| .. |
|
__init__.py
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
benchmark_attn.py
|
3669b25206
bwd benchmark + small fixes (#1129)
|
5 mēneši atpakaļ |
|
benchmark_flash_attention_fp8.py
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
4 mēneši atpakaļ |
|
block_info.h
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
epilogue_bwd_sm90_tma.hpp
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
epilogue_fwd_sm90_tma.hpp
|
a3a257c71d
Fix out-of-bound writes for var-seq-len zero-length KVs
|
5 mēneši atpakaļ |
|
flash.h
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_api.cpp
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_attn_interface.py
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_bwd_hdim128_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
flash_bwd_hdim128_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_bwd_hdim256_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_bwd_hdim64_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
flash_bwd_hdim64_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_bwd_hdim96_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
flash_bwd_hdim96_fp16_sm90.cu
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
flash_bwd_kernel.h
|
3f6ff1c1c5
Remove struct : cute::aligned_struct to avoid error with gcc 12
|
5 mēneši atpakaļ |
|
flash_bwd_launch_template.h
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_bwd_postprocess_kernel.h
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_bwd_preprocess_kernel.h
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
flash_fwd_hdim128_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
6 mēneši atpakaļ |
|
flash_fwd_hdim128_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
5 mēneši atpakaļ |
|
flash_fwd_hdim128_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_fwd_hdim256_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
6 mēneši atpakaļ |
|
flash_fwd_hdim256_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
5 mēneši atpakaļ |
|
flash_fwd_hdim256_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_fwd_hdim64_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
6 mēneši atpakaļ |
|
flash_fwd_hdim64_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
5 mēneši atpakaļ |
|
flash_fwd_hdim64_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
6 mēneši atpakaļ |
|
flash_fwd_kernel.h
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
4 mēneši atpakaļ |
|
flash_fwd_launch_template.h
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
kernel_traits.h
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
4 mēneši atpakaļ |
|
mainloop_bwd_sm90_tma_gmma_ws.hpp
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
mainloop_fwd_sm90_tma_gmma_ws.hpp
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
4 mēneši atpakaļ |
|
named_barrier.hpp
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
seq_len.h
|
dfe1a59e4b
Add var-seq-len to FA3 fp16 / bf16 fwd (#1072)
|
6 mēneši atpakaļ |
|
setup.py
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
softmax.h
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
4 mēneši atpakaļ |
|
static_switch.h
|
dfe1a59e4b
Add var-seq-len to FA3 fp16 / bf16 fwd (#1072)
|
6 mēneši atpakaļ |
|
test_flash_attn.py
|
496fdc4f6c
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
4 mēneši atpakaļ |
|
tile_scheduler.hpp
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
5 mēneši atpakaļ |
|
tile_scheduler_bwd.hpp
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
|
utils.h
|
bafe253042
[FA3] Bwd
|
5 mēneši atpakaļ |
Ying Zhang