MCPcopy Create free account

hub / github.com/fla-org/flash-linear-attention / functions

Functions2,989 in github.com/fla-org/flash-linear-attention

Functiontest_parallel
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, dtyp
tests/ops/test_simple_gla.py:618
Functiontest_parallel
( B: int, T: int, H: int, HQ: int, D: int, scale: float, dtype: torch.dtype, )
tests/ops/test_parallax.py:51
Functiontest_parallel
( B: int, T: int, H: int, HQ: int, D: int, scale: float, )
tests/ops/test_forgetting_attn.py:29
Functiontest_parallel
( B: int, T: int, H: int, K: int, expand_ratio: int, dtype: torch.dtype, )
tests/ops/test_retention.py:283
Functiontest_parallel_aggressive_gates_long_seq
Strong per-timestep decay; exact reference stays in fp32, kernel uses per-block R.
tests/ops/test_wall_attn.py:118
Functiontest_parallel_attn_sink_empty_row_matches_reference
()
tests/ops/test_attn_sink.py:243
Functiontest_parallel_attn_sink_matches_reference
(window_size, cu_seqlens)
tests/ops/test_attn_sink.py:175
Functiontest_parallel_attn_sink_with_g_matches_reference
(window_size, cu_seqlens)
tests/ops/test_attn_sink.py:299
Functiontest_parallel_bwd_full_value_reduction
Regression test: the backward must not split the value dim (NV == 1). On low-shared-memory GPUs (e.g. consumer RTX cards) the backward used to ca
tests/ops/test_attn.py:67
Functiontest_parallel_gqa_matches_reference
()
tests/ops/test_wall_attn.py:68
Functiontest_parallel_matches_reference
(B: int, T: int, H: int, HQ: int, K: int, V: int, window_size)
tests/ops/test_wall_attn.py:53
Functiontest_parallel_moba_backward
()
tests/ops/test_moba.py:111
Functiontest_parallel_moba_is_sparse_when_topk_small
()
tests/ops/test_moba.py:154
Functiontest_parallel_moba_matches_full_attn
(T, H, D, chunk_size, topk)
tests/ops/test_moba.py:48
Functiontest_parallel_moba_rejects_batch_gt_one_with_cu_seqlens
()
tests/ops/test_moba.py:138
Functiontest_parallel_moba_topk1_short_circuits
()
tests/ops/test_moba.py:67
Functiontest_parallel_moba_varlen_matches_full_attn
(cu_seqlens, H, D, chunk_size, topk)
tests/ops/test_moba.py:94
Functiontest_parallel_sink
(B: int, T: int, H: int, HQ: int, D: int)
tests/ops/test_attn.py:268
Functiontest_parallel_sink_bias_matches_reference
()
tests/ops/test_wall_attn.py:102
Functiontest_parallel_swa
( B: int, T: int, H: int, HQ: int, D: int, W: int, )
tests/ops/test_attn.py:221
Functiontest_parallel_swa
( B: int, T: int, H: int, HQ: int, D: int, W: int, dtype: torch.dtype, )
tests/ops/test_parallax.py:107
Functiontest_parallel_swa
( B: int, T: int, H: int, HQ: int, D: int, W: int, )
tests/ops/test_forgetting_attn.py:148
Functiontest_parallel_swa_varlen
( H: int, HQ: int, D: int, W: int, cu_seqlens: list[int], )
tests/ops/test_attn.py:312
Functiontest_parallel_swa_varlen
(H: int, HQ: int, D: int, W: int, cu_seqlens: list[int], dtype: torch.dtype)
tests/ops/test_parallax.py:205
Functiontest_parallel_varlen
( H: int, HQ: int, D: int, cu_seqlens: list[int], )
tests/ops/test_attn.py:164
Functiontest_parallel_varlen
( H: int, HQ: int, D: int, use_forget_gate: bool, cu_seqlens: list[int], dtype: torch.
