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