squall/flash-attention
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
704 Commits 1 Branch 0 Tags 7.9 MiB
c1d146cbd5
Commit Graph

125 Commits

Author SHA1 Message Date
Tri Dao
65c234ed90 Don't over-allocate dq_accum in case of varlen 2023-09-24 00:36:07 -07:00
Tri Dao
1879e089c7 Reduce number of templates for headdim > 128 2023-09-23 22:24:30 -07:00
Tri Dao
2d8ea9a530 Swap seqlen_q and ngroups when seqlen_q=1 (h/t Daniel Haziza) 2023-09-20 23:38:22 -07:00
Tri Dao
43617deab9 Remove template for (IsEvenMN=T, IsEvenK=F) to speed up compilation 2023-09-18 12:21:36 -07:00
Tri Dao
c984208ddb Set block size to 64 x 64 for kvcache to avoid nvcc segfaults 2023-09-17 16:14:58 -07:00
Tri Dao
ccbb14f38e Implement rotary embedding in flash_attn_with_kvcache 2023-09-16 01:20:16 -07:00
Tri Dao
56b7fc6ee0 Simplify the implementation of KVcache attn by appending KV first 2023-09-13 15:55:48 -07:00
Tri Dao
bb9beb3645 Remove some unused headers 2023-09-12 12:37:10 -07:00
Tri Dao
ee77b931b9 Swap seqlen_q and nheads for MQA to speed it up (h/t Daniel Haziza) 2023-09-10 22:56:33 -07:00
Tri Dao
37c6e05406 Implement flash_attn_with_kvcache 2023-09-04 00:11:44 -07:00
Tri Dao
6a89b2f121 Remove constexpr in launch template to fix CI compilation 2023-09-03 22:59:41 -07:00
Tri Dao
1dc1b6c8f2 Bump to v2.1.2 2023-09-03 22:23:05 -07:00
Tri Dao
5953c4f58c Remove unused sdPsum in dot_do_o function 2023-09-03 20:44:07 -07:00
Tri Dao
26d7d92f3d Fix splitKV combine function when local LSEs are all -inf 2023-09-03 11:39:09 -07:00
Sophia Wisdom
37e32febba
Remove commented out code in bwd (#512) 
* Remove lots of comments

* Remove unused traits
 2023-09-01 16:43:58 -07:00
Sophia Wisdom
dd8a754915
Remove old code in utils.h (#511) 2023-09-01 15:32:09 -07:00
Tri Dao
31920dda5f Fix typo with lse_max == -INFINITY 2023-08-29 21:48:59 -07:00
Tri Dao
b1fbbd8337 Implement splitKV attention 2023-08-29 00:58:29 -07:00
Tri Dao
7a983df742 Use generate_kernels.py script from Driss Guessous 2023-08-28 13:34:12 -07:00
Tri Dao
9e5e8bc91e Change causal mask to be aligned to bottom-right instead of top-left 2023-08-24 23:41:07 -07:00
BoxiangW
e07aa036db
Support flash attention 2 with causal masking when KV's seq length is longer than Q's seq length. (#436) 2023-08-24 16:42:34 -07:00
Tri Dao
bcfa7c9751 [FusedDense] Run black on fused_dense.py 2023-08-16 23:41:36 -07:00
Tri Dao
c65b5106ac Fix Bwd NaN for varlen when seqlen_q >> seqlen_k and causal 2023-08-16 15:12:36 -07:00
Tri Dao
dbd7923782 Prepare for Cutlass 3.2 2023-08-13 15:24:32 -07:00
Tri Dao
3524e13c11 Update to Cutlass 3.1 2023-08-13 13:53:17 -07:00
Tri Dao
1c41d2b0e5 Fix race condition in bwd (overwriting sK) 2023-08-01 09:00:10 -07:00
Tri Dao
a4f148b6ab Fix masking of bwd when seqlen is not divisible by 128 2023-07-31 17:46:34 -07:00
Kirthi Shankar Sivamani
a03f6f8e9e
Enable CUDA graphs (#386) 
* Add RNG state to kernel launch params

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Save seed and offset for backward

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Single thread write to global mem

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1colblock get seed and offset from launch params

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1rowblock get seed and offset from launch params

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change forward c++ APIs to save RNG state for backward

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change backward c++ APIs to set RNG state for bprop launcher

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fixes

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Python side API changes

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fix; only save seeds instead of full offset