tests/ops/test_path_attn.py:112
Functiontest_parallel_varlen
( H: int, HQ: int, D: int, S: int, block_size: int, cu_seqlens: list[int], dtype:
tests/ops/test_nsa.py:94
Functiontest_parallel_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_simple_gla.py:677
Functiontest_parallel_varlen
(H: int, HQ: int, D: int, cu_seqlens: list[int], dtype: torch.dtype)
tests/ops/test_parallax.py:161
Functiontest_parallel_varlen
( H: int, HQ: int, D: int, cu_seqlens: list[int], )
tests/ops/test_forgetting_attn.py:86
Functiontest_parallel_varlen_matches_reference
()
tests/ops/test_wall_attn.py:84
Functiontest_parallel_with_g
( B: int, T: int, H: int, HQ: int, D: int, scale: float, )
tests/ops/test_attn.py:113
Functiontest_per_layer_independent_counter_regression
Verify that each FLALayer maintains its own independent _seen_tokens counter. This is the mechanism that prevents the GitHub #766 bug.
tests/models/test_cache.py:550
Functiontest_powglu
(B: int, T: int, D: int, power: float, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:310
Functiontest_powglu_linear
(B: int, T: int, D: int, O: int, power: float, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:338
Functiontest_prepare_wy_repr_bwd_no_g
Regression for #862: prepare_wy_repr_bwd previously left b_dk uninitialized and missed the dk-side contribution to db when g is None. With g=
tests/ops/test_gdn.py:1132
Functiontest_public_utils_are_reexported
()
tests/test_utils.py:101
Functiontest_recurrent_fwd
( B: int, T: int, H: int, D: int, scale: float, dtype: torch.dtype, )
tests/ops/test_dplr_delta.py:150
Functiontest_rmsnorm
(B: int, H: int, T: int, D: int)
tests/modules/test_layernorm.py:95
Functiontest_rmsnorm_gated
(B: int, H: int, T: int, D: int, activation: str)
tests/modules/test_layernorm_gated.py:77
Functiontest_rmsnorm_gated_large_batch_offsets
()
tests/modules/test_layernorm_gated.py:102
Functiontest_rmsnorm_gated_large_batch_offsets_large_d
()
tests/modules/test_layernorm_gated.py:126
Functiontest_rmsnorm_linear
(N: int, D: int)
tests/modules/test_layernorm.py:197
Functiontest_rmsnorm_small_t
RMSNorm backward must handle T < SM_count without illegal memory access.
tests/modules/test_layernorm.py:247
Functiontest_rope_offset_consistency_simulation_regression
Simulate how RoPE offsets would be calculated during decode. This test demonstrates that with the fix, all layers compute the same posit
tests/models/test_cache.py:574
Functiontest_rotary
(B: int, T: int, H: int, G: int, D: int, dtype: torch.dtype)
tests/modules/test_rotary.py:21
Functiontest_rotary_varlen
(N: int, T: int, H: int, G: int, D: int, dtype: torch.dtype)
tests/modules/test_rotary.py:91
Functiontest_rotary_with_offsets
(B: int, T: int, H: int, G: int, D: int, dtype: torch.dtype)
tests/modules/test_rotary.py:48
Functiontest_scalar_gate_gradient_finite_differences
dL/dg_scalar for Wall + scalar gate via central differences.
tests/ops/test_wall_attn.py:254
Functiontest_scalar_gate_matches_reference
Wall + FoX-style additive scalar gate: Triton vs reference.
