david/flash-attention: flash-attention from https://github.com/Dao-AILab/flash-attention @ 6b1d059eda21c1bd421f3d352786fca2cab61954

Lu Fang 74aed78373 Replace c10::optional with std::optional in flash_attn		преди 3 седмици
..
instantiations	e94f7e89dc Always enable PackGQA is Split to reduce compilation and binary size	преди 4 седмици
__init__.py	7f67966cc7 FA3 initial code release	преди 7 месеца
benchmark_attn.py	5f525322ec Only pass sm_90a compile flag to Sm90 kernels, same w Sm89 kernels	преди 1 месец
benchmark_flash_attention_fp8.py	efbf19cd15 Fix incorrect torch dtype (#1399)	преди 4 седмици
benchmark_split_kv.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	преди 3 месеца
combine.h	478ee666cc Make namespace comment consistent (#1305)	преди 3 месеца
copy_sm90_bulk_reduce.hpp	7a802796e1 Big refactor and update	преди 1 месец
epilogue_bwd.hpp	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец
epilogue_fwd.hpp	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец
flash.h	a84a237d2a Split bwd softcap compilation units for Sm80	преди 4 седмици
flash_api.cpp	74aed78373 Replace c10::optional with std::optional in flash_attn	преди 3 седмици
flash_attn_interface.py	7a802796e1 Big refactor and update	преди 1 месец
flash_bwd_kernel_sm80.h	7a802796e1 Big refactor and update	преди 1 месец
flash_bwd_kernel_sm90.h	7a802796e1 Big refactor and update	преди 1 месец
flash_bwd_launch_template.h	a84a237d2a Split bwd softcap compilation units for Sm80	преди 4 седмици
flash_bwd_postprocess_kernel.h	7a802796e1 Big refactor and update	преди 1 месец
flash_bwd_preprocess_kernel.h	7a802796e1 Big refactor and update	преди 1 месец
flash_fwd_combine.cu	5f525322ec Only pass sm_90a compile flag to Sm90 kernels, same w Sm89 kernels	преди 1 месец
flash_fwd_combine_kernel.h	7a802796e1 Big refactor and update	преди 1 месец
flash_fwd_combine_launch_template.h	7a802796e1 Big refactor and update	преди 1 месец
flash_fwd_kernel_sm80.h	7a802796e1 Big refactor and update	преди 1 месец
flash_fwd_kernel_sm90.h	a93359a2bf If PackGQA, use producer threads instead of Mma threads to load Q	преди 4 седмици
flash_fwd_launch_template.h	1e3208566a Tune tile sizes for compilation	преди 4 седмици
generate_kernels.py	e94f7e89dc Always enable PackGQA is Split to reduce compilation and binary size	преди 4 седмици
heuristics.h	7a802796e1 Big refactor and update	преди 1 месец
mainloop_bwd_sm80.hpp	7a802796e1 Big refactor and update	преди 1 месец
mainloop_bwd_sm90_tma_gmma_ws.hpp	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец
mainloop_fwd_sm80.hpp	84f1287e42 Rename bool_constant<true> to true_type, same w bool_constant<false>	преди 4 седмици
mainloop_fwd_sm90_tma_gmma_ws.hpp	a93359a2bf If PackGQA, use producer threads instead of Mma threads to load Q	преди 4 седмици
mask.h	7a802796e1 Big refactor and update	преди 1 месец
named_barrier.hpp	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец
pack_gqa.h	7a802796e1 Big refactor and update	преди 1 месец
padding.py	7a802796e1 Big refactor and update	преди 1 месец
paged_kv.h	7a802796e1 Big refactor and update	преди 1 месец
rotary.h	7a802796e1 Big refactor and update	преди 1 месец
seqlen.h	7a802796e1 Big refactor and update	преди 1 месец
setup.py	22c0358f4b Fix nvcc_from_env not found	преди 3 седмици
sm90_pipeline_no_cluster.hpp	68bf390920 Update Cutlass to fix mem fence	преди 1 месец
softmax.h	7a802796e1 Big refactor and update	преди 1 месец
static_switch.h	180ff782dd Template for Sm86	преди 1 месец
test_attn_kvcache.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	преди 3 месеца
test_flash_attn.py	2ac6c986be Fix Sm80 tile_count_semaphore, adjust test tolerance	преди 4 седмици
test_kvcache.py	a5a75274bc FA3 kvcache + split kv + gqa parallelization (#1236)	преди 3 месеца
test_util.py	7a802796e1 Big refactor and update	преди 1 месец
tile_scheduler.hpp	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец
tile_size.h	1e3208566a Tune tile sizes for compilation	преди 4 седмици
utils.h	7bc3f031a4 Compile for both Sm80 and Sm90	преди 1 месец