.. |
__init__.py
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
benchmark_attn.py
|
3669b25206
bwd benchmark + small fixes (#1129)
|
4 ماه پیش |
benchmark_flash_attention_fp8.py
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
benchmark_split_kv.py
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
combine.h
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |
copy_paged_sm90_tma.hpp
|
284e2c6e5b
Make FA3 paged attention ready for upgrade to Cutlass 3.6 (#1331)
|
1 ماه پیش |
copy_paged_sm90_tma_cutlass35.hpp
|
0823cf7b5d
Fix FA3 Varlen Performance regression (#1361)
|
1 هفته پیش |
copy_paged_sm90_tma_cutlass36.hpp
|
0823cf7b5d
Fix FA3 Varlen Performance regression (#1361)
|
1 هفته پیش |
epilogue_bwd_sm90_tma.hpp
|
db80387343
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
2 ماه پیش |
epilogue_fwd_sm90_tma.hpp
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash.h
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
1 ماه پیش |
flash_api.cpp
|
1feb711f46
Fix compilation with clang on ARM64 (#1285)
|
1 هفته پیش |
flash_attn_interface.py
|
284e2c6e5b
Make FA3 paged attention ready for upgrade to Cutlass 3.6 (#1331)
|
1 ماه پیش |
flash_bwd_hdim128_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
4 ماه پیش |
flash_bwd_hdim128_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_bwd_hdim256_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_bwd_hdim64_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
4 ماه پیش |
flash_bwd_hdim64_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_bwd_hdim96_bf16_sm90.cu
|
bafe253042
[FA3] Bwd
|
4 ماه پیش |
flash_bwd_hdim96_fp16_sm90.cu
|
bafe253042
[FA3] Bwd
|
4 ماه پیش |
flash_bwd_kernel.h
|
1c9717d699
address comments
|
2 ماه پیش |
flash_bwd_launch_template.h
|
1c9717d699
address comments
|
2 ماه پیش |
flash_bwd_postprocess_kernel.h
|
db80387343
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
2 ماه پیش |
flash_bwd_preprocess_kernel.h
|
db80387343
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
2 ماه پیش |
flash_fwd_hdim128_bf16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_bf16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_bf16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_bf16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_bf16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
5 ماه پیش |
flash_fwd_hdim128_e4m3_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_e4m3_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_e4m3_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_e4m3_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_e4m3_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
4 ماه پیش |
flash_fwd_hdim128_fp16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_fp16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_fp16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_fp16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_fp16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim128_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_fwd_hdim256_bf16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_bf16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_bf16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_bf16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_bf16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
5 ماه پیش |
flash_fwd_hdim256_e4m3_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_e4m3_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_e4m3_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_e4m3_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_e4m3_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
4 ماه پیش |
flash_fwd_hdim256_fp16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_fp16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_fp16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_fp16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_fp16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim256_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_fwd_hdim64_bf16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_bf16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_bf16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_bf16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_bf16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_bf16_sm90.cu
|
74b0761ff7
[FA3] BF16 forward
|
5 ماه پیش |
flash_fwd_hdim64_e4m3_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_e4m3_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_e4m3_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_e4m3_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_e4m3_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_e4m3_sm90.cu
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
4 ماه پیش |
flash_fwd_hdim64_fp16_gqa16_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_fp16_gqa2_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_fp16_gqa32_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_fp16_gqa4_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_fp16_gqa8_sm90.cu
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_hdim64_fp16_sm90.cu
|
7f67966cc7
FA3 initial code release
|
5 ماه پیش |
flash_fwd_kernel.h
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
flash_fwd_launch_template.h
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
1 ماه پیش |
kernel_traits.h
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
mainloop_bwd_sm90_tma_gmma_ws.hpp
|
1c9717d699
address comments
|
2 ماه پیش |
mainloop_fwd_sm90_tma_gmma_ws.hpp
|
0823cf7b5d
Fix FA3 Varlen Performance regression (#1361)
|
1 هفته پیش |
named_barrier.hpp
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |
seq_len.h
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
1 ماه پیش |
setup.py
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
softmax.h
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |
static_switch.h
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
1 ماه پیش |
test_attn_kvcache.py
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
test_flash_attn.py
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
1 ماه پیش |
test_kvcache.py
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 ماه پیش |
tile_scheduler.hpp
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |
tile_scheduler_bwd.hpp
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |
utils.h
|
478ee666cc
Make namespace comment consistent (#1305)
|
1 ماه پیش |