Spaces:
Running
on
Zero
Running
on
Zero
Overhaul code for appropriate masking for full model instead of just attention layers
Browse files- llama_diffusion_model.py +54 -154
llama_diffusion_model.py
CHANGED
@@ -1,115 +1,57 @@
|
|
|
|
1 |
import torch.nn as nn
|
2 |
-
from transformers import AutoModelForCausalLM, PreTrainedModel, PretrainedConfig
|
3 |
-
from transformers.models.llama.modeling_llama import LlamaAttention
|
4 |
from torch.amp import autocast
|
|
|
|
|
5 |
from peft import LoraConfig, get_peft_model
|
6 |
-
from typing import Optional, Tuple
|
7 |
-
import torch
|
8 |
import os
|
9 |
-
|
10 |
-
|
11 |
|
12 |
hf_token = os.getenv("HF_TOKEN")
|
13 |
|
14 |
class BidirectionalLlamaAttention(LlamaAttention):
|
15 |
-
def __init__(self, original_layer, masking
|
16 |
super().__init__(original_layer.config, layer_idx=original_layer.layer_idx)
|
17 |
self.masking = masking
|
18 |
-
|
19 |
-
# Copy weights from original layer
|
20 |
self.q_proj.weight = original_layer.q_proj.weight
|
21 |
self.k_proj.weight = original_layer.k_proj.weight
|
22 |
self.v_proj.weight = original_layer.v_proj.weight
|
23 |
self.o_proj.weight = original_layer.o_proj.weight
|
24 |
|
25 |
def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
26 |
-
"""
|
27 |
-
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
28 |
-
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
29 |
-
"""
|
30 |
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
31 |
if n_rep == 1:
|
32 |
return hidden_states
|
33 |
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
34 |
-
|
35 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
36 |
|
37 |
-
def eager_attention_forward(self,
|
38 |
-
module: nn.Module,
|
39 |
-
query: torch.Tensor,
|
40 |
-
key: torch.Tensor,
|
41 |
-
value: torch.Tensor,
|
42 |
-
attention_mask: Optional[torch.Tensor],
|
43 |
-
scaling: float,
|
44 |
-
dropout: float = 0.0,
|
45 |
-
**kwargs,
|
46 |
-
):
|
47 |
key_states = self.repeat_kv(key, module.num_key_value_groups)
|
48 |
value_states = self.repeat_kv(value, module.num_key_value_groups)
|
49 |
-
|
50 |
-
# attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
51 |
-
# if attention_mask is not None:
|
52 |
-
# causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
53 |
-
# attn_weights = attn_weights + causal_mask
|
54 |
-
|
55 |
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
|
|
56 |
if attention_mask is not None:
|
57 |
-
# Convert bool -> float with -inf where masked
|
58 |
attn_mask = attention_mask.masked_fill(~attention_mask, float('-inf')).to(query.dtype)
|
59 |
attn_weights = attn_weights + attn_mask
|
60 |
|
61 |
-
|
62 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query.dtype)
|
63 |
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
64 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
65 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
66 |
-
|
67 |
return attn_output, attn_weights
|
68 |
|
69 |
def rotate_half(self, x):
|
70 |
-
"""Rotates half the hidden dims of the input."""
|
71 |
x1 = x[..., : x.shape[-1] // 2]
|
72 |
-
x2 = x[..., x.shape[-1] // 2
|
73 |
-
|
74 |
return torch.cat((-x2, x1), dim=-1)
|
75 |
|
76 |
-
|
77 |
-
def apply_rotary_pos_emb(self, q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
78 |
-
"""Applies Rotary Position Embedding to the query and key tensors.
|
79 |
-
|
80 |
-
Args:
|
81 |
-
q (`torch.Tensor`): The query tensor.
|
82 |
-
k (`torch.Tensor`): The key tensor.
|
83 |
-
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
84 |
-
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
85 |
-
position_ids (`torch.Tensor`, *optional*):
|
86 |
-
Deprecated and unused.
