 Tri Dao
|
4ead9bd7cc
[FA3] Varlen forward
|
hai 4 meses |
|
Tri Dao
|
74b0761ff7
[FA3] BF16 forward
|
hai 5 meses |
|
Tri Dao
|
898dd4bbf2
Pass seqused_k to _flash_attn_varlen_forward
|
hai 5 meses |
|
Tri Dao
|
7ef24848cf
Add FA3 image
|
hai 5 meses |
|
Tri Dao
|
7f67966cc7
FA3 initial code release
|
hai 5 meses |
|
Tri Dao
|
b4a9dd6c9c
Temporarily switch to cutlass fork for more shapes
|
hai 5 meses |
|
Tri Dao
|
7551202cb2
Bump to v2.6.1
|
hai 5 meses |
|
Tri Dao
|
844912dca0
[CI] Switch from CUDA 12.2 to 12.3
|
hai 5 meses |
|
Tri Dao
|
40e534a7f6
Implement cache_leftpad
|
hai 5 meses |
|
Tri Dao
|
116b05f9b0
[CI] Compile with pytorch 2.4.0.dev20240514
|
hai 5 meses |
|
Tri Dao
|
da11d1b853
Bump v2.6.0
|
hai 5 meses |
|
Tri Dao
|
d0787acc16
Relax dropout_fraction test
|
hai 5 meses |
|
Tri Dao
|
dca6d89da4
Don't support softcap and dropout at the same time
|
hai 5 meses |
|
Tri Dao
|
81e01efd4b
More typo fixes
|
hai 5 meses |
|
Tri Dao
|
72e27c6320
Fix typo with softcapping
|
hai 5 meses |
|
Tri Dao
|
3d41db3e2c
Only test backward if there's no softcapping
|
hai 5 meses |
|
Tri Dao
|
908511b2b6
Split into more .cu files to speed up compilation
|
hai 5 meses |
|
Tri Dao
|
1d536d7de5
Minor cleanup of softcapping
|
hai 5 meses |
|
Tri Dao
|
beb2bf2a32
Drop support for pytorch 1.12, 1.13, and python 3.7
|
hai 5 meses |
 Phil Wang
|
f4628b43ec
missing commas and backwards return arguments (#1032)
|
hai 5 meses |
 Nicolas Patry
|
8f873cc6ac
Implement softcapping. (#1025)
|
hai 5 meses |
 Jianwei Dong
|
4e8d60069f
Add the return_softmax_lse parameter to the flash_attn_with_kvcache function to allow returning the logsumexp of the attention scores. (#989)
|
hai 5 meses |
 muoshuosha
|
6df7e0a02e
Fix the varlen deterministic test (#1023)
|
hai 5 meses |
 66RING
|
9486635c92
Fix typos of comments about shape. (#837)
|
hai 5 meses |
 JDKWangGuan
|
0d810cfb73
Fix KeyError handling for non-existing key in state_dict.pop() (#898)
|
hai 5 meses |
 cao lei
|
6a2a16e994
fix typo (#974)
|
hai 5 meses |
|
Nicolas Patry
|
5bf201966a
Fixing argument checking when using `seqlenq_ngroups_swapped`. (#976)
|
hai 5 meses |
 Liang
|
ab59ec3590
remove swizzle part of `sV.data()` to get a completely non-swizzle `sVtNoSwizzle` (#984)
|
hai 5 meses |
 Grigory Sizov
|
f816dee63c
Support unpadded LSE layout (#970)
|
hai 5 meses |
|
Tri Dao
|
320fb59487
Update citation
|
hai 6 meses |
