Commit
·
cc809e5
1
Parent(s):
1e04a23
Update README.md
Browse files
README.md
CHANGED
@@ -5,33 +5,75 @@ license: apache-2.0
|
|
5 |
|
6 |
## INFERENCE
|
7 |
|
8 |
-
```
|
9 |
-
|
|
|
10 |
import torch
|
11 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
12 |
|
13 |
-
finetuned_model = AutoModelForCausalLM.from_pretrained("AquilaX-AI/QnA")
|
14 |
tokenizer = AutoTokenizer.from_pretrained("AquilaX-AI/QnA")
|
|
|
15 |
|
16 |
-
|
|
|
|
|
17 |
|
18 |
-
|
19 |
-
|
20 |
|
21 |
-
|
22 |
-
""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
|
29 |
-
|
30 |
-
inputs = encodeds.to(device)
|
31 |
|
32 |
-
|
33 |
-
generated_ids = finetuned_model.generate(inputs, max_new_tokens=256, temperature=0.5, top_p=0.90, do_sample=True,pad_token_id=50259,eos_token_id=50259,num_return_sequences=1)
|
34 |
-
print(tokenizer.decode(generated_ids[0]).split('### Response:')[1].split('<eos>')[0].strip())
|
35 |
-
e = time.time()
|
36 |
-
print(f'time taken:{e-s}')
|
37 |
```
|
|
|
5 |
|
6 |
## INFERENCE
|
7 |
|
8 |
+
```python
|
9 |
+
# Load model directly
|
10 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
11 |
import torch
|
|
|
12 |
|
|
|
13 |
tokenizer = AutoTokenizer.from_pretrained("AquilaX-AI/QnA")
|
14 |
+
model = AutoModelForCausalLM.from_pretrained("AquilaX-AI/QnA")
|
15 |
|
16 |
+
prompt = """
|
17 |
+
<|im_start|>system\nYou are a helpful AI assistant named Securitron<|im_end|>
|
18 |
+
"""
|
19 |
|
20 |
+
# Keep a list for the last one conversation exchanges
|
21 |
+
conversation_history = []
|
22 |
|
23 |
+
while True:
|
24 |
+
user_prompt = input("\nUser Question: ")
|
25 |
+
if user_prompt.lower() == 'break':
|
26 |
+
break
|
27 |
+
|
28 |
+
# Format the user's input
|
29 |
+
user = f"""<|im_start|>user
|
30 |
+
{user_prompt}<|im_end|>
|
31 |
+
<|im_start|>assistant"""
|
32 |
+
|
33 |
+
# Add the user's question to the conversation history
|
34 |
+
conversation_history.append(user)
|
35 |
+
|
36 |
+
# Ensure conversation starts with a user's input and keep only the last 2 exchanges (4 turns)
|
37 |
+
conversation_history = conversation_history[-5:]
|
38 |
+
|
39 |
+
# Build the full prompt
|
40 |
+
current_prompt = prompt + "\n".join(conversation_history)
|
41 |
+
|
42 |
+
# Tokenize the prompt
|
43 |
+
encodeds = tokenizer(current_prompt, return_tensors="pt", truncation=True).input_ids
|
44 |
+
|
45 |
+
# Move model and inputs to the appropriate device
|
46 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
47 |
+
model.to(device)
|
48 |
+
inputs = encodeds.to(device)
|
49 |
+
|
50 |
+
# Create an empty list to store generated tokens
|
51 |
+
generated_ids = inputs
|
52 |
|
53 |
+
# Start generating tokens one by one
|
54 |
+
assistant_response = ""
|
55 |
+
for _ in range(512): # Specify a max token limit for streaming
|
56 |
+
next_token = model.generate(
|
57 |
+
generated_ids,
|
58 |
+
max_new_tokens=1,
|
59 |
+
pad_token_id=151644,
|
60 |
+
eos_token_id=151645,
|
61 |
+
num_return_sequences=1,
|
62 |
+
do_sample=False,
|
63 |
+
# top_k=5,
|
64 |
+
# temperature=0.2,
|
65 |
+
# top_p=0.90
|
66 |
+
)
|
67 |
+
|
68 |
+
generated_ids = torch.cat([generated_ids, next_token[:, -1:]], dim=1)
|
69 |
+
token_id = next_token[0, -1].item()
|
70 |
+
token = tokenizer.decode([token_id], skip_special_tokens=True)
|
71 |
+
|
72 |
+
assistant_response += token
|
73 |
+
print(token, end="", flush=True)
|
74 |
|
75 |
+
if token_id == 151645: # EOS token
|
76 |
+
break
|
|
|
77 |
|
78 |
+
conversation_history.append(f"{assistant_response.strip()}<|im_end|>")
|
|
|
|
|
|
|
|
|
79 |
```
|