Vladislav Sokolovskii
commited on
Commit
•
e97bce5
1
Parent(s):
7c046a6
Add handler and reqs
Browse files- handler.py +71 -0
- requirements.txt +8 -0
handler.py
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from typing import Dict, List, Any
|
3 |
+
from unsloth import FastLanguageModel
|
4 |
+
from unsloth.chat_templates import get_chat_template
|
5 |
+
import torch
|
6 |
+
from huggingface_hub import login
|
7 |
+
import os
|
8 |
+
|
9 |
+
class EndpointHandler:
|
10 |
+
def __init__(self, path=""):
|
11 |
+
# access_token = os.environ["HUGGINGFACE_TOKEN"]
|
12 |
+
# login(token=access_token)
|
13 |
+
# Load the model and tokenizer
|
14 |
+
self.model, self.tokenizer = FastLanguageModel.from_pretrained(
|
15 |
+
model_name = path, # Use the current directory path
|
16 |
+
max_seq_length = 2048,
|
17 |
+
dtype = None,
|
18 |
+
load_in_4bit = True,
|
19 |
+
)
|
20 |
+
FastLanguageModel.for_inference(self.model)
|
21 |
+
|
22 |
+
# Set up the chat template
|
23 |
+
self.tokenizer = get_chat_template(
|
24 |
+
self.tokenizer,
|
25 |
+
chat_template="llama-3",
|
26 |
+
mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"}
|
27 |
+
)
|
28 |
+
|
29 |
+
def __call__(self, data: Dict[str, Any]) -> List[str]:
|
30 |
+
inputs = data.pop("inputs", data)
|
31 |
+
parameters = data.pop("parameters", {})
|
32 |
+
|
33 |
+
# Extract parameters or use defaults
|
34 |
+
max_tokens = parameters.get("max_new_tokens", 512)
|
35 |
+
temperature = parameters.get("temperature", 0.2)
|
36 |
+
top_p = parameters.get("top_p", 0.5)
|
37 |
+
system_message = parameters.get("system_message", "")
|
38 |
+
|
39 |
+
# Prepare messages
|
40 |
+
messages = [{"from": "human", "value": system_message}]
|
41 |
+
if isinstance(inputs, str):
|
42 |
+
messages.append({"from": "human", "value": inputs})
|
43 |
+
elif isinstance(inputs, list):
|
44 |
+
for msg in inputs:
|
45 |
+
role = "human" if msg["role"] == "user" else "gpt"
|
46 |
+
messages.append({"from": role, "value": msg["content"]})
|
47 |
+
|
48 |
+
# Tokenize input
|
49 |
+
tokenized_input = self.tokenizer.apply_chat_template(
|
50 |
+
messages,
|
51 |
+
tokenize=True,
|
52 |
+
add_generation_prompt=True,
|
53 |
+
return_tensors="pt"
|
54 |
+
).to("cuda")
|
55 |
+
|
56 |
+
# Generate output
|
57 |
+
with torch.no_grad():
|
58 |
+
output = self.model.generate(
|
59 |
+
input_ids=tokenized_input,
|
60 |
+
max_new_tokens=max_tokens,
|
61 |
+
temperature=temperature,
|
62 |
+
top_p=top_p,
|
63 |
+
use_cache=True
|
64 |
+
)
|
65 |
+
|
66 |
+
# Decode and process the output
|
67 |
+
full_response = self.tokenizer.decode(output[0], skip_special_tokens=True)
|
68 |
+
response_lines = [line.strip() for line in full_response.split('\n') if line.strip()]
|
69 |
+
last_response = response_lines[-1] if response_lines else ""
|
70 |
+
|
71 |
+
return [last_response]
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torchvision
|
2 |
+
xformers<0.0.27
|
3 |
+
trl==0.8.6
|
4 |
+
transformers==4.44.2
|
5 |
+
bitsandbytes==0.43.3
|
6 |
+
peft==0.12.0
|
7 |
+
accelerate>=0.34.2
|
8 |
+
git+https://github.com/unslothai/unsloth.git
|