Update configuration_rwkv5.py
Browse files- configuration_rwkv5.py +23 -36
configuration_rwkv5.py
CHANGED
@@ -21,79 +21,67 @@ from transformers.utils import logging
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
}
|
27 |
|
28 |
|
29 |
class Rwkv5Config(PretrainedConfig):
|
30 |
"""
|
31 |
-
This is the configuration class to store the configuration of a [`
|
32 |
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
33 |
defaults will yield a similar configuration to that of the RWVK-4
|
34 |
-
[RWKV/rwkv-
|
35 |
-
|
36 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
37 |
documentation from [`PretrainedConfig`] for more information.
|
38 |
-
|
39 |
-
|
40 |
Args:
|
41 |
-
vocab_size (`int`, *optional*, defaults to
|
42 |
-
Vocabulary size of the
|
43 |
-
`inputs_ids` passed when calling [`
|
44 |
-
|
45 |
-
The maximum sequence length that this model can be be used with in a single forward (using it in RNN mode
|
46 |
-
lets use any sequence length).
|
47 |
-
hidden_size (`int`, *optional*, defaults to 4096):
|
48 |
Dimensionality of the embeddings and hidden states.
|
49 |
-
num_hidden_layers (`int`, *optional*, defaults to
|
50 |
Number of hidden layers in the model.
|
51 |
attention_hidden_size (`int`, *optional*):
|
52 |
Dimensionality of the attention hidden states. Will default to `hidden_size` if unset.
|
|
|
|
|
|
|
53 |
intermediate_size (`int`, *optional*):
|
54 |
Dimensionality of the inner feed-forward layers. Will default to 4 times `hidden_size` if unset.
|
55 |
-
|
56 |
The epsilon to use in the layer normalization layers.
|
57 |
bos_token_id (`int`, *optional*, defaults to 0):
|
58 |
-
The id of the beginning of sentence token in the vocabulary. Defaults to 0 as
|
59 |
as GPTNeoX.
|
60 |
eos_token_id (`int`, *optional*, defaults to 0):
|
61 |
-
The id of the end of sentence token in the vocabulary. Defaults to 0 as
|
62 |
GPTNeoX.
|
63 |
-
rescale_every (`int`, *optional*,
|
64 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
65 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
66 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
67 |
Whether or not to tie the word embeddings with the input token embeddings.
|
68 |
use_cache (`bool`, *optional*, defaults to `True`):
|
69 |
Whether or not the model should return the last state.
|
70 |
-
|
71 |
-
|
72 |
Example:
|
73 |
-
|
74 |
```python
|
75 |
-
>>> from transformers import
|
76 |
-
|
77 |
-
>>>
|
78 |
-
>>> configuration = RwkvConfig()
|
79 |
-
|
80 |
>>> # Initializing a model (with random weights) from the configuration
|
81 |
-
>>> model =
|
82 |
-
|
83 |
>>> # Accessing the model configuration
|
84 |
>>> configuration = model.config
|
85 |
```"""
|
86 |
|
87 |
model_type = "rwkv5"
|
88 |
-
attribute_map = {"max_position_embeddings": "context_length"}
|
89 |
|
90 |
-
def __init__(
|
91 |
self,
|
92 |
vocab_size=65536,
|
93 |
-
context_length=4096,
|
94 |
hidden_size=768,
|
95 |
num_hidden_layers=24,
|
96 |
attention_hidden_size=None,
|
|
|
97 |
head_size=64,
|
98 |
intermediate_size=None,
|
99 |
layer_norm_epsilon=1e-5,
|
@@ -102,14 +90,13 @@ class Rwkv5Config(PretrainedConfig):
|
|
102 |
rescale_every=6,
|
103 |
tie_word_embeddings=False,
|
104 |
use_cache=True,
|
105 |
-
model_version="5_2",
|
106 |
**kwargs,
|
107 |
):
|
108 |
self.vocab_size = vocab_size
|
109 |
-
self.context_length = context_length
|
110 |
self.hidden_size = hidden_size
|
111 |
self.num_hidden_layers = num_hidden_layers
|
112 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
|
|
113 |
self.head_size = head_size
|
114 |
self.intermediate_size = None
|
115 |
self.layer_norm_epsilon = layer_norm_epsilon
|
@@ -118,8 +105,8 @@ class Rwkv5Config(PretrainedConfig):
|
|
118 |
|
119 |
self.bos_token_id = bos_token_id
|
120 |
self.eos_token_id = eos_token_id
|
121 |
-
self.model_version = model_version
|
122 |
|
123 |
super().__init__(
|
124 |
tie_word_embeddings=tie_word_embeddings, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs
|
125 |
)
|
|
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
24 |
+
RWKV5_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
|
|
|
|
|
25 |
|
26 |
|
27 |
class Rwkv5Config(PretrainedConfig):
|
28 |
"""
|
29 |
+
This is the configuration class to store the configuration of a [`Rwkv5Model`]. It is used to instantiate a RWKV5
|
30 |
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
31 |
defaults will yield a similar configuration to that of the RWVK-4
|
32 |
+
[RWKV/rwkv-5-world-1b5](https://huggingface.co/RWKV/rwkv-5-world-1b5) architecture.
