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