MCPcopy Create free account

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

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

Functiontest_chunk
( B: int, T: int, H: int, D: int, dtype: torch.dtype, gate_logit_normalizer: float, )
tests/ops/test_gla.py:262
Functiontest_chunk
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, dtyp
tests/ops/test_simple_gla.py:258
Functiontest_chunk
( B: int, T: int, H: int, D: int, gate_logit_normalizer: float, dtype: torch.dtype, )
tests/ops/test_rwkv6.py:38
Functiontest_chunk
( B: int, T: int, H: int, HV: int, D: int, scale: float, gate_logit_normalizer: fl
tests/ops/test_gdn.py:104
Functiontest_chunk
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, mask
tests/ops/test_oja.py:321
Functiontest_chunk
( B: int, T: int, H: int, D: int, scale: float, num_householder: int, gate_logit_n
tests/ops/test_gdp.py:39
Functiontest_chunk
( B: int, T: int, H: int, D: int, scale: float, dtype: torch.dtype, )
tests/ops/test_iplr_delta.py:202
Functiontest_chunk
( B: int, T: int, H: int, K: int, expand_ratio: int, dtype: torch.dtype, )
tests/ops/test_retention.py:31
Functiontest_chunk
( B: int, T: int, H: int, D: int, gate_range: tuple[float, float], dtype: torch.dtype,
tests/ops/test_mesa.py:36
Functiontest_chunk_beta_sigmoid_in_kernel
`use_beta_sigmoid_in_kernel=True` (raw beta logits) matches manual sigmoid + autograd.
tests/ops/test_gdn.py:187
Functiontest_chunk_bwd
( B: int, T: int, H: int, D: int, dtype: torch.dtype, )
tests/ops/test_log_linear_attn.py:64
Functiontest_chunk_forward
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, mask
tests/ops/test_oja.py:264
Functiontest_chunk_gate_in_kernel
Test use_gate_in_kernel=True path: fused gate activation + chunk cumsum inside kernel.
tests/ops/test_gdn.py:779
Functiontest_chunk_gate_in_kernel_gqa
Test use_gate_in_kernel=True with grouped value attention (HV > H).
tests/ops/test_gdn.py:881
Functiontest_chunk_gate_in_kernel_varlen
Test use_gate_in_kernel=True with variable-length sequences.
tests/ops/test_gdn.py:982
Functiontest_chunk_gdn_intracard_gqa
E2E: chunk_gated_delta_rule intracard path produces correct results with GQA (Hq < H). Uses a long varlen sequence to exercise the intracard spli
tests/ops/test_intracard_cache.py:131
Functiontest_chunk_kda_intracard_cache_hit_same_cu_seqlens_object
E2E: chunk_kda should reuse intracard precompute cache on second call. This test intentionally uses a very long varlen sequence so that: 1) i
tests/ops/test_intracard_cache.py:30
Functiontest_chunk_matches_fused_recurrent
The two production kernels must agree with each other (and the naive).
tests/ops/test_gdn2.py:393
Functiontest_chunk_offsets_correctness
(batch_size, max_seq_len, chunk_size)
tests/ops/utils/test_index.py:138
Functiontest_chunk_return_intermediate_states
return_intermediate_states yields per-chunk pre-states h; the output must still match the normal path.
tests/ops/test_gdn2.py:407
Functiontest_chunk_return_intermediate_states
Test that return_intermediate_states=True works in inference mode and returns h with correct shape.