|
87 |
-
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
88 |
-
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
89 |
-
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
90 |
-
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
91 |
-
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
92 |
-
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
93 |
-
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
94 |
-
Returns:
|
95 |
-
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
96 |
-
"""
|
97 |
cos = cos.unsqueeze(unsqueeze_dim)
|
98 |
sin = sin.unsqueeze(unsqueeze_dim)
|
99 |
q_embed = (q * cos) + (self.rotate_half(q) * sin)
|
100 |
k_embed = (k * cos) + (self.rotate_half(k) * sin)
|
101 |
-
|
102 |
return q_embed, k_embed
|
103 |
|
104 |
-
def forward(
|
105 |
-
self,
|
106 |
-
hidden_states: torch.Tensor,
|
107 |
-
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
108 |
-
attention_mask: Optional[torch.Tensor],
|
109 |
-
past_key_value: Optional[torch.Tensor] = None,
|
110 |
-
cache_position: Optional[torch.LongTensor] = None,
|
111 |
-
**kwargs,
|
112 |
-
):
|
113 |
input_shape = hidden_states.shape[:-1]
|
114 |
hidden_shape = (*input_shape, -1, self.head_dim)
|
115 |
|
@@ -117,7 +59,6 @@ class BidirectionalLlamaAttention(LlamaAttention):
|
|
117 |
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
118 |
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
119 |
|
120 |
-
# Apply rotary embeddings
|
121 |
cos, sin = position_embeddings
|
122 |
query_states, key_states = self.apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
123 |
|
@@ -125,58 +66,17 @@ class BidirectionalLlamaAttention(LlamaAttention):
|
|
125 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
126 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
127 |
|
128 |
-
# 🔄 **Modify the Attention Mask**
|
129 |
-
seq_len = hidden_states.shape[1]
|
130 |
-
batch_size = hidden_states.shape[0]
|
131 |
-
if self.masking == 'bidirectional':
|
132 |
-
base_mask = torch.ones((seq_len, seq_len), device=hidden_states.device, dtype=torch.bool)
|
133 |
-
attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone() # ✅ Copy for each batch
|
134 |
-
elif self.masking == 'bidirectional_masked':
|
135 |
-
base_mask = torch.ones((seq_len, seq_len), device=hidden_states.device, dtype=torch.bool)
|
136 |
-
base_mask[:, :].fill_diagonal_(False) # ✅ Apply diagonal masking only in 2D
|
137 |
-
attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone() # ✅ Copy for each batch
|
138 |
-
else: # unidirectional
|
139 |
-
# 🚀 Standard autoregressive (causal) mask
|
140 |
-
attn_mask = torch.tril(torch.ones(seq_len, seq_len, device=hidden_states.device, dtype=torch.bool))
|
141 |
-
attn_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone() # ✅ Copy for each batch
|
142 |
-
|
143 |
-
|
144 |
-
# Call the default attention function
|
145 |
attn_output, attn_weights = self.eager_attention_forward(
|
146 |
-
self,
|
147 |
-
|
148 |
-
key_states,
|
149 |
-
value_states,
|
150 |
-
attn_mask, # ✅ Custom mask is applied here
|
151 |
-
dropout=0.0 if not self.training else self.attention_dropout,
|
152 |
-
scaling=self.scaling,
|
153 |
-
**kwargs,
|
154 |
)
|
155 |
|
156 |
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
157 |
-
|
158 |
-
|
159 |
-
return attn_output, attn_weights
|
160 |
-
|
161 |
-
|
162 |
-
def _split_heads(self, tensor, num_heads, attn_head_size):
|
163 |
-
"""
|
164 |
-
Splits hidden_size dim into attn_head_size and num_heads
