squall/flash-attention
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
562 Commits 1 Branch 0 Tags 7.9 MiB
000b67f5d8
Commit Graph

166 Commits

Author SHA1 Message Date
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
Aman Gupta Karmani
866a9d33f9
bump cutlass submodule (#504) 2023-08-30 10:32:04 -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
dan_the_3rd
c3f2a632aa
[ft_attention] Fix for seqlen=8136 (#488) 
When seqlen=8136, `smem_sz = 48840`, and apparently starting the kernel returns an `invalid argument` CUDA error.

`48840 < 48 * 1024` but apparently it's still above the limit somehow..?
Tested on A100
 2023-08-28 10:00:22 -07:00
Tri Dao
757058d4d3 Update Cutlass to v3.2.0 2023-08-27 23:47:28 -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
Joel Lamy-Poirier
767b71ccf0
Fix random state for dropout_layer_norm (#315) 2023-07-23 15:05:13 -07:00
Tri Dao
a157cc8c9b [FT] Implement MQA/GQA 2023-07-22 23:47:01 -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
Ikko Eltociear Ashimine
dfc60f6b7d
[LayerNorm] Fix typo in ln_api.cpp 
unintialized -> uninitialized
 2023-07-20 01:16:16 +09:00
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
Tri Dao
2800efc71f [FT] rotary_cos/sin should have batch_size dimension 2023-07-06 15:33:33 -07:00
Tri Dao
3a9bfd076f [FT] rotary_cos/sin should have shape (dim) instead of (seqlen, dim) 2023-07-03 09:41:04 -07:00
Tri Dao
62e9814466 [Rotary] Make sure frequency calculation is in fp32 2023-07-02 16:39:39 -07:00
Tri Dao
27f8f890df [FusedDense] Allocate lt_workspace on input device 2023-05-30 14:17:26 -07:00
Tri Dao
48bc6eacd6 [Gen] Add rotary base as an argument to FT attention kernel 2023-05-30 13:38:34 -07:00
Tri Dao
ad113948a6 [Docs] Clearer error message for bwd d > 64, bump to v1.0.4 2023-04-26 09:19:48 -07:00
Tri Dao
311d6606bf [Gen] Fix FT kernel smem size, CG when batch size changed 2023-04-20 17:03:13 -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
7d25a4ec4f Handle FlashAttnQKVPackedSplitFunc by making rng_state optional in backward 
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
 2023-04-13 06:25:52 +00:00
Kirthi Shankar Sivamani
315fd31f0c
Merge branch 'HazyResearch:main' into enable_cuda_graph_capture 2023-04-12 22:42:24 -07:00
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
dec4f2e910 [FusedDense] Set workspace size to 32M for Hopper and 4M for others 2023-04-06 23:40:15 -07:00
Tri Dao
393882bc08 [LayerNorm] Implement LN with parallel residual, support dim 8k 2023-03-31 14:23:45 -07:00
Tri Dao
f5d0fbd468 [FT] Fix FT's single query attention for bf16 hdim128 rotary 2023-03-28 21:27:00 -07:00
Tri Dao
dc08ea1c33 Support H100 for other CUDA extensions 2023-03-15 16:59:27 -07:00
Tri Dao
1b18f1b7a1 Support H100 2023-03-15 14:59:02 -07:00
Tri Dao
e45a46a5b7 [Rotary] Implement GPT-J style (interleaved) rotary 2023-03-14 14:35:53 -07:00
Tri Dao
6b4a48218e [FA] Remove unused variable rng_engine_inputs 2023-01-25 15:32:40 -08:00
Tri Dao
eb33e587e9 [LayerNorm] Rename x1 -> residual 2023-01-19 13:07:27 -08:00
Tri Dao
88173a1aaf [FusedDense] Support relu, rename FusedDenseGeluDense -> FusedMLP 2023-01-17 18:12:27 -08:00
Tri Dao
f1e01c27ba [Gen] Pass qkv_stride to ft_attention kernel for batched generation 2023-01-15 15:20:01 -08:00
Tri Dao
7c2191542a [Gen] Make generation work with Tensor Parallel 2023-01-15 11:34:27 -08:00
Tri Dao
6738d9477d [LayerNorm] Implement RMS Norm 2023-01-06 17:34:22 -08:00
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
be1afaa276 [Gen, FT] Use fp32 accum for FMA 2023-01-03 22:09:22 -08:00
Tri Dao
f266fc7262 [Gen, FT] Use tlength instead of params.timestep for rotary 2023-01-03 17:46:55 -08:00
Tri Dao
a01d1213d7 [Gen] Add kernel from FasterTransformer for benchmarking 2023-01-03 17:37:43 -08:00
Tri Dao
a8cfe51551 Implement Tensor Parallel for transformer Block 2022-12-25 14:08:21 -08:00
Tri Dao
1e712ea8b0 Implement TensorParallel for MHA 2022-12-25 11:39:55 -08:00
Tri Dao
226a1b721d Implement TensorParallel for FusedDense and FusedDenseGeluDense 2022-12-24 11:48:56 -08:00
Tri Dao
dff68c2b22 Add smoothing for CrossEntropyParallel, rename to CrossEntropyLoss 2022-12-23 14:51:08 -08:00
Tri Dao
e68ebbe89a Simplify FusedDense 2022-12-22 21:25:31 -08:00
Tri Dao
5db330519a [LayerNorm] Support taking subset of input or subset of output 2022-12-12 22:16:14 -08:00
Tri Dao
ae137ed17a [LayerNorm] Fuse LayerScale 2022-12-10 23:28:23 -08:00
Tri Dao
8c6609ae1a [LayerNorm] Support all dimensions up to 6k (if divisible by 8) 2022-12-09 02:06:22 -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
0bf5e50038 Release training code 2022-11-28 17:34:40 -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
39ed597b28 [LayerNorm] Compile for both sm70 and sm80 2022-11-17 11:45:11 -08:00
Tri Dao
43ab0b5205 Mention that some CUDA extensions have only been tested on A100s 2022-11-15 07:10:25 -08:00
Tri Dao
e4d3013e15 [LayerNorm] Check cuda error after querying ctas_per_sm 2022-11-15 07:05:13 -08:00
Tri Dao
2e33fc8e36 Add GPT and ViT models 2022-11-13 22:30:23 -08:00
Tri Dao
fa6d1ce44f Add fused_dense and dropout_add_layernorm CUDA extensions 2022-11-13 21:59:20 -08:00
Tri Dao
7c9953815a Add fused cross entropy loss 2022-11-12 21:58:41 -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
ca81f32e04 Implement rotary embedding in CUDA 2022-11-04 22:42:01 -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
97e13de2b4 Cast q.get_device() to char to avoid compiler warning (narrowing) 2022-10-24 15:59:49 -07:00
Tri Dao
ed553e9238 Add Megatron attention implementation for benchmarking 2022-10-23 23:04:16 -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
YangShu
ff07250e8f
fix typo in function mha_fwd 
as title.
 2022-10-17 16:13:47 +08:00
Tri Dao
52fb4b729b Fix #54: set device for multi-GPU case 2022-10-16 12:51:26 -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
c0daa62eaa Add type check (fp16) in the forward pass 2022-06-26 11:41:30 -07:00
Tri Dao
ea38d3d261 Fix race condition in backward pass (smem_dq) 2022-06-25 18:02:30 -07:00