MCPcopy Create free account

hub / github.com/deepseek-ai/FlashMLA / functions

Functions363 in github.com/deepseek-ai/FlashMLA

Functionref_sparse_attn_fwd
Returns: - o: [s_q, h_q, dv] - o_fp32: [s_q, h_q, dv] - max_logits: [s_q, h_q] - lse: [s_q, h_q]
tests/ref.py:19
Functionrescale_O
csrc/sm100/prefill/sparse/common_subroutine.h:146
Methodrestage
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:219
Methodrestage
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:218
Functionretrieve_mask_and_reduce_p
csrc/sm100/prefill/sparse/common_subroutine.h:74
Methodrun
Primary run() entry point API that is static allowing users to create and manage their own params. Supplied params struct must be construct by calling
csrc/sm100/prefill/dense/device/fmha.hpp:205
Methodrun
Primary run() entry point API that is static allowing users to create and manage their own params. Supplied params struct must be construct by calling
csrc/sm100/prefill/dense/device/fmha_device_bwd.hpp:286
Methodrun_
csrc/api/sparse_decode.h:69
Methodrun_
csrc/api/sparse_decode.h:102
Methodrun_
csrc/api/sparse_decode.h:136
Methodrun_
csrc/api/sparse_decode.h:178
Methodrun_
csrc/api/sparse_fwd.h:41
Methodrun_
csrc/api/sparse_fwd.h:61
Methodrun_
csrc/api/sparse_fwd.h:79
Methodrun_
csrc/api/sparse_fwd.h:96
Functionrun_flash_infer
(q, block_table, blocked_k, max_seqlen_pad, block_size, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype
benchmark/bench_flash_mla.py:82
Functionrun_flash_mla
(q, block_table, blocked_k, max_seqlen_pad, block_size, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype
benchmark/bench_flash_mla.py:63
Functionrun_flash_mla_decode
(p: TestParam, t: TestcaseForDecode, tile_scheduler_metadata, num_splits)
tests/lib.py:319
Functionrun_flash_mla_sparse_fwd
(p: TestParam, t: Testcase, return_p_sum: bool)
tests/lib.py:310
Functionrun_flash_mla_triton
(q, block_table, blocked_k, max_seqlen_pad, block_size, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype
benchmark/bench_flash_mla.py:381
Functionrun_torch_mla
(q, block_table, blocked_k, max_seqlen_pad, block_size, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype
benchmark/bench_flash_mla.py:36
Methodsafe_div
csrc/sm100/prefill/dense/common/gather_tensor.hpp:121
Methodsave_rPb_to_sP
csrc/sm90/decode/sparse_fp8/config.h:159
Methodsave_rS_to_sS
csrc/sm90/prefill/sparse/config.h:123
Functionsdpa
(query, key, value, attn_bias, softmax_scale=None)
tests/test_fmha_sm100.py:31
Functionset_random_seed
(seed: int)
tests/kernelkit/utils.py:33
Methodsoftmax
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:735
Methodsoftmax_step
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp:514
Methodsoftmax_step
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:532
Functionsparse_attn_decode_interface
csrc/api/sparse_decode.h:183
Functionsparse_attn_prefill_interface
csrc/api/sparse_fwd.h:101
Functionst_async_128b
csrc/sm90/decode/sparse_fp8/components/helpers.h:78
Methodstore
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_epilogue_tma_warpspecialized.hpp:155
Methodstore
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:951
Methodstore_o
csrc/sm90/decode/sparse_fp8/config.h:183
Methodsync_all_threads_in_cluster
csrc/sm90/decode/sparse_fp8/config.h:145
Methodto_bwd_arguments
csrc/sm100/prefill/dense/device/fmha_device_bwd.hpp:172
Methodto_convert_arguments
csrc/sm100/prefill/dense/device/fmha_device_bwd.hpp:153
Methodto_sum_OdO_arguments
csrc/sm100/prefill/dense/device/fmha_device_bwd.hpp:130
Methodto_underlying_arguments
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp:217
Methodto_underlying_arguments
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_epilogue_tma_warpspecialized.hpp:107
Methodto_underlying_arguments
csrc/sm100/prefill/dense/collective/sm100_fmha_load_tma_warpspecialized.hpp:87
Methodto_underlying_arguments
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:235
Methodto_underlying_arguments
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_load_tma_warpspecialized.hpp:94
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp:66
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp:143
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:372
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:241
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_convert.hpp:102
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:365
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_sum_OdO.hpp:103
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_tile_scheduler.hpp:56
Methodto_underlying_arguments
csrc/sm100/prefill/dense/kernel/fmha_tile_scheduler.hpp:106
Functionupcast
csrc/sm100/prefill/dense/common/gather_tensor.hpp:175
Methodupdate
csrc/utils.h:62
Methodupdate
Update API is preserved in 3.0, but does not guarantee a lightweight update of params.
csrc/sm100/prefill/dense/device/fmha.hpp:190
Methodwarp_idx_to_WarpRole
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:64
Methodwarp_idx_to_WarpRole
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:103
Methodwarp_idx_to_WarpRole
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:143
Methodwarp_idx_to_role
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:95
Methodwarp_idx_to_role
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:94
Methodwhat
csrc/kerutils/include/kerutils/host/host.h:30
Method~ImplBase
csrc/api/common.h:173
← previous301–363 of 363, ranked by callers