File size: 4,883 Bytes
bc3697f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
# -*- coding: utf-8 -*-
# Copyright 2026 EngineerGL Research.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from transformers import PretrainedConfig

class AlinlightConfig(PretrainedConfig):
    """
    Configuration class for Alinlight model.
    
    Args:
        vocab_size (int): Vocabulary size of the model.
        hidden_size (int): Dimensionality of the encoder layers and the pooler layer.
        intermediate_size (int): Dimensionality of the "intermediate" (i.e., feed-forward) layer.
        num_hidden_layers (int): Number of hidden layers in the Transformer encoder.
        num_attention_heads (int): Number of attention heads for each attention layer.
        num_key_value_heads (int): Number of key/value heads for Grouped Query Attention.
        max_position_embeddings (int): The maximum sequence length that this model might ever be used with.
        rope_theta (float): The base period of the RoPE embeddings.
        rope_scaling (dict, optional): Dictionary containing the scaling configuration for the RoPE embeddings.
        sliding_window (int, optional): Sliding window size for local attention. None to disable.
        attention_dropout (float): The dropout ratio for the attention probabilities.
        use_qk_norm (bool): Whether to apply RMSNorm to Query and Key matrices.
        attn_logit_softcapping (float, optional): If set, applies tanh soft-capping to attention logits (Gemma-2 style).
        rms_norm_eps (float): The epsilon used by the rms normalization layers.
        initializer_range (float): The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        resid_pdrop (float): The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        embed_pdrop (float): The dropout probability for the embedding layer.
        embed_scale (bool): Whether to scale embeddings by sqrt(hidden_size).
        final_logit_softcapping (float, optional): If set, applies tanh soft-capping to final LM head logits.
        z_loss_weight (float): Coefficient for the Z-loss regularization term (stabilizes final logits).
    """
    model_type = "alinlight"
    
    def __init__(
        self,
        # Architecture
        vocab_size=128000,
        hidden_size=2048,
        intermediate_size=5632,
        num_hidden_layers=22,
        num_attention_heads=32,
        num_key_value_heads=8,
        
        # Positional Encoding
        max_position_embeddings=4096,
        rope_theta=10000.0,
        rope_scaling=None,
        
        # Attention
        sliding_window=None,
        attention_dropout=0.0,
        use_qk_norm=True,
        attn_logit_softcapping=50.0,
        
        # Normalization & Regularization
        rms_norm_eps=1e-6,
        initializer_range=0.02,
        resid_pdrop=0.0,
        embed_pdrop=0.0,
        
        # Stability Features
        embed_scale=True,
        final_logit_softcapping=30.0,
        z_loss_weight=1e-4,
        
        # System
        use_cache=True,
        pad_token_id=0,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=True,
        
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.max_position_embeddings = max_position_embeddings
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.sliding_window = sliding_window
        self.attention_dropout = attention_dropout
        self.use_qk_norm = use_qk_norm
        self.attn_logit_softcapping = attn_logit_softcapping
        self.rms_norm_eps = rms_norm_eps
        self.initializer_range = initializer_range
        self.resid_pdrop = resid_pdrop
        self.embed_pdrop = embed_pdrop
        self.embed_scale = embed_scale
        self.final_logit_softcapping = final_logit_softcapping
        self.z_loss_weight = z_loss_weight
        self.use_cache = use_cache

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs
        )