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Account for 3D grid size

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
 2023-07-27 16:11:34 -07:00
Tri Dao
9ee0ff1d9b Fix using dO stride for O, which can cause memory error in bwd 2023-07-20 17:39:57 -07:00
danthe3rd
538d570c96 Fix compile error on MSVC 
See also: https://stackoverflow.com/questions/55136414/constexpr-variable-captured-inside-lambda-loses-its-constexpr-ness
 2023-07-19 08:04:57 +00:00
Tri Dao
4f285b3547 FlashAttention-2 release 2023-07-17 06:21:34 -07:00
Kirthi Shankar Sivamani
45567a25a2 only 1 thread writes to global mem in fprop 
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
 2023-04-15 06:09:41 +00:00
Kirthi Shankar Sivamani
31018c5fa0 Support CUDA graph capture 
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
 2023-04-12 16:53:22 -07:00
Tri Dao
1b18f1b7a1 Support H100 2023-03-15 14:59:02 -07:00
Tri Dao
a1f49a2b92 [Compilation] Change BOOL_SWITCH to fix Windows compilation 
Follow xFormers's DISTPATCH_BOOL. Haven't tested it on Windows.
 2023-01-06 14:40:58 -08:00
Tri Dao
8a2ece89f7 Simplify BOOL_SWITCH macro to fix compiling error on gcc 7 2022-12-06 14:38:32 -08:00
Tri Dao
9bc63d1e2d Fix typo in comments 2022-11-25 16:35:08 -08:00
Tri Dao
d95ee1a95d Speed up compilation by splitting into separate .cu files 2022-11-25 16:30:18 -08:00
Tri Dao
6998e0ecdb Fix out-of-bound memory read 2022-11-09 09:34:14 -08:00
Tri Dao
557781933d Parallelize CUDA bwd along seqlen_k instead of seqlen_q 
This is faster since we only need to do atomic adds on dq, instead of atomic
adds on both dk and dv.
 2022-11-05 16:26:17 -07:00
Tri Dao
c422fee377 Get rid of o_rows_are_valid since we don't have headdim=16 anymore 2022-10-24 17:29:36 -07:00
Tri Dao
46fd2a20b2 Support all head dims that are multiples of 8, up to 128 2022-10-24 16:04:21 -07:00
Tri Dao
9e92a1f2d2 Attempt to use atomicCAS to replace atomicAdd(bfloat16) 2022-10-23 16:22:43 -07:00
Tri Dao
a5a8806d1a Split bwd on the seqlen_q dimension 2022-10-23 11:35:15 -07:00
Tri Dao
871db47941 Don't need to run configure for the forward pass 2022-10-21 18:22:27 -07:00
Tri Dao
7fc39832e2 Use block_size=128 for headdim=128 on SM80 
Previously we were using block_size=256.
 2022-10-21 13:19:54 -07:00
Tri Dao
a44f48df5a Split fwd on the seqlen_q dimension 2022-10-21 12:04:27 -07:00
Tri Dao
1aa6d7d9b6 Rework dropout to decouple forward and backward 
They don't have to have the same block size, number of threads, etc.
 2022-10-21 12:04:27 -07:00
Tri Dao
5badfb7848 Implement attention kernel that splits the batch into two 2022-10-13 20:49:02 -07:00
Eric Engelhart
2211db5fab Fixed switch statement, thanks @yocabon 2022-10-04 21:31:39 -04:00
Eric Engelhart
9d7fd5b6e7 Replace BOOL_SWITCH with FP16_SWITCH to work around MSVC bug with constexpr variables and templates 2022-10-04 21:31:39 -04:00
Tri Dao
8166063a55 Use block_size=128 for d=128 on SM86 to avoid exceeding smem limit 2022-09-12 14:21:29 -07:00
Tri Dao
bc2c210254 Don't nest BOOL_SWITCH to work around gcc 7 bug 2022-07-11 10:28:46 -07:00
Tri Dao
de19de7ab1 Implement for bf16 2022-07-09 23:31:56 -07:00
Tri Dao
6a77a6da10 Refactor gemm_cl to template on either __half or __nv_bfloat16 2022-07-09 23:18:26 -07:00
Tri Dao
e518a4b327 Refactor to template on __half, implement bf16 util functions 2022-07-09 23:18:26 -07:00
Tri Dao
2dc1b205f6 Fix Illegal Memory Access bug in fwd when d=16 2022-07-09 23:17:14 -07:00
Tri Dao
5b838a8bef Apply dropout scaling to dQ and dK instead of to V (in bwd) 
Theoretically this might have lower numerical error since the scaling is in
fp32 instead of fp16 (not sure, I haven't thought too carefully about it).
However, in practice, the numerical errors seem about the same.
 2022-07-03 17:53:37 -07:00
Tri Dao
a5559a0e75 Do P * dP (pointwise) in the bwd in fp32 instead of fp16 2022-07-03 17:52:05 -07:00
Tri Dao
6c3a8c65af Implement cross attention 2022-07-03 17:48:12 -07:00
Tri Dao
f66603cb6f Support batch size > 64K by swapping grid.x and grid.y 2022-06-29 23:16:24 -07:00
Tri Dao
ea38d3d261 Fix race condition in backward pass (smem_dq) 2022-06-25 18:02:30 -07:00
Tri Dao
eeca63a72a Bug fix: wrong smem_o write pointer for d=16 2022-06-25 15:18:33 -07:00
Tri Dao
5d07483bbc Refactor Gmem code to store q, k, v pointers separately 2022-06-12 16:37:32 -07:00
Tri Dao
d3e6440958 Implement bwd for head dim 128 2022-06-11 17:52:36 -07:00
Tri Dao
0d854692c6 Implement fwd for head dim 128 2022-06-11 17:52:36 -07:00
Tri Dao
321c57d07d Set block size of SM75 fwd to 256 if there's no dropout 
This speeds up the fwd by 1.5x.
 2022-06-04 16:51:28 -07:00
Tri Dao
d380e87fb6 Don't use Smem_dp_sum in backward pass 
To reduce smem usage for SM75
 2022-06-04 16:01:36 -07:00
Tri Dao
b17c6fe235 Reduce smem usage for Q and dO in the backward pass 
From 4KB per buffer to 2KB per buffer. This saves us 8KB of smem (each Q and dO
have 2 buffers)
 2022-06-03 16:59:11 -07:00
Tri Dao
2712aa4c8d Support Turing mma instructions 2022-06-03 16:58:44 -07:00
Tri Dao
050873327e Remove softmax fp16 max 2022-06-02 14:09:46 -07:00
Tri Dao
14dc326e59 Use Cutlass gemm as WarpMma 2022-06-02 10:33:32 -07:00
Tri Dao
e78e7c9553 Remove old backward 2022-06-02 10:13:44 -07:00
Tri Dao
c41479d66d Support SM86 GPUs 2022-06-01 18:49:47 -07:00
Tri Dao
9dbc491aa5 Rename, add benchmarking script 2022-05-26 13:57:38 -07:00