Update configuration_neollm.py
Browse files- configuration_neollm.py +16 -8
configuration_neollm.py
CHANGED
|
@@ -7,9 +7,9 @@ logger = logging.get_logger(__name__)
|
|
| 7 |
|
| 8 |
class NeoLLMConfig(PretrainedConfig):
|
| 9 |
r"""
|
| 10 |
-
This is the configuration class to store the configuration of a
|
| 11 |
NeoLLM model according to the specified arguments, defining the model architecture.
|
| 12 |
-
Configuration objects inherit from
|
| 13 |
"""
|
| 14 |
model_type = "neollm"
|
| 15 |
keys_to_ignore_at_inference = []
|
|
@@ -33,10 +33,13 @@ class NeoLLMConfig(PretrainedConfig):
|
|
| 33 |
attention_bias=False,
|
| 34 |
attention_dropout=0.1,
|
| 35 |
head_dim=64,
|
| 36 |
-
|
| 37 |
fan_ratio=0.125,
|
| 38 |
-
fan_ratio_ffn=0.0625,
|
| 39 |
dropout_rate=0.1,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
**kwargs,
|
| 41 |
):
|
| 42 |
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
|
|
@@ -56,19 +59,24 @@ class NeoLLMConfig(PretrainedConfig):
|
|
| 56 |
self.attention_bias = attention_bias
|
| 57 |
self.attention_dropout = attention_dropout
|
| 58 |
self.head_dim = head_dim
|
| 59 |
-
|
| 60 |
rope_config_validation(self)
|
| 61 |
|
| 62 |
# FANformer parameters
|
| 63 |
-
self.fan_ratio = fan_ratio
|
| 64 |
-
self.fan_ratio_ffn = fan_ratio_ffn
|
| 65 |
|
| 66 |
self.dropout_rate = dropout_rate
|
| 67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
self.auto_map = {
|
| 69 |
"AutoConfig": "configuration_neollm.NeoLLMConfig",
|
| 70 |
"AutoModel": "modeling_neollm.NeoLLMModel",
|
| 71 |
"AutoModelForCausalLM": "modeling_neollm.NeoLLMForCausalLM"
|
| 72 |
}
|
| 73 |
|
| 74 |
-
__all__ = ["NeoLLMConfig"]
|
|
|
|
| 7 |
|
| 8 |
class NeoLLMConfig(PretrainedConfig):
|
| 9 |
r"""
|
| 10 |
+
This is the configuration class to store the configuration of a `NeoLLMModel`. It is used to instantiate a
|
| 11 |
NeoLLM model according to the specified arguments, defining the model architecture.
|
| 12 |
+
Configuration objects inherit from `PretrainedConfig` and can be used to control the model outputs.
|
| 13 |
"""
|
| 14 |
model_type = "neollm"
|
| 15 |
keys_to_ignore_at_inference = []
|
|
|
|
| 33 |
attention_bias=False,
|
| 34 |
attention_dropout=0.1,
|
| 35 |
head_dim=64,
|
|
|
|
| 36 |
fan_ratio=0.125,
|
| 37 |
+
fan_ratio_ffn=0.0625,
|
| 38 |
dropout_rate=0.1,
|
| 39 |
+
use_stack=True,
|
| 40 |
+
num_stack_heads=4,
|
| 41 |
+
stack_slots=24,
|
| 42 |
+
stack_d_model=16,
|
| 43 |
**kwargs,
|
| 44 |
):
|
| 45 |
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
|
|
|
|
| 59 |
self.attention_bias = attention_bias
|
| 60 |
self.attention_dropout = attention_dropout
|
| 61 |
self.head_dim = head_dim
|
|
|
|
| 62 |
rope_config_validation(self)
|
| 63 |
|
| 64 |
# FANformer parameters
|
| 65 |
+
self.fan_ratio = fan_ratio
|
| 66 |
+
self.fan_ratio_ffn = fan_ratio_ffn
|
| 67 |
|
| 68 |
self.dropout_rate = dropout_rate
|
| 69 |
|
| 70 |
+
# StackMemory parameters
|
| 71 |
+
self.use_stack = use_stack
|
| 72 |
+
self.num_stack_heads = num_stack_heads
|
| 73 |
+
self.stack_slots = stack_slots
|
| 74 |
+
self.stack_d_model = stack_d_model
|
| 75 |
+
|
| 76 |
self.auto_map = {
|
| 77 |
"AutoConfig": "configuration_neollm.NeoLLMConfig",
|
| 78 |
"AutoModel": "modeling_neollm.NeoLLMModel",
|
| 79 |
"AutoModelForCausalLM": "modeling_neollm.NeoLLMForCausalLM"
|
| 80 |
}
|
| 81 |
|
| 82 |
+
__all__ = ["NeoLLMConfig"]
|