MCPcopy Create free account

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

Functions363 in github.com/deepseek-ai/FlashMLA

↓ 1 callersFunctionscaled_dot_product_attention
(query, key, value, h_q, h_kv, is_causal=False)
benchmark/bench_flash_mla.py:15
↓ 1 callersFunctionscaled_dot_product_attention
( batch_idx: int, query: torch.Tensor, # [h_q, s_q, d] kv: torch.Tensor, # [h_
tests/test_flash_mla_dense_decoding.py:85
↓ 1 callersFunctionsdpa_checkpoint
(*args, **kwargs)
tests/test_fmha_sm100.py:46
↓ 1 callersFunctiontest_flash_attention
(b, mean_sq, mean_sk, varlen, h, h_k, d, dv, causal, window, has_bwd, check_correctness: bool = True)
tests/test_fmha_sm100.py:50
↓ 1 callersFunctiontest_flash_mla
(p: TestParam)
tests/test_flash_mla_sparse_decoding.py:142
↓ 1 callersFunctiontest_flash_mla
(t: TestParam)
tests/test_flash_mla_dense_decoding.py:145
↓ 1 callersMethodto_test_param
(self)
tests/lib.py:74
↓ 1 callersFunctiontorch_attn
()
tests/test_fmha_sm100.py:98
MethodArch
csrc/api/common.h:27
MethodCUTE_DEVICE Pow2
csrc/sm100/prefill/dense/common/pow_2.hpp:45
MethodCausalIndividualTileScheduler
csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp:62
MethodIndexedGather
csrc/sm100/prefill/dense/common/gather_tensor.hpp:52
MethodIndividualTileScheduler
csrc/sm100/prefill/dense/kernel/fmha_tile_scheduler.hpp:52
MethodKUException
csrc/kerutils/include/kerutils/host/host.h:22
MethodNoGather
csrc/sm100/prefill/dense/common/gather_tensor.hpp:45
FunctionPYBIND11_MODULE
csrc/api/api.cpp:8
MethodSm100FmhaFwdEpilogueTmaWarpspecialized
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_epilogue_tma_warpspecialized.hpp:152
MethodStridedGather
csrc/sm100/prefill/dense/common/gather_tensor.hpp:74
Method__enter__
(self)
tests/kernelkit/bench.py:10
Method__enter__
(self)
tests/kernelkit/precision.py:9
Method__exit__
(self, *_)
tests/kernelkit/bench.py:13
Method__exit__
(self, exc_type, exc_value, traceback)
tests/kernelkit/precision.py:13
Method__init__
(self)
tests/kernelkit/utils.py:45
Method__init__
(self, enabled: bool = True)
tests/kernelkit/precision.py:6
Function_check_optional_tensor
csrc/kerutils/include/kerutils/supplemental/torch_tensors.h:15
Function_mla_attn_kernel
( Q_nope, Q_pe, Kv_c_cache, K_pe_cache, Req_to_tokens, B_seq_len, O, sm_scale,
benchmark/bench_flash_mla.py:136
Function_mla_softmax_reducev_kernel
( Logits, B_seq_len, O, stride_l_b, stride_l_h, stride_l_s, stride_o_b, stride
benchmark/bench_flash_mla.py:274
Functionabs_indices2indices_in_kvcache
Convert abs_indices (logical index, ranging from 0 to s_k-1) to index expected by the sparse attn kernel Equivalent to: b, s_q, topk
tests/quant.py:126
Methodapply_batch
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:250
Methodapply_mask
csrc/sm100/prefill/dense/collective/fmha_fusion.hpp:114
Methodapply_mask
csrc/sm100/prefill/dense/collective/fmha_fusion.hpp:166
Methodapply_mask
csrc/sm100/prefill/dense/collective/fmha_fusion.hpp:244
Methodapply_mask
csrc/sm100/prefill/dense/collective/fmha_fusion.hpp:285
Methodapply_perm
Apply a batch permutation to this KVScope. Used for batch-invariance test
tests/lib.py:202
Functionbench_by_cuda_events
(kernels: List[Callable], num_warmups_each: int, num_runs_each: int)
tests/kernelkit/bench.py:161
Functionbench_kineto
Run `fn` for `num_tests` times under `bench_kineto` (CUPTI), and returns a BenchKinetoRawResult
tests/kernelkit/bench.py:103
Methodbfloat16_t>
csrc/api/common.h:16
Methodcan_implement
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp:212
Methodcan_implement
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:230
Methodcan_implement
Determines whether the GEMM can execute the given problem.
csrc/sm100/prefill/dense/device/fmha.hpp:86
Methodcan_implement
Determines whether the GEMM can execute the given problem.
csrc/sm100/prefill/dense/device/fmha_device_bwd.hpp:197
Methodcan_implement
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:354
Methodcan_implement
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:228
Methodcan_implement
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_convert.hpp:88
Methodcan_implement
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:347
Methodcan_implement
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_sum_OdO.hpp:89
Functioncdiv
(a: int, b: int)
tests/kernelkit/utils.py:18
Functionceil
csrc/kerutils/include/kerutils/host/host.h:77
Functionceil_div
csrc/utils.h:37
Methodcheck_if_all_features_are_supported_and_abort
csrc/api/common.h:192
Functioncheck_is_allclose_comparator
