david/flash-attention: flash-attention from https://github.com/Dao-AILab/flash-attention @ c272619bf5dd845bcd70b28fa3c8ea8c47ffc834

Jay Shah c272619bf5 disable clusters		1 ヶ月前
..
__init__.py	7f67966cc7 FA3 initial code release	5 ヶ月前
benchmark_attn.py	3669b25206 bwd benchmark + small fixes (#1129)	4 ヶ月前
benchmark_flash_attention_fp8.py	483b26e541 add working fp8 varlen	1 ヶ月前
benchmark_split_kv.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
combine.h	478ee666cc Make namespace comment consistent (#1305)	1 ヶ月前
epilogue_bwd_sm90_tma.hpp	db80387343 Add seqused_q in fwd / bwd and seqused_k in bwd.	3 ヶ月前
epilogue_fwd_sm90_tma.hpp	bb169e2a54 revert to single tile scheduler	1 ヶ月前
flash.h	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_api.cpp	bb169e2a54 revert to single tile scheduler	1 ヶ月前
flash_attn_interface.py	483b26e541 add working fp8 varlen	1 ヶ月前
flash_bwd_hdim128_bf16_sm90.cu	bafe253042 [FA3] Bwd	4 ヶ月前
flash_bwd_hdim128_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_bwd_hdim256_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_bwd_hdim64_bf16_sm90.cu	bafe253042 [FA3] Bwd	4 ヶ月前
flash_bwd_hdim64_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_bwd_hdim96_bf16_sm90.cu	bafe253042 [FA3] Bwd	4 ヶ月前
flash_bwd_hdim96_fp16_sm90.cu	bafe253042 [FA3] Bwd	4 ヶ月前
flash_bwd_kernel.h	1c9717d699 address comments	2 ヶ月前
flash_bwd_launch_template.h	1c9717d699 address comments	2 ヶ月前
flash_bwd_postprocess_kernel.h	db80387343 Add seqused_q in fwd / bwd and seqused_k in bwd.	3 ヶ月前
flash_bwd_preprocess_kernel.h	db80387343 Add seqused_q in fwd / bwd and seqused_k in bwd.	3 ヶ月前
flash_fwd_hdim128_bf16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_bf16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_bf16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_bf16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_bf16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_bf16_sm90.cu	74b0761ff7 [FA3] BF16 forward	5 ヶ月前
flash_fwd_hdim128_e4m3_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_e4m3_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_e4m3_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_e4m3_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_e4m3_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_e4m3_sm90.cu	5018ac6ac5 Fp8 kernel with "in-kernel" transpose of V in producer (#1100)	4 ヶ月前
flash_fwd_hdim128_fp16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_fp16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_fp16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_fp16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_fp16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim128_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_fwd_hdim256_bf16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_bf16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_bf16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_bf16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_bf16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_bf16_sm90.cu	74b0761ff7 [FA3] BF16 forward	5 ヶ月前
flash_fwd_hdim256_e4m3_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_e4m3_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_e4m3_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_e4m3_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_e4m3_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_e4m3_sm90.cu	5018ac6ac5 Fp8 kernel with "in-kernel" transpose of V in producer (#1100)	4 ヶ月前
flash_fwd_hdim256_fp16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_fp16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_fp16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_fp16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_fp16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim256_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_fwd_hdim64_bf16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_bf16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_bf16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_bf16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_bf16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_bf16_sm90.cu	74b0761ff7 [FA3] BF16 forward	5 ヶ月前
flash_fwd_hdim64_e4m3_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_e4m3_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_e4m3_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_e4m3_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_e4m3_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_e4m3_sm90.cu	5018ac6ac5 Fp8 kernel with "in-kernel" transpose of V in producer (#1100)	4 ヶ月前
flash_fwd_hdim64_fp16_gqa16_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_fp16_gqa2_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_fp16_gqa32_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_fp16_gqa4_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_fp16_gqa8_sm90.cu	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
flash_fwd_hdim64_fp16_sm90.cu	7f67966cc7 FA3 initial code release	5 ヶ月前
flash_fwd_kernel.h	c272619bf5 disable clusters	1 ヶ月前
flash_fwd_launch_template.h	c272619bf5 disable clusters	1 ヶ月前
kernel_traits.h	bb169e2a54 revert to single tile scheduler	1 ヶ月前
mainloop_bwd_sm90_tma_gmma_ws.hpp	1c9717d699 address comments	2 ヶ月前
mainloop_fwd_sm90_tma_gmma_ws.hpp	bb169e2a54 revert to single tile scheduler	1 ヶ月前
named_barrier.hpp	bb169e2a54 revert to single tile scheduler	1 ヶ月前
seq_len.h	483b26e541 add working fp8 varlen	1 ヶ月前
setup.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
softmax.h	478ee666cc Make namespace comment consistent (#1305)	1 ヶ月前
static_switch.h	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
test_attn_kvcache.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
test_flash_attn.py	c272619bf5 disable clusters	1 ヶ月前
test_kvcache.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	2 ヶ月前
tile_scheduler.hpp	478ee666cc Make namespace comment consistent (#1305)	1 ヶ月前
tile_scheduler_bwd.hpp	478ee666cc Make namespace comment consistent (#1305)	1 ヶ月前
utils.h	478ee666cc Make namespace comment consistent (#1305)	1 ヶ月前