Coverage for  / home / jenkins / .local / lib / python3.10 / site-packages / hyper_parallel / components / functional / dsa_indexer.py: 0%

1# Copyright 2026 Huawei Technologies Co., Ltd 

2# 

3# Licensed under the Apache License, Version 2.0 (the "License"); 

4# you may not use this file except in compliance with the License. 

5# You may obtain a copy of the License at 

6# 

7# http://www.apache.org/licenses/LICENSE-2.0 

8# 

9# Unless required by applicable law or agreed to in writing, software 

10# distributed under the License is distributed on an "AS IS" BASIS, 

11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 

12# See the License for the specific language governing permissions and 

13# limitations under the License. 

14# ============================================================================ 

15"""DSA sparse-attention indexer.""" 

16 

17from typing import Tuple 

18 

19import torch # pylint: disable=forbidden-backend-import 

20from einops import rearrange 

21import omni_training_custom_ops # noqa: F401 # pylint: disable=unused-import 

22 

23 

24def dsa_indexer( 

25 index_query: torch.Tensor, 

26 index_key: torch.Tensor, 

27 merge_weight: torch.Tensor, 

28 actual_seq_qlen: torch.Tensor, 

29 actual_seq_kvlen: torch.Tensor, 

30 sparse_count: int, 

31) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: 

32 """Generate top-k token indices for DSA sparse attention. 

33 

34 Args: 

35 index_query: Indexer query in BSND layout. 

36 index_key: Indexer key in BSND layout. 

37 merge_weight: Per-head indexer weights in BSN layout. 

38 actual_seq_qlen: Cumulative query lengths as an int32 NPU tensor. 

39 actual_seq_kvlen: Cumulative key lengths as an int32 NPU tensor. 

40 sparse_count: Number of token indices selected for each query. 

41 

42 Returns: 

43 Top-k indices in ``[T, N_kv, K]`` int32 layout, followed by the query, 

44 key, and merge-weight tensors in TND, TND, and TN layouts. Positions 

45 without enough causal candidates are filled with ``-1``. The flattened 

46 tensors can be passed directly to :func:`dsa_kl_loss`. 

47 """ 

48 index_query_tnd = rearrange(index_query, "b s n d -> (b s) n d") 

49 index_key_tnd = rearrange(index_key, "b s n d -> (b s) n d") 

50 merge_weight_tnd = rearrange(merge_weight, "b s n -> (b s) n") 

51 topk_indices, _ = torch.ops.custom.npu_lightning_indexer_enhance( 

52 index_query_tnd, 

53 index_key_tnd, 

54 merge_weight_tnd, 

55 actual_seq_lengths_query=actual_seq_qlen, 

56 actual_seq_lengths_key=actual_seq_kvlen, 

57 block_table=None, 

58 layout_query="TND", 

59 layout_key="TND", 

60 sparse_count=sparse_count, 

61 sparse_mode=3, 

62 return_value=False, 

63 ) 

64 return topk_indices, index_query_tnd, index_key_tnd, merge_weight_tnd