup? AMD 9950X 96GB RAM and RTX 5090

  #1519 opened a month ago by FurkanGozukara
  3

• AMD 9070 and XT launch

  #1529 opened a month ago by bennmann
  0

• How To Get Attention Map with Flash-Attn

  #1520 opened a month ago by i11box
  2

• Is there a function that can be used solely to compute the dot product of q and k?

  #1526 opened a month ago by liushuhao1130
  1

• Flash Attention 3 compile fails (able to compile 2) : TypeError: _write_ninja_file() got an unexpected keyword argument 'sycl_cflags'

  #1524 opened a month ago by FurkanGozukara
  13

• Why rotary embedding is only implemented for Split KV kernels

  #1521 opened a month ago by phantaurus
  7

• Does it support 2080Ti?

  #1514 opened a month ago by Lsnxiaoxiong
  3

• Installing flash-attn & torch with requirements.txt raises build error

  #1515 opened a month ago by erenirmak
  1

• [QST] Why are Layout<Shape<Int<kNWarps>, _1, _1>> and Tile<Int<16 * kNWarps>, _16, _16>> used in kernel_traits.h for TiledMma?

  #1512 opened a month ago by March-H
  6

• Where is the flash_attn_2_cuda

  #1506 opened a month ago by VirgoAsumita
  0

• I have compiled for Windows but where is the WHL file? Finished processing dependencies for flash-attn==2.7.4.post1 - compiled for CU128 for RTX 5000 series

  #1510 opened a month ago by FurkanGozukara
  1

• na

  #1505 opened a month ago by Sk3W3d
  0