matthewkenney commited on
Commit
f661692
1 Parent(s): 075337d

Create configuration_mixformer_sequential.py

Browse files
configuration_mixformer_sequential.py ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Microsoft Corporation.
2
+ # Licensed under the MIT license.
3
+
4
+ import math
5
+ from typing import Any, Dict, List, Optional, Union
6
+
7
+ from transformers import PretrainedConfig
8
+
9
+
10
+ class MixFormerSequentialConfig(PretrainedConfig):
11
+ """MixFormer (sequential for DeepSpeed) configuration."""
12
+
13
+ model_type = "mixformer-sequential"
14
+
15
+ attribute_map = {
16
+ "max_position_embeddings": "n_positions",
17
+ "hidden_size": "n_embd",
18
+ "num_attention_heads": "n_head",
19
+ "num_hidden_layers": "n_layer",
20
+ "input_emb_layer": "embd_layer", # `input_emb_layer` key is for backward compatibility
21
+ "blocks": "architecture", # `blocks` key is for backward compatibility
22
+ }
23
+
24
+ def __init__(
25
+ self,
26
+ vocab_size: Optional[int] = 50304,
27
+ n_positions: Optional[int] = 2048,
28
+ n_embd: Optional[int] = 1024,
29
+ n_layer: Optional[int] = 20,
30
+ n_inner: Optional[int] = None,
31
+ n_head: Optional[int] = 16,
32
+ rotary_dim: Optional[int] = 32,
33
+ activation_function: Optional[str] = "gelu_new",
34
+ embd_layer: Optional[str] = "default",
35
+ architecture: Union[Dict[str, Any], List[Dict[str, Any]]] = None,
36
+ embd_pdrop: Optional[float] = 0.0,
37
+ resid_pdrop: Optional[float] = 0.0,
38
+ layer_norm_epsilon: Optional[float] = 1e-5,
39
+ initializer_range: Optional[float] = 0.02,
40
+ tie_word_embeddings: Optional[bool] = False,
41
+ pad_vocab_size_multiple: Optional[int] = 64,
42
+ **kwargs
43
+ ) -> None:
44
+ self.vocab_size = int(math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple)
45
+ self.n_positions = n_positions
46
+ self.n_embd = n_embd
47
+ self.n_layer = n_layer
48
+ self.n_inner = n_inner
49
+ self.n_head = n_head
50
+ self.rotary_dim = min(rotary_dim, n_embd // n_head)
51
+ self.activation_function = activation_function
52
+ self.embd_layer = embd_layer
53
+ self.architecture = architecture
54
+ self.embd_pdrop = embd_pdrop
55
+ self.resid_pdrop = resid_pdrop
56
+ self.layer_norm_epsilon = layer_norm_epsilon
57
+ self.initializer_range = initializer_range
58
+
59
+ super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)