|
|
|
33 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
34 |
documentation from [`PretrainedConfig`] for more information.
|
|
|
|
|
35 |
Args:
|
36 |
+
vocab_size (`int`, *optional*, defaults to 65536):
|
37 |
+
Vocabulary size of the RWKV5 model. Defines the number of different tokens that can be represented by the
|
38 |
+
`inputs_ids` passed when calling [`Rwkv5Model`].
|
39 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
|
|
|
|
|
|
40 |
Dimensionality of the embeddings and hidden states.
|
41 |
+
num_hidden_layers (`int`, *optional*, defaults to 24):
|
42 |
Number of hidden layers in the model.
|
43 |
attention_hidden_size (`int`, *optional*):
|
44 |
Dimensionality of the attention hidden states. Will default to `hidden_size` if unset.
|
45 |
+
num_attention_heads (`int`, *optional*, defaults to 64):
|
46 |
+
The attention heads to use in rwkv5 self_attention module.
|
47 |
+
head_size (`int`, *optional*, defaults to 64): head_size of rwkv5 self_attention module.
|
48 |
intermediate_size (`int`, *optional*):
|
49 |
Dimensionality of the inner feed-forward layers. Will default to 4 times `hidden_size` if unset.
|
50 |
+
layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
|
51 |
The epsilon to use in the layer normalization layers.
|
52 |
bos_token_id (`int`, *optional*, defaults to 0):
|
53 |
+
The id of the beginning of sentence token in the vocabulary. Defaults to 0 as RWKV5 uses the same tokenizer
|
54 |
as GPTNeoX.
|
55 |
eos_token_id (`int`, *optional*, defaults to 0):
|
56 |
+
The id of the end of sentence token in the vocabulary. Defaults to 0 as RWKV5 uses the same tokenizer as
|
57 |
GPTNeoX.
|
58 |
+
rescale_every (`int`, *optional*, defaults to 6):
|
59 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
60 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
61 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
62 |
Whether or not to tie the word embeddings with the input token embeddings.
|
63 |
use_cache (`bool`, *optional*, defaults to `True`):
|
64 |
Whether or not the model should return the last state.
|
|
|
|
|
65 |
Example:
|
|
|
66 |
```python
|
67 |
+
>>> from transformers import Rwkv5Config, Rwkv5Model
|
68 |
+
>>> # Initializing a Rwkv5 configuration
|
69 |
+
>>> configuration = Rwkv5Config()
|
|
|
|
|
70 |
>>> # Initializing a model (with random weights) from the configuration
|
71 |
+
>>> model = Rwkv5Model(configuration)
|
|
|
72 |
>>> # Accessing the model configuration
|
73 |
>>> configuration = model.config
|
74 |
```"""
|
75 |
|
76 |
model_type = "rwkv5"
|
|
|
77 |
|
78 |
+
def __init__(
|
79 |
self,
|
80 |
vocab_size=65536,
|
|
|
81 |
hidden_size=768,
|
82 |
num_hidden_layers=24,
|
83 |
attention_hidden_size=None,
|
84 |
+
num_attention_heads=64,
|
85 |
head_size=64,
|
86 |
intermediate_size=None,
|
87 |
layer_norm_epsilon=1e-5,
|
|
|
90 |
rescale_every=6,
|
91 |
tie_word_embeddings=False,
|
92 |
use_cache=True,
|
|
|
93 |
**kwargs,
|
94 |
):
|
95 |
self.vocab_size = vocab_size
|
|
|
96 |
self.hidden_size = hidden_size
|
97 |
self.num_hidden_layers = num_hidden_layers
|
98 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
99 |
+
self.num_attention_heads = num_attention_heads
|
100 |
self.head_size = head_size
|
101 |
self.intermediate_size = None
|
102 |
self.layer_norm_epsilon = layer_norm_epsilon
|
|
|
105 |
|
106 |
self.bos_token_id = bos_token_id
|
107 |
self.eos_token_id = eos_token_id
|
|
|
108 |
|
109 |
super().__init__(
|
110 |
tie_word_embeddings=tie_word_embeddings, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs
|
111 |
)
|
112 |
+
|