Tri Dao 4f0640d534 Move writing P to smem as separate function hai 1 día
..
instantiations 8668823635 Split hdimdiff into a separate translation unit hai 3 semanas
__init__.py 7f67966cc7 FA3 initial code release hai 7 meses
benchmark_attn.py 5458c78e6d Remove sink token hai 4 días
benchmark_flash_attention_fp8.py efbf19cd15 Fix incorrect torch dtype (#1399) hai 1 mes
benchmark_mla_decode.py 08f4c802c4 Add FLOPS to MLA decode benchmark hai 5 días
benchmark_split_kv.py ea3ecea97a Add tp_degree to benchmark_split_kv hai 2 semanas
block.h 6752d62aa4 Add dynamic splits hai 1 semana
copy_sm90_bulk_reduce.hpp 7a802796e1 Big refactor and update hai 1 mes
epilogue_bwd.hpp 6752d62aa4 Add dynamic splits hai 1 semana
epilogue_fwd.hpp 6752d62aa4 Add dynamic splits hai 1 semana
flash.h 5458c78e6d Remove sink token hai 4 días
flash_api.cpp 45c48afb2b Add option for WG1 to use RS MMA but WG2 using SS MMA hai 2 días
flash_attn_interface.py 5458c78e6d Remove sink token hai 4 días
flash_bwd_kernel_sm80.h 15cf7ee435 Rename collective_mainloop -> mainloop, move tile_scheduler variable hai 3 semanas
flash_bwd_kernel_sm90.h 15cf7ee435 Rename collective_mainloop -> mainloop, move tile_scheduler variable hai 3 semanas
flash_bwd_launch_template.h 5458c78e6d Remove sink token hai 4 días
flash_bwd_postprocess_kernel.h 7a802796e1 Big refactor and update hai 1 mes
flash_bwd_preprocess_kernel.h 7a802796e1 Big refactor and update hai 1 mes
flash_fwd_combine.cu 5378bc3204 Tile fwd_combine kernel along headdim, don't need kBlockM > 128 hai 3 semanas
flash_fwd_combine_kernel.h 6752d62aa4 Add dynamic splits hai 1 semana
flash_fwd_combine_launch_template.h 6752d62aa4 Add dynamic splits hai 1 semana
flash_fwd_kernel_sm80.h 15cf7ee435 Rename collective_mainloop -> mainloop, move tile_scheduler variable hai 3 semanas
flash_fwd_kernel_sm90.h 15cf7ee435 Rename collective_mainloop -> mainloop, move tile_scheduler variable hai 3 semanas
flash_fwd_launch_template.h 5458c78e6d Remove sink token hai 4 días
flash_prepare_scheduler.cu 20b84d6363 Don't use IntraWGOverlap for hdim 64,512 hai 4 días
generate_kernels.py 8668823635 Split hdimdiff into a separate translation unit hai 3 semanas
heuristics.h 6752d62aa4 Add dynamic splits hai 1 semana
mainloop_bwd_sm80.hpp 5458c78e6d Remove sink token hai 4 días
mainloop_bwd_sm90_tma_gmma_ws.hpp 5458c78e6d Remove sink token hai 4 días
mainloop_fwd_sm80.hpp 5458c78e6d Remove sink token hai 4 días
mainloop_fwd_sm90_tma_gmma_ws.hpp 4f0640d534 Move writing P to smem as separate function hai 1 día
mask.h 7a802796e1 Big refactor and update hai 1 mes
named_barrier.hpp f0f25239bd Implement the case of LargeHeadDimV hai 1 mes
pack_gqa.h 7a802796e1 Big refactor and update hai 1 mes
padding.py 7a802796e1 Big refactor and update hai 1 mes
paged_kv.h e3b2400a31 Fix loading paged V when kHeadDimV != kHeadDim hai 1 mes
rotary.h 7a802796e1 Big refactor and update hai 1 mes
seqlen.h 7a802796e1 Big refactor and update hai 1 mes
setup.py 3edf7e0daa Add kwargs to _write_ninja_file for compatibility with new torch hai 1 día
sm90_pipeline_no_cluster.hpp 68bf390920 Update Cutlass to fix mem fence hai 1 mes
softmax.h 7a802796e1 Big refactor and update hai 1 mes
static_switch.h 180ff782dd Template for Sm86 hai 1 mes
test_attn_kvcache.py a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236) hai 4 meses
test_flash_attn.py 5458c78e6d Remove sink token hai 4 días
test_kvcache.py a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236) hai 4 meses
test_util.py fa445ff6c2 Fix FP8 test hai 3 semanas
tile_scheduler.hpp 20b84d6363 Don't use IntraWGOverlap for hdim 64,512 hai 4 días
tile_size.h 20b84d6363 Don't use IntraWGOverlap for hdim 64,512 hai 4 días
utils.h 45c48afb2b Add option for WG1 to use RS MMA but WG2 using SS MMA hai 2 días