tests/ops/test_kda.py:1092
Functiontest_chunk_state_v_first
state_v_first must give the same output and a transposed final state. The state-first flag flips the layout of *both* the initial and final state
tests/ops/test_gdn2.py:304
Functiontest_chunk_state_v_first
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, dtyp
tests/ops/test_kda.py:647
Functiontest_chunk_state_v_first
( B: int, T: int, H: int, D: int, dtype: torch.dtype, )
tests/ops/test_gla.py:334
Functiontest_chunk_state_v_first
( B: int, T: int, H: int, D: int, dtype: torch.dtype, )
tests/ops/test_simple_gla.py:329
Functiontest_chunk_state_v_first
( B: int, T: int, H: int, D: int, scale: float, gate_logit_normalizer: float, dtyp
tests/ops/test_gdn.py:271
Functiontest_chunk_use_beta_sigmoid_in_kernel
( B: int, T: int, H: int, D: int, allow_neg_eigval: bool, dtype: torch.dtype, )
tests/ops/test_kda.py:722
Functiontest_chunk_varlen
Packed varlen chunk run (fwd + grads) must equal per-sequence reference.
tests/ops/test_gdn2.py:337
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_ttt.py:211
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_log_linear_attn.py:117
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_delta.py:105
Functiontest_chunk_varlen
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_comba.py:223
Functiontest_chunk_varlen
( H: int, D: int, num_householder: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_delta_product.py:105
Functiontest_chunk_varlen
( H: int, D: int, mask_p: float, gate_logit_normalizer: float, safe_gate: bool, cu_seq
tests/ops/test_dplr_delta.py:386
Functiontest_chunk_varlen
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, use_gate_i
tests/ops/test_kda.py:824
Functiontest_chunk_varlen
( H: int, D: int, M: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_gsa.py:311
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_gla.py:394
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_simple_gla.py:389
Functiontest_chunk_varlen
( H: int, D: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_rwkv6.py:125
Functiontest_chunk_varlen
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_gdn.py:614
Functiontest_chunk_varlen
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_oja.py:403
Functiontest_chunk_varlen
( H: int, D: int, num_householder: int, mask_p: float, cu_seqlens: list[int], dtype: t
tests/ops/test_gdp.py:121
Functiontest_chunk_varlen
( H: int, K: int, expand_ratio: int, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_retention.py:84
Functiontest_chunk_varlen
( H: int, D: int, gate_range: tuple[float, float], cu_seqlens: list[int], dtype: torch.dty
tests/ops/test_mesa.py:125
Functiontest_chunk_varlen_prefill
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, use_gate_i
tests/ops/test_kda.py:944
Functiontest_chunk_varlen_prefill
( H: int, D: int, mask_p: float, cu_seqlens: list[int], dtype: torch.dtype, )
tests/ops/test_gdn.py:705
Functiontest_conv
( B: int, T: int, D: int, W: int, activation: str, has_bias: bool, has_residual: b
tests/modules/test_conv.py:139
Functiontest_conv_cache_backward
( B: int, T: int, D: int, W: int, has_bias: bool, has_residual: bool, activation:
tests/modules/test_conv.py:889
Functiontest_conv_cache_backward_no_final_state
Test backward with initial_state but WITHOUT output_final_state. This tests the 'if not USE_FINAL_STATE' branch in causal_conv1d_bwd_kernel,
tests/modules/test_conv.py:1050
Functiontest_conv_decoding
( B: int, T: int, D: int, W: int, activation: str, has_bias: bool, has_residual: b
tests/modules/test_conv.py:292
Functiontest_conv_decoding_non_contiguous_x
Test decoding with non-contiguous input x.
tests/modules/test_conv.py:720
Functiontest_conv_decoding_with_cache
( B: int, D: int, W: int, activation: str, has_bias: bool, has_residual: bool, dty
tests/modules/test_conv.py:536
Functiontest_conv_non_contiguous_dy
Test that backward produces correct gradients when dy is non-contiguous. This simulates the split -> conv -> cat pattern used in fused-QKV attent
tests/modules/test_conv.py:1405
Functiontest_conv_non_contiguous_qkv
Test non-contiguous input from QKV concatenated tensor (non-varlen mode).
tests/modules/test_conv.py:1141
Functiontest_conv_varlen
( N: int, T: int, D: int, W: int, activation: str, has_bias: bool, has_residual: b
tests/modules/test_conv.py:211
Functiontest_conv_varlen_decoding
Test varlen mode decoding with causal_conv1d_update.