|
165 |
-
"""
|
166 |
-
new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
|
167 |
-
tensor = tensor.view(*new_shape)
|
168 |
-
return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
|
169 |
-
|
170 |
-
def _merge_heads(self, tensor, num_heads, attn_head_size):
|
171 |
-
"""
|
172 |
-
Merges attn_head_size dim and num_attn_heads dim into hidden_size
|
173 |
-
"""
|
174 |
-
tensor = tensor.permute(0, 2, 1, 3).contiguous()
|
175 |
-
new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
|
176 |
-
return tensor.view(new_shape)
|
177 |
|
178 |
class CustomTransformerConfig(PretrainedConfig):
|
179 |
-
def __init__(self, vocab_size=128256, hidden_size=4096, num_layers=32, num_heads=32, prediction_chunk=256, dropout=0,
|
|
|
180 |
super().__init__(**kwargs)
|
181 |
self.vocab_size = vocab_size
|
182 |
self.hidden_size = hidden_size
|
@@ -186,72 +86,72 @@ class CustomTransformerConfig(PretrainedConfig):
|
|
186 |
self.prediction_chunk = prediction_chunk
|
187 |
self.max_position_embeddings = max_position_embeddings
|
188 |
self.input_size = prediction_chunk
|
|
|
189 |
|
190 |
class CustomTransformerModel(PreTrainedModel):
|
191 |
config_class = CustomTransformerConfig
|
192 |
|
193 |
def __init__(self, config):
|
194 |
super().__init__(config)
|
195 |
-
|
196 |
-
# Load pre-trained Llama model (excluding its original lm_head)
|
197 |
-
self.llama = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B", torch_dtype=torch.float16, device_map="auto", token = hf_token)
|
198 |
-
|
199 |
self.llama.resize_token_embeddings(config.vocab_size)
|
200 |
|
201 |
for i, layer in enumerate(self.llama.model.layers):
|
202 |
-
layer.self_attn = BidirectionalLlamaAttention(layer.self_attn, masking=
|
203 |
|
204 |
-
# Freeze Llama to retain pre-trained knowledge
|
205 |
for param in self.llama.parameters():
|
206 |
param.requires_grad = False
|
207 |
-
|
208 |
for param in self.llama.lm_head.parameters():
|
209 |
param.requires_grad = True
|
210 |
|
211 |
lora_config = LoraConfig(
|
212 |
-
r=512,
|
213 |
-
|
214 |
-
|
215 |
-
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], # Llama-3 uses these attention modules
|
216 |
-
bias="none",
|
217 |
-
task_type=None
|
218 |
)
|
219 |
|
220 |
self.llama = get_peft_model(self.llama, lora_config)
|
221 |
-
self.llama.print_trainable_parameters()
|
222 |
self.llama = self.llama.to(torch.float16)
|
223 |
|
224 |
-
|
225 |
def forward(self, input_ids, labels=None, **kwargs):
|
226 |
-
batch_size,
|
227 |
-
assert
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
if labels is not None:
|
242 |
-
assert labels.shape == (batch_size,
|
243 |
-
|
244 |
-
# Compute loss
|
245 |
-
loss_fct = torch.nn.CrossEntropyLoss()
|
246 |
-
|
247 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
248 |
|
249 |
-
|
250 |
return {"loss": loss, "logits": logits} if loss is not None else {"logits": logits}
|
251 |
|
252 |
-
|
253 |
def disable_dropout(model):
|
254 |
for name, module in model.named_modules():
|
255 |
if isinstance(module, nn.Dropout):
|
256 |
setattr(model, name, nn.Identity())
|
257 |
-
return model
|
|
|
1 |
+
import torch
|
2 |
import torch.nn as nn
|
|
|
|
|
3 |
from torch.amp import autocast
|
4 |
+
from transformers import AutoModelForCausalLM, PreTrainedModel, PretrainedConfig
|
5 |
+
from transformers.models.llama.modeling_llama import LlamaAttention
|
6 |
from peft import LoraConfig, get_peft_model
|
|
|
|
|
7 |
import os
|
8 |
+
from typing import Optional, Tuple
