Create handler.py
Browse files- handler.py +59 -0
handler.py
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
import base64
|
4 |
+
from io import BytesIO
|
5 |
+
from transformers import AutoModel, AutoTokenizer
|
6 |
+
|
7 |
+
class EndpointHandler:
|
8 |
+
def __init__(self, path="/repository"):
|
9 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
10 |
+
|
11 |
+
# Load the model
|
12 |
+
self.model = AutoModel.from_pretrained(
|
13 |
+
path,
|
14 |
+
trust_remote_code=True,
|
15 |
+
attn_implementation='sdpa',
|
16 |
+
torch_dtype=torch.bfloat16 if self.device.type == "cuda" else torch.float32,
|
17 |
+
).to(self.device)
|
18 |
+
self.model.eval()
|
19 |
+
|
20 |
+
# Load the tokenizer
|
21 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
22 |
+
path,
|
23 |
+
trust_remote_code=True,
|
24 |
+
)
|
25 |
+
|
26 |
+
def __call__(self, data):
|
27 |
+
# Extract image and text from the input data
|
28 |
+
image_data = data.get("inputs", {}).get("image", "")
|
29 |
+
text_prompt = data.get("inputs", {}).get("text", "")
|
30 |
+
|
31 |
+
if not image_data or not text_prompt:
|
32 |
+
return {"error": "Both 'image' and 'text' must be provided in the input data."}
|
33 |
+
|
34 |
+
# Process the image data
|
35 |
+
try:
|
36 |
+
image_bytes = base64.b64decode(image_data)
|
37 |
+
image = Image.open(BytesIO(image_bytes)).convert("RGB")
|
38 |
+
except Exception as e:
|
39 |
+
return {"error": f"Failed to process image data: {e}"}
|
40 |
+
|
41 |
+
# Prepare the messages for the model
|
42 |
+
msgs = [{'role': 'user', 'content': [image, text_prompt]}]
|
43 |
+
|
44 |
+
# Generate output
|
45 |
+
with torch.no_grad():
|
46 |
+
res = self.model.chat(
|
47 |
+
image=None,
|
48 |
+
msgs=msgs,
|
49 |
+
tokenizer=self.tokenizer,
|
50 |
+
sampling=True,
|
51 |
+
temperature=0.7,
|
52 |
+
top_p=0.95,
|
53 |
+
max_length=2000,
|
54 |
+
)
|
55 |
+
|
56 |
+
# The result is the generated text
|
57 |
+
output_text = res
|
58 |
+
|
59 |
+
return {"generated_text": output_text}
|