tests/modules/test_conv.py:596
Functiontest_conv_varlen_initial_state_backward_random
()
tests/modules/test_conv.py:959
Functiontest_conv_varlen_non_contiguous_qkv
Test non-contiguous input from QKV concatenated tensor (varlen mode).
tests/modules/test_conv.py:1256
Functiontest_conv_varlen_with_cache_prefill_fwd
( N: int, T: int, D: int, W: int, activation: str, has_bias: bool, has_residual: b
tests/modules/test_conv.py:440
Functiontest_conv_with_cache_prefill_fwd
( B: int, T: int, D: int, W: int, activation: str, has_bias: bool, has_residual: b
tests/modules/test_conv.py:365
Functiontest_core_only_import_with_repo_root_on_path
fla-core installed alone should still be importable from the repo root. This guards against the case where the source tree (containing fla/layers
tests/test_split_package_release.py:203
Functiontest_core_then_extension_install_sequence_without_runtime_deps
(tmp_path: Path)
tests/test_split_package_release.py:150
Functiontest_cp2_boundary_aligned
CP2: sequence boundaries aligned with rank boundaries.
tests/context_parallel/test_cp_token_shift.py:209
Functiontest_cp2_boundary_aligned
CP2: sequence boundaries aligned with rank boundaries.
tests/context_parallel/test_cp_dplr.py:371
Functiontest_cp2_boundary_aligned
Test Case 2: CP2 with sequence boundaries aligned with rank boundaries. Scenario: - world_size=2, T=1024, chunk_size=512 - lengths=[
tests/context_parallel/test_cp_conv.py:279
Functiontest_cp2_boundary_aligned
CP2: sequence boundaries aligned with rank boundaries.
tests/context_parallel/test_cp_kda.py:463
Functiontest_cp2_boundary_aligned
CP2: sequence boundaries aligned with rank boundaries.
tests/context_parallel/test_cp_gdn.py:328
Functiontest_cp2_disable_recompute
CP2: disable_recompute=True.
tests/context_parallel/test_cp_kda.py:538
Functiontest_cp2_global_len1_sequence
CP2: a globally length-1 sequence sits right at the rank boundary. lengths=[512, 1, 511] → seq1 is globally length 1, entirely on rank 1
tests/context_parallel/test_cp_conv.py:480
Functiontest_cp2_gqa_sequence_cut
CP2 GQA: sequences cut across rank boundary, Hq < H.
tests/context_parallel/test_cp_gdn.py:398
Functiontest_cp2_gqa_single_sequence
CP2 GQA: single long sequence with Hq < H.
tests/context_parallel/test_cp_gdn.py:412
Functiontest_cp2_many_short_sequences
CP2: many short sequences.
tests/context_parallel/test_cp_token_shift.py:251
Functiontest_cp2_many_short_sequences
CP2: many short sequences.
tests/context_parallel/test_cp_dplr.py:413
Functiontest_cp2_many_short_sequences
Test Case 5: CP2 with many short sequences. Scenario: - world_size=2, T=1024, chunk_size=512 - lengths=[100, 150, 200, 250, 124, 100
tests/context_parallel/test_cp_conv.py:353
Functiontest_cp2_many_short_sequences
CP2: many short sequences.
tests/context_parallel/test_cp_kda.py:523
Functiontest_cp2_many_short_sequences
CP2: many short sequences.
tests/context_parallel/test_cp_gdn.py:384
Functiontest_cp2_multiple_short_tails
CP2: multiple sequences each end 1 token into rank 1. lengths=[200, 313, 511] → seq0 ends at 200 (rank 0), seq1 ends at 513 (rank 1 gets 1 t
tests/context_parallel/test_cp_conv.py:457
Functiontest_cp2_safe_gate
CP2: safe_gate=True with sequence cut.
tests/context_parallel/test_cp_dplr.py:427
Functiontest_cp2_sequence_cut
CP2: sequences cut across rank boundary.