tests/ops/test_wall_attn.py:238
Functiontest_sigmoid
(B: int, T: int, D: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:58
Functiontest_sigmoidglu
(B: int, T: int, D: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:249
Functiontest_sigmoidglu_linear
(B: int, T: int, D: int, O: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:277
Functiontest_simple_gla_to_mamba2
(vary_A, dtype)
tests/ops/test_simple_gla.py:738
Functiontest_single_vs_multi_segment_specializes_once
(name, impl, ref)
tests/ops/utils/test_index.py:187
Functiontest_solve_tril
(B, T, H, chunk_size)
tests/ops/test_solve_tril.py:36
Functiontest_solve_tril_large_batch_offsets
()
tests/ops/test_solve_tril.py:101
Functiontest_solve_tril_varlen
( H: int, D: int, chunk_size: int, cu_seqlens: list[int], )
tests/ops/test_solve_tril.py:74
Functiontest_split_cu_seqlens_correctness
(batch_size, seq_len, split_size)
tests/ops/utils/test_index.py:149
Functiontest_split_namespace_across_sys_path_entries_without_runtime_deps
(tmp_path: Path)
tests/test_split_package_release.py:247
Functiontest_split_wheels_match_release_contract
(tmp_path: Path)
tests/test_split_package_release.py:112
Functiontest_sqrelu
(B: int, T: int, D: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:162
Methodtest_stage1_gk_per_head_sensitivity
Verify stage 1 (dh computation) is sensitive to per-head gk values. The backward pre-processing kernel computes: dh = ex
tests/context_parallel/test_cp_bwd_gk_offset.py:73
Functiontest_swiglu
(B: int, T: int, D: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:188
Functiontest_swiglu_linear
(B: int, T: int, D: int, O: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:216
Functiontest_swish
(B: int, T: int, D: int, noncontiguous: bool, compile: bool)
tests/modules/test_activation.py:110
Functiontest_tensor_cache_uses_identity_until_disabled
(monkeypatch)
tests/test_utils.py:26
Functiontest_token_shift
(B, T, H, cu_seqlens_val, dtype)
tests/modules/test_token_shift.py:33
Functiontest_token_shift_varlen_long_kernel_matches_reference
()
tests/modules/test_token_shift.py:163
Functiontest_token_shift_varlen_uses_max_doc_len_for_short_kernel
()
tests/modules/test_token_shift.py:150
Functiontest_top_level_exports_layers_and_non_config_models
()
tests/test_public_api.py:13
Functiontest_unpack_sequence
( B: int, T: int, H: int, D: int, padding_side: str, dtype: torch.dtype, )
tests/ops/utils/test_pack.py:73
Functiontest_varlen_d_debug
(cu_seqlens, D)
tests/ops/test_attn.py:366
Methodtie_weights
Tie weights for the model. Accepts any arguments for transformers version compatibility.
fla/models/deltaformer/modeling_deltaformer.py:311
Methodto_cpu
(x)
fla/models/utils.py:156
Methodto_dev
(x)
fla/models/utils.py:171
Methodto_dict
(self)
scripts/utils/autotune_export.py:170
Methodto_legacy_cache
(self)
fla/models/utils.py:336
Methodto_legacy_cache
(self)
fla/models/utils.py:444
Functiontoken_shift_bwd_kernel_long
( dx, dy, cu_seqlens, chunk_indices, grad_cache_in, grad_cache_out, T, D: tl.c
fla/modules/token_shift.py:314
Functiontoken_shift_bwd_kernel_short
( dx, dy, cu_seqlens, grad_cache_in, grad_cache_out, T, D: tl.constexpr, BD: t
fla/modules/token_shift.py:236
Functiontoken_shift_fwd_kernel_long
( x, y, cu_seqlens, chunk_indices, cache, cache_out, T, D: tl.constexpr, B
fla/modules/token_shift.py:160
Functiontoken_shift_fwd_kernel_short
( x, y, cu_seqlens, cache, cache_out, T, D: tl.constexpr, BD: tl.constexpr,
fla/modules/token_shift.py:68
Functiontransform_q_fwd_kernel
( q, q_new, w1, w2, cu_seqlens, indices, T, S: tl.constexpr, G: tl.constex
fla/ops/path_attn/transform_q.py:19
Functionwrapped_func
(*args, **kwargs)
fla/utils/_decorators.py:270
Functionwrapper
(*args, **kwargs)
fla/ops/backends/__init__.py:171
Functionwrapper
(*args: Any, **kwargs: Any)
fla/utils/_decorators.py:65
Functionwu_fwd_kernel
( w, u, a, k, v, A, cu_seqlens, chunk_indices, T, H: tl.constexpr,
fla/ops/generalized_delta_rule/iplr/wy_fast.py:166
Functionwu_fwd_kernel
( w, u, ag, v, A_ab_inv, A_ak, cu_seqlens, chunk_indices, T, H: tl.con
fla/ops/generalized_delta_rule/dplr/wy_fast_fwd.py:156
← previous2,901–2,989 of 2,989, ranked by callers