Create inference.py
Browse files- inference.py +101 -0
inference.py
ADDED
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
3 |
+
from utils import load_config
|
4 |
+
from tokenization import get_tokenizer
|
5 |
+
|
6 |
+
class CustomConfig(PretrainedConfig):
|
7 |
+
"""Configuration class for the custom language model."""
|
8 |
+
model_type = "custom_llm"
|
9 |
+
|
10 |
+
def __init__(
|
11 |
+
self,
|
12 |
+
vocab_size: int = 50000,
|
13 |
+
n_embd: int = 640,
|
14 |
+
n_head: int = 10,
|
15 |
+
n_layer: int = 12,
|
16 |
+
n_positions: int = 512,
|
17 |
+
tie_word_embeddings: bool = True,
|
18 |
+
**kwargs
|
19 |
+
):
|
20 |
+
self.vocab_size = vocab_size
|
21 |
+
self.n_embd = n_embd
|
22 |
+
self.n_head = n_head
|
23 |
+
self.n_layer = n_layer
|
24 |
+
self.n_positions = n_positions
|
25 |
+
self.tie_word_embeddings = tie_word_embeddings
|
26 |
+
super().__init__(**kwargs)
|
27 |
+
|
28 |
+
def generate_text(
|
29 |
+
prompt: str,
|
30 |
+
model_path: str = "outputs/hf_model",
|
31 |
+
max_length: int = 200,
|
32 |
+
temperature: float = 0.8,
|
33 |
+
top_k: int = 50,
|
34 |
+
top_p: float = 0.9,
|
35 |
+
repetition_penalty: float = 1.2,
|
36 |
+
no_repeat_ngram_size: int = 3
|
37 |
+
):
|
38 |
+
"""Generate text using the model."""
|
39 |
+
# Load config and tokenizer
|
40 |
+
config = load_config()
|
41 |
+
tokenizer = get_tokenizer(config)
|
42 |
+
|
43 |
+
# Load model
|
44 |
+
from inference import CustomModelForCausalLM # Import here to avoid circular imports
|
45 |
+
model = CustomModelForCausalLM.from_pretrained(model_path)
|
46 |
+
|
47 |
+
# Move model to GPU if available
|
48 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
49 |
+
model = model.to(device)
|
50 |
+
model.eval()
|
51 |
+
|
52 |
+
# Encode prompt
|
53 |
+
encoded = tokenizer.batch_encode(
|
54 |
+
[prompt],
|
55 |
+
return_tensors="pt"
|
56 |
+
)
|
57 |
+
input_ids = encoded["input_ids"].to(device)
|
58 |
+
|
59 |
+
# Generate
|
60 |
+
with torch.no_grad():
|
61 |
+
output_ids = model.generate(
|
62 |
+
input_ids=input_ids,
|
63 |
+
max_length=max_length,
|
64 |
+
temperature=temperature,
|
65 |
+
top_k=top_k,
|
66 |
+
top_p=top_p,
|
67 |
+
repetition_penalty=repetition_penalty,
|
68 |
+
no_repeat_ngram_size=no_repeat_ngram_size
|
69 |
+
)
|
70 |
+
|
71 |
+
# Decode and return
|
72 |
+
generated_text = tokenizer.decode(output_ids[0].tolist())
|
73 |
+
return generated_text
|
74 |
+
|
75 |
+
if __name__ == "__main__":
|
76 |
+
# Example prompts to test
|
77 |
+
prompts = [
|
78 |
+
"Once upon a time",
|
79 |
+
"The meaning of life is",
|
80 |
+
"In the distant future",
|
81 |
+
"The best way to learn programming is",
|
82 |
+
"Today I learned that"
|
83 |
+
]
|
84 |
+
|
85 |
+
print("\nGenerating text from multiple prompts:")
|
86 |
+
print("=" * 50)
|
87 |
+
|
88 |
+
for prompt in prompts:
|
89 |
+
generated_text = generate_text(
|
90 |
+
prompt=prompt,
|
91 |
+
max_length=200,
|
92 |
+
temperature=0.8, # Adjust for creativity (higher = more creative)
|
93 |
+
top_k=50, # Limit to top 50 tokens
|
94 |
+
top_p=0.9, # Nucleus sampling threshold
|
95 |
+
repetition_penalty=1.2, # Penalize repetition
|
96 |
+
no_repeat_ngram_size=3 # Prevent 3-gram repetition
|
97 |
+
)
|
98 |
+
|
99 |
+
print(f"\nPrompt: {prompt}")
|
100 |
+
print(f"Generated: {generated_text}")
|
101 |
+
print("-" * 50)
|