Laurent Mazare 0ace420e66 Flash attention without padding (varlen). (#281) ...

* Expose the seqlen variable for flash-attn without padding.

* Fix the batched call.

* Adapt for the varlen variant.

* No need to set the batch strides when in varlen mode.

* Add a test (disabled at the moment).

* Get the test to work properly.

2023-07-31 09:45:39 +01:00

..

block_info.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

flash_api.cu

Flash attention without padding (varlen). (#281)

2023-07-31 09:45:39 +01:00

flash_fwd_hdim32_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim32_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim64_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim64_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim96_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim96_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim128_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim128_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim160_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim160_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim192_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim192_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim224_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim224_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim256_bf16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_hdim256_fp16_sm80.cu

Again set a few extra params in flash-attn. (#245)

2023-07-26 14:16:37 +01:00

flash_fwd_kernel.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

flash_fwd_launch_template.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

flash.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

kernel_traits.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

philox.cuh

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

softmax.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

static_switch.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00

utils.h

Add flash attention (#241)

2023-07-26 07:48:10 +01:00