Create modeling_DiffusionLLM.py
Browse files- modeling_DiffusionLLM.py +126 -0
modeling_DiffusionLLM.py
ADDED
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
4 |
+
|
5 |
+
class DiffusionConfig(PretrainedConfig):
|
6 |
+
"""Configuration class for Diffusion-LLM model."""
|
7 |
+
model_type = "diffusionLM"
|
8 |
+
|
9 |
+
def __init__(
|
10 |
+
self,
|
11 |
+
vocab_size: int = 50257,
|
12 |
+
hidden_size: int = 768,
|
13 |
+
num_hidden_layers: int = 12,
|
14 |
+
num_attention_heads: int = 12,
|
15 |
+
intermediate_size: int = 3072,
|
16 |
+
hidden_dropout_prob: float = 0.1,
|
17 |
+
attention_probs_dropout_prob: float = 0.1,
|
18 |
+
max_position_embeddings: int = 1024,
|
19 |
+
initializer_range: float = 0.02,
|
20 |
+
layer_norm_eps: float = 1e-12,
|
21 |
+
pad_token_id: int = 0,
|
22 |
+
mask_token_id: int = 50256,
|
23 |
+
eos_token_id: int = 50256,
|
24 |
+
num_timesteps: int = 100,
|
25 |
+
time_embed_dim: int = 128,
|
26 |
+
**kwargs
|
27 |
+
):
|
28 |
+
super().__init__(pad_token_id=pad_token_id, **kwargs)
|
29 |
+
self.vocab_size = vocab_size
|
30 |
+
self.hidden_size = hidden_size
|
31 |
+
self.num_hidden_layers = num_hidden_layers
|
32 |
+
self.num_attention_heads = num_attention_heads
|
33 |
+
self.intermediate_size = intermediate_size
|
34 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
35 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
36 |
+
self.max_position_embeddings = max_position_embeddings
|
37 |
+
self.initializer_range = initializer_range
|
38 |
+
self.layer_norm_eps = layer_norm_eps
|
39 |
+
self.mask_token_id = mask_token_id
|
40 |
+
self.eos_token_id = eos_token_id
|
41 |
+
self.num_timesteps = num_timesteps
|
42 |
+
self.time_embed_dim = time_embed_dim
|
43 |
+
|
44 |
+
class DiffusionLLM(PreTrainedModel):
|
45 |
+
"""Main Diffusion-LLM model class"""
|
46 |
+
config_class = DiffusionConfig
|
47 |
+
base_model_prefix = "diffusionLM"
|
48 |
+
|
49 |
+
def __init__(self, config: DiffusionConfig):
|
50 |
+
super().__init__(config)
|
51 |
+
self.model = LLaDAModel(config)
|
52 |
+
self.init_weights()
|
53 |
+
|
54 |
+
def forward(
|
55 |
+
self,
|
56 |
+
input_ids=None,
|
57 |
+
attention_mask=None,
|
58 |
+
timesteps=None,
|
59 |
+
labels=None,
|
60 |
+
return_dict=True,
|
61 |
+
):
|
62 |
+
outputs = self.model(
|
63 |
+
input_ids=input_ids,
|
64 |
+
attention_mask=attention_mask,
|
65 |
+
timesteps=timesteps,
|
66 |
+
labels=labels,
|
67 |
+
)
|
68 |
+
|
69 |
+
return outputs
|
70 |
+
|
71 |
+
def generate(
|
72 |
+
self,
|
73 |
+
prompt=None,
|
74 |
+
max_length=100,
|
75 |
+
num_inference_steps=50,
|
76 |
+
temperature=1.0,
|
77 |
+
strategy='random',
|
78 |
+
top_p=0.9,
|
79 |
+
top_k=50,
|
80 |
+
num_beams=5,
|
81 |
+
return_scores=False,
|
82 |
+
use_streaming=False,
|
83 |
+
callback_fn=None
|
84 |
+
):
|
85 |
+
"""Unified generation interface"""
|
86 |
+
if use_streaming:
|
87 |
+
return self.generate_stream(
|
88 |
+
prompt=prompt,
|
89 |
+
max_length=max_length,
|
90 |
+
num_inference_steps=num_inference_steps,
|
91 |
+
temperature=temperature,
|
92 |
+
strategy=strategy,
|
93 |
+
top_p=top_p,
|
94 |
+
top_k=top_k,
|
95 |
+
num_beams=num_beams,
|
96 |
+
callback_fn=callback_fn
|
97 |
+
)
|
98 |
+
else:
|
99 |
+
return self.model.generate(
|
100 |
+
prompt=prompt,
|
101 |
+
max_length=max_length,
|
102 |
+
num_inference_steps=num_inference_steps,
|
103 |
+
temperature=temperature,
|
104 |
+
strategy=strategy,
|
105 |
+
top_p=top_p,
|
106 |
+
top_k=top_k,
|
107 |
+
num_beams=num_beams,
|
108 |
+
return_scores=return_scores
|
109 |
+
)
|
110 |
+
|
111 |
+
def generate_stream(self, **kwargs):
|
112 |
+
"""Streaming generation wrapper"""
|
113 |
+
return self.model.generate_stream(**kwargs)
|
114 |
+
|
115 |
+
def prepare_inputs_for_generation(self, input_ids, **kwargs):
|
116 |
+
"""Prepare inputs for generation compatibility"""
|
117 |
+
return {
|
118 |
+
"input_ids": input_ids,
|
119 |
+
"attention_mask": kwargs.get("attention_mask", None),
|
120 |
+
"timesteps": kwargs.get("timesteps", None),
|
121 |
+
}
|
122 |
+
|
123 |
+
@staticmethod
|
124 |
+
def _reorder_cache(past, beam_idx):
|
125 |
+
"""Reorder cache for beam search compatibility"""
|
126 |
+
return past
|