|
|
|
9 |
|
10 |
hf_token = os.getenv("HF_TOKEN")
|
11 |
|
12 |
class BidirectionalLlamaAttention(LlamaAttention):
|
13 |
+
def __init__(self, original_layer, masking='unidirectional'):
|
14 |
super().__init__(original_layer.config, layer_idx=original_layer.layer_idx)
|
15 |
self.masking = masking
|
|
|
|
|
16 |
self.q_proj.weight = original_layer.q_proj.weight
|
17 |
self.k_proj.weight = original_layer.k_proj.weight
|
18 |
self.v_proj.weight = original_layer.v_proj.weight
|
19 |
self.o_proj.weight = original_layer.o_proj.weight
|
20 |
|
21 |
def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
22 |
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
23 |
if n_rep == 1:
|
24 |
return hidden_states
|
25 |
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
|
|
26 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
27 |
|
28 |
+
def eager_attention_forward(self, module: nn.Module, query, key, value, attention_mask, scaling, dropout=0.0, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
key_states = self.repeat_kv(key, module.num_key_value_groups)
|
30 |
value_states = self.repeat_kv(value, module.num_key_value_groups)
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
32 |
+
|
33 |
if attention_mask is not None:
|
|
|
34 |
attn_mask = attention_mask.masked_fill(~attention_mask, float('-inf')).to(query.dtype)
|
35 |
attn_weights = attn_weights + attn_mask
|
36 |
|
|
|
37 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query.dtype)
|
38 |
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
39 |
+
attn_output = torch.matmul(attn_weights, value_states).transpose(1, 2).contiguous()
|
|
|
|
|
40 |
return attn_output, attn_weights
|
41 |
|
42 |
def rotate_half(self, x):
|
|
|
43 |
x1 = x[..., : x.shape[-1] // 2]
|
44 |
+
x2 = x[..., x.shape[-1] // 2:]
|
|
|
45 |
return torch.cat((-x2, x1), dim=-1)
|
46 |
|
47 |
+
def apply_rotary_pos_emb(self, q, k, cos, sin, unsqueeze_dim=1):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
cos = cos.unsqueeze(unsqueeze_dim)
|
49 |
sin = sin.unsqueeze(unsqueeze_dim)
|
50 |
q_embed = (q * cos) + (self.rotate_half(q) * sin)
|
51 |
k_embed = (k * cos) + (self.rotate_half(k) * sin)
|
|
|
52 |
return q_embed, k_embed
|
53 |
|
54 |
+
def forward(self, hidden_states, position_embeddings, attention_mask=None, past_key_value=None, cache_position=None, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
input_shape = hidden_states.shape[:-1]
|
56 |
hidden_shape = (*input_shape, -1, self.head_dim)
|
57 |
|
|
|
59 |
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
60 |
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
61 |
|
|
|
62 |
cos, sin = position_embeddings
|
63 |
query_states, key_states = self.apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
64 |
|
|
|
66 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
67 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
attn_output, attn_weights = self.eager_attention_forward(
|
70 |
+
self, query_states, key_states, value_states, attention_mask,
|
71 |
+
dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
)
|
73 |
|
74 |
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
75 |
+
return self.o_proj(attn_output), attn_weights
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
class CustomTransformerConfig(PretrainedConfig):
|
78 |
+
def __init__(self, vocab_size=128256, hidden_size=4096, num_layers=32, num_heads=32, prediction_chunk=256, dropout=0,
|
79 |
+
max_position_embeddings=4096, masking_type="bidirectional_masked", **kwargs):