(name: str, ans: torch.Tensor, ref: torch.Tensor, out: torch.Tensor, abs_tol: float = 1e-5, rel_tol: float = 1
tests/kernelkit/compare.py:94
Functioncheck_is_bitwise_equal
(name: str, ans: torch.Tensor, ref: torch.Tensor, quiet: bool = False)
tests/kernelkit/compare.py:13
Functioncheck_is_bitwise_equal_comparator
Return if two tensors are bitwise equal Return a bool if avoid_sync is False, else return a tensor
tests/kernelkit/compare.py:5
Methodcompute
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:1125
Methodcompute
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:1119
Functionconsumer_release
csrc/sm100/prefill/dense/common/pipeline_mla.hpp:213
Functionconsumer_wait
csrc/sm100/prefill/dense/common/pipeline_mla.hpp:208
Methodcopy
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_convert.hpp:107
Methodcorrection
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:962
Methodcorrection_empty
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:1149
Methodcorrection_epilogue
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp:778
Methodcorrection_epilogue
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:786
Methodcorrection_rescale
csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp:868
Methodcorrection_rescale
csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp:876
Functioncount_flop_and_mem_vol
(p: TestParam, t: Testcase)
tests/lib.py:345
Functioncount_flop_and_mem_vol_for_decode
(p: TestParam, t: TestcaseForDecode)
tests/lib.py:367
Functioncp_async_bulk_shared_cta_shared_cluster
csrc/sm90/decode/sparse_fp8/components/helpers.h:90
Functioncp_async_cacheglobal_l2_prefetch_256B
csrc/sm90/helpers.h:8
Functioncreatepolicy_evict_first
csrc/sm90/helpers.h:35
Functioncreatepolicy_evict_last
csrc/sm90/helpers.h:25
Functioncvt_fp8x8_bf16x8
csrc/sm90/decode/sparse_fp8/components/dequant.h:20
Functiondense_attn_decode_interface
csrc/api/dense_decode.h:13
Functiondequantize_k_cache
De-quantize the k-cache
tests/quant.py:81
Methodepilogue
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:1021
Methodepilogue
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:1015
Methodepilogue_clear
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:976
Methodepilogue_clear
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:971
Functionflash_attn_varlen_kvpacked_func
( q: torch.Tensor, kv: torch.Tensor, cu_seqlens_qo: torch.Tensor, cu_seqlens_kv: torch.Tensor,
flash_mla/flash_mla_interface.py:415
Functionflash_attn_varlen_qkvpacked_func
( qkv: torch.Tensor, cu_seqlens: torch.Tensor, max_seqlen: int, head_dim_qk: int, dropout_
flash_mla/flash_mla_interface.py:395
Functionflash_mla_sparse_fwd
Sparse attention prefill kernel Args: q: [s_q, h_q, d_qk], bfloat16 kv: [s_kv, h_kv, d_qk], bfloat16 indices: [s_q,
flash_mla/flash_mla_interface.py:176
Methodforward
( ctx, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, cu_seqlens_q
flash_mla/flash_mla_interface.py:329
Functiongemm
csrc/sm90/decode/sparse_fp8/components/helpers.h:22
Functiongemm_rs
csrc/sm90/helpers.h:116
Functiongemm_ss
csrc/sm90/helpers.h:97
Functiongen_non_contiguous_randn_tensor
(shape, *args, **kwargs)
tests/kernelkit/generate.py:10
Functiongenerate_testcase
(t: TestParam)
tests/lib.py:121
Functiongenerate_testcase_for_decode
(t: TestParam)
tests/lib.py:227
Functionget_AorC_col_idx
csrc/sm90/helpers.h:53
Functionget_AorC_row_idx
csrc/sm90/helpers.h:46
Functionget_AorC_row_idx
In the layout of fragment A and fragment C during WGMMA, data each thread holds resides in two particular rows. This function converts the local_row_i
csrc/sm90/decode/sparse_fp8/components/helpers.h:14
Methodget_block_coord
csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp:87
Methodget_block_coord
csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp:178
Methodget_block_coord
csrc/sm100/prefill/dense/kernel/fmha_tile_scheduler.hpp:141
Methodget_block_shape
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_mla_kernel_tma_warpspecialized.hpp:1816
Methodget_block_shape
csrc/sm100/prefill/dense/kernel/sm100_fmha_fwd_kernel_tma_warpspecialized.hpp:236
Methodget_block_shape
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_convert.hpp:97
Methodget_block_shape
csrc/sm100/prefill/dense/kernel/sm100_fmha_bwd_kernel_tma_warpspecialized.hpp:1822
Methodget_block_shape
csrc/sm100/prefill/dense/kernel/fmha_kernel_bwd_sum_OdO.hpp:98
Functionget_enum_max
csrc/api/common.h:124
Methodget_grid_shape
Computes the grid shape
csrc/sm100/prefill/dense/device/fmha.hpp:105
← previousnext →101–200 of 363, ranked by callers