ROCm flash-attention issues

ROCm / flash-attention

Fast and memory-efficient exact attention

BSD 3-Clause "New" or "Revised" License

141 stars 46 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Added Benchmark for Rotary Decode Kernel + Performance Speed Up for Rotary Kernel

#102 alexkranias-amd opened 2 days ago
0
Dropout

#101 micmelesse opened 1 week ago
0
Enable MQA/GQA in backward

#100 micmelesse closed 1 week ago
0
Added Support for Rotary Positional Embeddings

#99 alexkranias-amd closed 4 days ago
0
[CK_TILE] Fix fmha fwd splitkv block table read out-of-bound

#98 poyenc closed 1 week ago
0
[CK_TILE] Fix fmha fwd splitkv block table read out-of-bound

#97 poyenc closed 1 week ago
0
Fix README

#96 micmelesse closed 2 weeks ago
0
Added Dropout BWD

#95 alexkranias-amd opened 2 weeks ago
1
[Issue]: Memory access fault in tests/test_flash_attn_ck.py::test_flash_attn_qkvpacked[0.17-97-80-False-False-False-False-dtype1]

#94 IMbackK closed 1 week ago
3
[Issue]: gfx1100 is invalid or not supported by Flash-Attention

#93 linqingxu opened 2 weeks ago
1
update Triton commit readme

#92 micmelesse closed 3 weeks ago
0
Update Triton Version

#91 micmelesse closed 3 weeks ago
1
Autotune off by default

#90 micmelesse closed 3 weeks ago
0
Enable sequence_parallel in bwd

#89 micmelesse closed 3 weeks ago
0
[Issue]: Installation of flash-attention failed

#88 Kingmeng-Stack closed 2 weeks ago
7
[Issue]: v2.6.3-cktile and latest commit `53a4f34` tag fail some of the `test_flash_attn_ck.py`

#87 tjtanaa closed 3 weeks ago
0
Enable Vanilla Bwd and Refactor

#86 micmelesse closed 3 weeks ago
1
Support dropout seed offset as pointer

#85 rocking5566 closed 2 weeks ago
0
Update ck backend

#84 rocking5566 closed 1 month ago
0
Integrated Rotary Positional Embeddings (RoPEs) into flash_attn_kvcache

#83 alexkranias-amd closed 2 days ago
0
[Documentation]: Which branch, tag or commit sha should I use to build flash attention for AMD devices ?

#82 etiennemlb closed 2 months ago
6
Clean up for Upstream

#81 micmelesse closed 2 months ago
0
Install flash-attention failed

#80 sdfasfsdfasfasafd opened 2 months ago
15
[Issue]: is scaled_dot_product_attention part of flash attention?

#79 unclemusclez opened 2 months ago
21
Change rounding of bf16 to rtn

#78 rocking5566 closed 2 months ago
0
Sync Ck tile compile flag with rocm6.2

#77 rocking5566 closed 2 months ago
0
Upstream Triton Internal Review

#76 micmelesse closed 2 months ago
0
try sjw/reorder-barrier

#75 micmelesse closed 2 months ago
0
Ck tile/kvcache

#74 rocking5566 closed 2 months ago
2
[Issue]: 2.6.2-cktile - Won't build from source.

#73 nktice closed 1 week ago
3
enable packed layouts and all configs

#72 micmelesse closed 2 months ago
0
[Feature]: Flash Attention 3 Support for MI300X GPUs

#71 codinggosu opened 3 months ago
0
Improve FMHA bwd

#70 rocking5566 closed 3 months ago
0
[Issue]: Installing ROCm Flash-Attention on RHEL

#69 varshaprasad96 opened 4 months ago
4
enable flash_attn_with_kvcache

#68 micmelesse closed 3 months ago
0
[WIP] update to latest ck

#67 rocking5566 closed 4 months ago
0
Use same python as build flash-attn to generate ck kernel

#66 oraluben closed 4 months ago
0
Integrate ck tile backward

#65 rocking5566 closed 5 months ago
0
[Issue]: memory format option is only supported by strided tensors

#64 hyattpd closed 1 week ago
3
Enable fwd and varlen_fwd on AMD

#63 micmelesse closed 5 months ago
0
Feature/seqlenq ngroups swap

#62 poyenc closed 5 months ago
0
Ck tile/flash attention

#61 rocking5566 closed 5 months ago
0
Enable fwd and varlen_fwd on AMD

#60 micmelesse closed 5 months ago
1
Do not return lse if return_softmax == false

#59 rocking5566 closed 5 months ago
0
Fix stride issues in flash_attn_interface

#58 clintg6 opened 5 months ago
0
Integrate with ck

#57 rocking5566 closed 6 months ago
0
Integrate ck with FA

#56 rocking5566 closed 6 months ago
0
update from upstream

#55 rocking5566 closed 6 months ago
0
[Issue]: Expected dout_seq_stride == out_seq_stride to be true, but got false

#54 ehartford closed 6 months ago
2
[Feature]: Support for newer flash-attention versions (e.g. ≥2.1.0)

#53 JiahuaZhao opened 6 months ago
2