Tri Dao 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
..
instantiations fc2fd95a18 Renable FP8 kernels 3 недель назад
__init__.py 7f67966cc7 FA3 initial code release 5 месяцев назад
benchmark_attn.py 82c1aa3514 Move PackGQA epilogue code to pack_gqa.h 1 месяц назад
benchmark_flash_attention_fp8.py c92ca63268 FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173) 3 месяцев назад
copy_sm90_bulk_reduce.hpp 29cdfedd80 Use Bulk reduce instead of TMA for dQaccum, split across WGs 1 неделя назад
epilogue_bwd_sm90_tma.hpp 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
epilogue_fwd_sm90_tma.hpp 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
flash.h fb9c9cbbe9 Support qkv_descale of shape (batch_size, nheads_kv) 3 дней назад
flash_api.cpp 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
flash_attn_interface.py 0c49ac9a07 Implement rotary non-interleaved 3 недель назад
flash_bwd_kernel.h 29cdfedd80 Use Bulk reduce instead of TMA for dQaccum, split across WGs 1 неделя назад
flash_bwd_launch_template.h 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
flash_bwd_postprocess_kernel.h 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
flash_bwd_preprocess_kernel.h 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
flash_fwd_combine_kernel.h 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
flash_fwd_combine_launch_template.h 9fd6b977bb Precompute the pointers in mha_combine kernel 3 недель назад
flash_fwd_combine_sm80.cu 9fd6b977bb Precompute the pointers in mha_combine kernel 3 недель назад
flash_fwd_kernel.h fb9c9cbbe9 Support qkv_descale of shape (batch_size, nheads_kv) 3 дней назад
flash_fwd_launch_template.h 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
generate_kernels.py fc2fd95a18 Renable FP8 kernels 3 недель назад
heuristics.h 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
mainloop_bwd_sm90_tma_gmma_ws.hpp ae3c1fb3e0 Simplify bwd by setting NumdQWarpGroups = NumMmaWarpGroups 1 неделя назад
mainloop_fwd_sm90_tma_gmma_ws.hpp fb9c9cbbe9 Support qkv_descale of shape (batch_size, nheads_kv) 3 дней назад
mask.h 3b6ac2b954 Use compile time constants in local mask 1 неделя назад
named_barrier.hpp 29cdfedd80 Use Bulk reduce instead of TMA for dQaccum, split across WGs 1 неделя назад
pack_gqa.h fe412d6b36 Redo rotary when contiguous 3 недель назад
paged_kv.h 94657af3e8 Add option for not doing intra-WG overlapping of gemm and softmax 3 недель назад
rotary.h 82dc825759 Don't use the unsafe convert_type function 2 недель назад
seqlen.h 2c996ca25f Use SeqlenInfo for bwd and epilogue 1 неделя назад
setup.py 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
softmax.h 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
static_switch.h 42fc4962f0 Uncomment tanh softcapping 2 недель назад
test_flash_attn.py 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
tile_scheduler.hpp 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
tile_size.h 6807b1ea37 Longest-processing-time-first scheduler for causal 21 часов назад
utils.h e8a1edbeb2 Clean up some #include 2 недель назад