tests/context_parallel/test_cp_token_shift.py:195
Functiontest_cp2_sequence_cut
CP2: sequences cut across rank boundary.
tests/context_parallel/test_cp_dplr.py:357
Functiontest_cp2_sequence_cut
Test Case 1: CP2 with sequences cut in the middle. Scenario: - world_size=2, T=1024, chunk_size=512 - lengths=[300, 400, 324] -> seq
tests/context_parallel/test_cp_conv.py:255
Functiontest_cp2_sequence_cut
CP2: sequences cut across rank boundary.
tests/context_parallel/test_cp_kda.py:448
Functiontest_cp2_sequence_cut
CP2: sequences cut across rank boundary.
tests/context_parallel/test_cp_gdn.py:314
Functiontest_cp2_short_tail_len1
CP2: rank 1 gets a length-1 tail.
tests/context_parallel/test_cp_token_shift.py:266
Functiontest_cp2_short_tail_len1
CP2: seq0 has 513 tokens, so rank 1 gets a length-1 tail (T=1 < W=4). Rank 0: [0, 512) → 512 tokens of seq0 Rank 1: [512, 1024) → 1 tok
tests/context_parallel/test_cp_conv.py:387
Functiontest_cp2_short_tail_len2
CP2: rank 1 gets a length-2 tail.
tests/context_parallel/test_cp_token_shift.py:280
Functiontest_cp2_short_tail_len2
CP2: seq0 has 514 tokens, so rank 1 gets a length-2 tail (T=2 < W=4). Rank 1: [512, 1024) → 2 tokens of seq0 (T_local=2!) + 510 tokens of se
tests/context_parallel/test_cp_conv.py:409
Functiontest_cp2_state_v_first
CP2: state_v_first=True with sequence cut.
tests/context_parallel/test_cp_kda.py:558
Functiontest_cp2_state_v_first
CP2: state_v_first=True with sequence cut.
tests/context_parallel/test_cp_gdn.py:430
Functiontest_cp4_complex
CP4: complex sequence distribution.
tests/context_parallel/test_cp_token_shift.py:223
Functiontest_cp4_complex
CP4: complex sequence distribution, first sequence spans 3 ranks.
tests/context_parallel/test_cp_dplr.py:385
Functiontest_cp4_complex
Test Case 3: CP4 with complex sequence distribution. Scenario: - world_size=4, T=1024, chunk_size=256 - lengths=[700, 324] -> first
tests/context_parallel/test_cp_conv.py:304
Functiontest_cp4_complex
CP4: complex sequence distribution, first sequence spans 3 ranks.
tests/context_parallel/test_cp_kda.py:478
Functiontest_cp4_complex
CP4: complex sequence distribution, first sequence spans 3 ranks.
tests/context_parallel/test_cp_gdn.py:342
Functiontest_cp4_every_rank_gets_short_tail
CP4: every non-first rank gets a length-1 local sequence tail. lengths=[257, 255, 257, 255], T=1024, chunk_size=256 Rank 0: [0, 256) →
tests/context_parallel/test_cp_conv.py:430
Functiontest_cp4_multiple_short_tails
CP4: multiple sequences each end 1-2 tokens past a rank boundary, creating short local segments on multiple ranks. lengths=[257, 253, 25
tests/context_parallel/test_cp_conv.py:505
Functiontest_cp4_single_sequence
CP4: single long sequence spanning all ranks.
tests/context_parallel/test_cp_token_shift.py:237
Functiontest_cp4_single_sequence
CP4: single long sequence spanning all ranks.
tests/context_parallel/test_cp_dplr.py:399
Functiontest_cp4_single_sequence
Test Case 4: CP4 with a single long sequence spanning all ranks. Scenario: - world_size=4, T=1024, chunk_size=256 - lengths=[1024] -
tests/context_parallel/test_cp_conv.py:330
Functiontest_cp4_single_sequence
CP4: single long sequence spanning all ranks.
tests/context_parallel/test_cp_kda.py:493
← previousnext →2,601–2,700 of 2,989, ranked by callers