|
80 |
super().__init__(**kwargs)
|
81 |
self.vocab_size = vocab_size
|
82 |
self.hidden_size = hidden_size
|
|
|
86 |
self.prediction_chunk = prediction_chunk
|
87 |
self.max_position_embeddings = max_position_embeddings
|
88 |
self.input_size = prediction_chunk
|
89 |
+
self.masking_type = masking_type
|
90 |
|
91 |
class CustomTransformerModel(PreTrainedModel):
|
92 |
config_class = CustomTransformerConfig
|
93 |
|
94 |
def __init__(self, config):
|
95 |
super().__init__(config)
|
96 |
+
self.llama = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-3B", torch_dtype=torch.float16, device_map="auto", token=hf_token)
|
|
|
|
|
|
|
97 |
self.llama.resize_token_embeddings(config.vocab_size)
|
98 |
|
99 |
for i, layer in enumerate(self.llama.model.layers):
|
100 |
+
layer.self_attn = BidirectionalLlamaAttention(layer.self_attn, masking=config.masking_type)
|
101 |
|
|
|
102 |
for param in self.llama.parameters():
|
103 |
param.requires_grad = False
|
|
|
104 |
for param in self.llama.lm_head.parameters():
|
105 |
param.requires_grad = True
|
106 |
|
107 |
lora_config = LoraConfig(
|
108 |
+
r=512, lora_alpha=512, lora_dropout=0.0,
|
109 |
+
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
|
110 |
+
bias="none", task_type=None
|
|
|
|
|
|
|
111 |
)
|
112 |
|
113 |
self.llama = get_peft_model(self.llama, lora_config)
|
114 |
+
self.llama.print_trainable_parameters()
|
115 |
self.llama = self.llama.to(torch.float16)
|
116 |
|
|
|
117 |
def forward(self, input_ids, labels=None, **kwargs):
|
118 |
+
batch_size, seq_len = input_ids.shape
|
119 |
+
assert seq_len == self.config.prediction_chunk, f"Expected input length {self.config.prediction_chunk}, got {seq_len}"
|
120 |
+
|
121 |
+
# Build attention mask
|
122 |
+
device = input_ids.device
|
123 |
+
if self.config.masking_type == 'bidirectional':
|
124 |
+
base_mask = torch.ones(seq_len, seq_len, dtype=torch.bool, device=device)
|
125 |
+
elif self.config.masking_type == 'bidirectional_masked':
|
126 |
+
base_mask = torch.ones(seq_len, seq_len, dtype=torch.bool, device=device)
|
127 |
+
base_mask.fill_diagonal_(False)
|
128 |
+
elif self.config.masking_type == 'unidirectional':
|
129 |
+
base_mask = torch.tril(torch.ones(seq_len, seq_len, dtype=torch.bool, device=device))
|
130 |
+
else:
|
131 |
+
raise ValueError(f"Unknown masking type: {self.config.masking_type}")
|
132 |
+
|
133 |
+
attention_mask = base_mask.unsqueeze(0).unsqueeze(1).expand(batch_size, 1, seq_len, seq_len).clone()
|
134 |
+
|
135 |
+
with autocast("cuda", dtype=torch.float16):
|
136 |
+
outputs = self.llama(
|
137 |
+
input_ids,
|
138 |
+
attention_mask=attention_mask,
|
139 |
+
output_hidden_states=True,
|
140 |
+
**kwargs
|
141 |
+
)
|
142 |
+
|
143 |
+
logits = outputs.logits[:, :, :self.config.vocab_size].view(batch_size, seq_len, self.config.vocab_size)
|
144 |
+
|
145 |
+
loss = None
|
146 |
if labels is not None:
|
147 |
+
assert labels.shape == (batch_size, seq_len), f"Labels shape mismatch: expected ({batch_size}, {seq_len}), got {labels.shape}"
|
148 |
+
loss_fct = nn.CrossEntropyLoss()
|
|
|
|
|
|
|
149 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
150 |
|
|
|
151 |
return {"loss": loss, "logits": logits} if loss is not None else {"logits": logits}
|
152 |
|
|
|
153 |
def disable_dropout(model):
|
154 |
for name, module in model.named_modules():
|
155 |
if isinstance(module, nn.Dropout):
|
156 |
setattr(model, name, nn.Identity())
|
157 |
+
return model
|