Tri Dao 0519920e23 Deal with the case where q or k/v have length 0 il y a 3 semaines
..
instantiations 7f5d73a162 Add env var to disable specific hdim il y a 1 mois
__init__.py 7f67966cc7 FA3 initial code release il y a 6 mois
benchmark_attn.py 82c1aa3514 Move PackGQA epilogue code to pack_gqa.h il y a 2 mois
benchmark_flash_attention_fp8.py c92ca63268 FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173) il y a 5 mois
copy_sm90_bulk_reduce.hpp 29cdfedd80 Use Bulk reduce instead of TMA for dQaccum, split across WGs il y a 2 mois
epilogue_bwd.hpp 0890032358 Implement backward pass for Sm80 il y a 4 semaines
epilogue_fwd.hpp da264e5742 Change file names and class names to include sm90 suffix il y a 1 mois
flash.h 76f14c61c9 Tune fwd tile sizes for Sm86 and Sm89 il y a 4 semaines
flash_api.cpp 0519920e23 Deal with the case where q or k/v have length 0 il y a 3 semaines
flash_attn_interface.py a609d82315 Change extension name to flash_attn_3_cuda il y a 4 semaines
flash_bwd_kernel_sm80.h 0890032358 Implement backward pass for Sm80 il y a 4 semaines
flash_bwd_kernel_sm90.h 659a631f4c Rename bwd classes to include Sm90 suffix il y a 1 mois
flash_bwd_launch_template.h 0890032358 Implement backward pass for Sm80 il y a 4 semaines
flash_bwd_postprocess_kernel.h 0890032358 Implement backward pass for Sm80 il y a 4 semaines
flash_bwd_preprocess_kernel.h 2c996ca25f Use SeqlenInfo for bwd and epilogue il y a 1 mois
flash_fwd_combine_kernel.h 2c996ca25f Use SeqlenInfo for bwd and epilogue il y a 1 mois
flash_fwd_combine_launch_template.h 9fd6b977bb Precompute the pointers in mha_combine kernel il y a 2 mois
flash_fwd_combine_sm80.cu 9fd6b977bb Precompute the pointers in mha_combine kernel il y a 2 mois
flash_fwd_kernel_sm80.h 76f14c61c9 Tune fwd tile sizes for Sm86 and Sm89 il y a 4 semaines
flash_fwd_kernel_sm90.h 659a631f4c Rename bwd classes to include Sm90 suffix il y a 1 mois
flash_fwd_launch_template.h f907a13187 Tune tile sizes for fwd varlen on Sm80 and Sm86 il y a 4 semaines
generate_kernels.py 7f5d73a162 Add env var to disable specific hdim il y a 1 mois
heuristics.h 147ac33a2e Tune num_splits for local, don't split when num_n_blocks is small il y a 1 mois
mainloop_bwd_sm80.hpp 5acb532214 Switch to cutlass v3.6.0, fix perf regression for hdim 128 causal il y a 4 semaines
mainloop_bwd_sm90_tma_gmma_ws.hpp 0890032358 Implement backward pass for Sm80 il y a 4 semaines
mainloop_fwd_sm80.hpp 5acb532214 Switch to cutlass v3.6.0, fix perf regression for hdim 128 causal il y a 4 semaines
mainloop_fwd_sm90_tma_gmma_ws.hpp 5acb532214 Switch to cutlass v3.6.0, fix perf regression for hdim 128 causal il y a 4 semaines
mask.h 51484a7b56 Make backward epilogue work for Sm80 il y a 1 mois
named_barrier.hpp 29cdfedd80 Use Bulk reduce instead of TMA for dQaccum, split across WGs il y a 2 mois
pack_gqa.h 5acb532214 Switch to cutlass v3.6.0, fix perf regression for hdim 128 causal il y a 4 semaines
padding.py 0519920e23 Deal with the case where q or k/v have length 0 il y a 3 semaines
paged_kv.h 5acb532214 Switch to cutlass v3.6.0, fix perf regression for hdim 128 causal il y a 4 semaines
rotary.h 82dc825759 Don't use the unsafe convert_type function il y a 2 mois
seqlen.h 2c996ca25f Use SeqlenInfo for bwd and epilogue il y a 1 mois
setup.py a609d82315 Change extension name to flash_attn_3_cuda il y a 4 semaines
sm90_pipeline_no_cluster.hpp a609d82315 Change extension name to flash_attn_3_cuda il y a 4 semaines
softmax.h 6807b1ea37 Longest-processing-time-first scheduler for causal il y a 1 mois
static_switch.h 42fc4962f0 Uncomment tanh softcapping il y a 2 mois
test_flash_attn.py 0519920e23 Deal with the case where q or k/v have length 0 il y a 3 semaines
test_util.py 0519920e23 Deal with the case where q or k/v have length 0 il y a 3 semaines
tile_scheduler.hpp a901c7eeda Make Sm80 forward pass work with persistent scheduler il y a 1 mois
tile_size.h f907a13187 Tune tile sizes for fwd varlen on Sm80 and Sm86 il y a 4 semaines
utils.h 51484a7b56 Make backward epilogue work for Sm80 il y a 1 mois