Create modeling_theta.py
Browse files- modeling_theta.py +117 -0
modeling_theta.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast
|
4 |
+
from transformers import PreTrainedModel
|
5 |
+
from configuration_theta import ThetaConfig
|
6 |
+
|
7 |
+
class ThetaAttention(nn.Module):
|
8 |
+
def __init__(self, config: ThetaConfig):
|
9 |
+
super().__init__()
|
10 |
+
self.num_heads = config.num_attention_heads
|
11 |
+
self.head_dim = config.hidden_size // config.num_attention_heads
|
12 |
+
self.scale = self.head_dim ** -0.5
|
13 |
+
|
14 |
+
self.q_proj = nn.Linear(config.hidden_size, config.hidden_size)
|
15 |
+
self.k_proj = nn.Linear(config.hidden_size, config.hidden_size)
|
16 |
+
self.v_proj = nn.Linear(config.hidden_size, config.hidden_size)
|
17 |
+
self.out_proj = nn.Linear(config.hidden_size, config.hidden_size)
|
18 |
+
|
19 |
+
def forward(self, hidden_states):
|
20 |
+
batch_size, seq_length, embed_dim = hidden_states.size()
|
21 |
+
|
22 |
+
query = self.q_proj(hidden_states)
|
23 |
+
key = self.k_proj(hidden_states)
|
24 |
+
value = self.v_proj(hidden_states)
|
25 |
+
|
26 |
+
query = query.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
27 |
+
key = key.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
28 |
+
value = value.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
29 |
+
|
30 |
+
attn_scores = torch.matmul(query, key.transpose(-2, -1)) * self.scale
|
31 |
+
attn_probs = nn.functional.softmax(attn_scores, dim=-1)
|
32 |
+
attn_output = torch.matmul(attn_probs, value)
|
33 |
+
|
34 |
+
attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, seq_length, embed_dim)
|
35 |
+
attn_output = self.out_proj(attn_output)
|
36 |
+
|
37 |
+
return attn_output
|
38 |
+
|
39 |
+
class ThetaMLP(nn.Module):
|
40 |
+
def __init__(self, config: ThetaConfig):
|
41 |
+
super().__init__()
|
42 |
+
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
|
43 |
+
self.act = nn.SiLU()
|
44 |
+
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
|
45 |
+
|
46 |
+
def forward(self, hidden_states):
|
47 |
+
hidden_states = self.fc1(hidden_states)
|
48 |
+
hidden_states = self.act(hidden_states)
|
49 |
+
hidden_states = self.fc2(hidden_states)
|
50 |
+
return hidden_states
|
51 |
+
|
52 |
+
class ThetaBlock(nn.Module):
|
53 |
+
def __init__(self, config: ThetaConfig):
|
54 |
+
super().__init__()
|
55 |
+
self.attention = ThetaAttention(config)
|
56 |
+
self.mlp = ThetaMLP(config)
|
57 |
+
self.norm1 = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps)
|
58 |
+
self.norm2 = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps)
|
59 |
+
|
60 |
+
def forward(self, hidden_states):
|
61 |
+
hidden_states = hidden_states + self.attention(self.norm1(hidden_states))
|
62 |
+
hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
|
63 |
+
return hidden_states
|
64 |
+
|
65 |
+
class ThetaModel(PreTrainedModel):
|
66 |
+
config_class = ThetaConfig
|
67 |
+
|
68 |
+
def __init__(self, config: ThetaConfig):
|
69 |
+
super().__init__(config)
|
70 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
|
71 |
+
self.layers = nn.ModuleList([ThetaBlock(config) for _ in range(config.num_hidden_layers)])
|
72 |
+
self.norm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps)
|
73 |
+
|
74 |
+
def forward(
|
75 |
+
self,
|
76 |
+
input_ids=None,
|
77 |
+
attention_mask=None,
|
78 |
+
**kwargs,
|
79 |
+
):
|
80 |
+
hidden_states = self.embed_tokens(input_ids)
|
81 |
+
|
82 |
+
for layer in self.layers:
|
83 |
+
hidden_states = layer(hidden_states)
|
84 |
+
|
85 |
+
hidden_states = self.norm(hidden_states)
|
86 |
+
|
87 |
+
return BaseModelOutputWithPast(
|
88 |
+
last_hidden_state=hidden_states,
|
89 |
+
hidden_states=None,
|
90 |
+
past_key_values=None,
|
91 |
+
)
|
92 |
+
|
93 |
+
class ThetaForCausalLM(PreTrainedModel):
|
94 |
+
config_class = ThetaConfig
|
95 |
+
|
96 |
+
def __init__(self, config: ThetaConfig):
|
97 |
+
super().__init__(config)
|
98 |
+
self.model = ThetaModel(config)
|
99 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
100 |
+
|
101 |
+
def forward(
|
102 |
+
self,
|
103 |
+
input_ids=None,
|
104 |
+
labels=None,
|
105 |
+
**kwargs,
|
106 |
+
):
|
107 |
+
outputs = self.model(input_ids=input_ids, **kwargs)
|
108 |
+
logits = self.lm_head(outputs.last_hidden_state)
|
109 |
+
|
110 |
+
loss = None
|
111 |
+
if labels is not None:
|
112 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
113 |
+
shift_labels = labels[..., 1:].contiguous()
|
114 |
+
loss_fct = nn.CrossEntropyLoss()
|
115 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
116 |
+
|
117 |
+
return {"loss": loss, "logits": logits} if loss is not None else {"logits": logits}
|