|
14 | 14 |
|
15 | 15 | import re
|
16 | 16 | from pathlib import Path
|
17 |
| -from typing import List, Optional, Tuple, Union |
| 17 | +from typing import List, Optional, Union |
18 | 18 |
|
19 | 19 | import torch
|
20 | 20 | from huggingface_hub import HfApi, HfFolder
|
|
23 | 23 | MULTI_QUERY_ATTN_MODELS = {"falcon", "gpt_bigcode"}
|
24 | 24 |
|
25 | 25 |
|
26 |
| -# Modified from transformers.models.bloom.modeling_bloom._make_causal_mask |
27 |
| -def _make_causal_mask( |
28 |
| - input_ids_shape: torch.Size, |
29 |
| - device: torch.device, |
30 |
| - past_key_values_length: int, |
31 |
| - dtype: torch.dtype = torch.bool, |
32 |
| -) -> torch.BoolTensor: |
33 |
| - """ |
34 |
| - Make causal mask used for bi-directional self-attention. |
35 |
| - """ |
36 |
| - batch_size, target_length = input_ids_shape |
37 |
| - mask = torch.zeros((target_length, target_length + past_key_values_length), dtype=dtype, device=device) |
38 |
| - seq_ids = torch.arange(target_length, device=device) |
39 |
| - |
40 |
| - mask[:, past_key_values_length:] = ( |
41 |
| - (seq_ids[:, None] < seq_ids[None, :]) * torch.finfo(dtype).min |
42 |
| - if torch.is_floating_point(mask) |
43 |
| - else seq_ids[:, None] < seq_ids[None, :] |
44 |
| - ) |
45 |
| - |
46 |
| - return mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length) |
47 |
| - |
48 |
| - |
49 |
| -# Modified from transformers.models..bloom.modeling_bloom._prepare_attn_mask |
50 |
| -def _prepare_attn_mask( |
51 |
| - attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int |
52 |
| -) -> torch.BoolTensor: |
53 |
| - from transformers.models.bloom.modeling_bloom import _expand_mask |
54 |
| - |
55 |
| - # create causal mask |
56 |
| - # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length] |
57 |
| - combined_attention_mask = None |
58 |
| - device = attention_mask.device |
59 |
| - _, src_length = input_shape |
60 |
| - |
61 |
| - combined_attention_mask = _make_causal_mask( |
62 |
| - input_shape, device=device, past_key_values_length=past_key_values_length |
63 |
| - ) |
64 |
| - # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]_prepare_decoder_attention_mask |
65 |
| - expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length) |
66 |
| - combined_attention_mask = ( |
67 |
| - expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask |
68 |
| - ) |
69 |
| - |
70 |
| - return combined_attention_mask |
71 |
| - |
72 |
| - |
73 |
| -# Modified from transformers.models.llama.modeling_llama._prepare_decoder_attention_mask |
74 |
| -def _prepare_decoder_attention_mask(attention_mask, input_shape, inputs_embeds, past_key_values_length): |
75 |
| - from transformers.models.llama.modeling_llama import _expand_mask |
76 |
| - |
77 |
| - # create causal mask |
78 |
| - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] |
79 |
| - combined_attention_mask = None |
80 |
| - |
81 |
| - combined_attention_mask = _make_causal_mask( |
82 |
| - input_shape, |
83 |
| - device=inputs_embeds.device, |
84 |
| - past_key_values_length=past_key_values_length, |
85 |
| - dtype=inputs_embeds.dtype, |
86 |
| - ) |
87 |
| - |
88 |
| - if attention_mask is not None: |
89 |
| - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] |
90 |
| - expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to( |
91 |
| - inputs_embeds.device |
92 |
| - ) |
93 |
| - combined_attention_mask = ( |
94 |
| - expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask |
95 |
| - ) |
96 |
| - |
97 |
| - return combined_attention_mask |
98 |
| - |
99 |
| - |
100 |
| -# Modified from transformers.models.mistral.modeling_mistral._prepare_decoder_sliding_window_attention_mask |
101 |
| -def _prepare_decoder_sliding_window_attention_mask( |
102 |
| - attention_mask: torch.Tensor, |
103 |
| - input_shape: Tuple[int, int], |
104 |
| - inputs_embeds: torch.Tensor, |
105 |
| - past_key_values_length: int, |
106 |
| - sliding_window: int, |
107 |
| -): |
108 |
| - from transformers.models.mistral.modeling_mistral import _expand_mask, _make_sliding_window_causal_mask |
109 |
| - |
110 |
| - # create causal mask |
111 |
| - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] |
112 |
| - combined_attention_mask = None |
113 |
| - |
114 |
| - combined_attention_mask = _make_sliding_window_causal_mask( |
115 |
| - input_shape, |
116 |
| - device=inputs_embeds.device, |
117 |
| - dtype=inputs_embeds.dtype, |
118 |
| - past_key_values_length=past_key_values_length, |
119 |
| - sliding_window=sliding_window, |
120 |
| - ) |
121 |
| - |
122 |
| - if attention_mask is not None: |
123 |
| - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] |
124 |
| - expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to( |
125 |
| - inputs_embeds.device |
126 |
| - ) |
127 |
| - combined_attention_mask = ( |
128 |
| - expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask |
129 |
| - ) |
130 |
| - |
131 |
| - return combined_attention_mask |
132 |
| - |
133 |
| - |
134 | 26 | def get_model_device(model: torch.nn.Module) -> torch.device:
|
135 | 27 | """
|
136 | 28 | Determines the device on which a PyTorch model is currently residing.
|
|
0 commit comments