Create handler.py
Browse files- handler.py +62 -0
handler.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
import requests
|
4 |
+
from io import BytesIO
|
5 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
6 |
+
from qwen_vl_utils import process_vision_info
|
7 |
+
|
8 |
+
class EndpointHandler:
|
9 |
+
def __init__(self, path=""):
|
10 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
11 |
+
|
12 |
+
# Load model and processor
|
13 |
+
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
|
14 |
+
path, torch_dtype="auto", device_map="auto"
|
15 |
+
)
|
16 |
+
self.processor = AutoProcessor.from_pretrained(path)
|
17 |
+
|
18 |
+
def __call__(self, data):
|
19 |
+
# Extract image and text from the input data
|
20 |
+
image_url = data.get("image", "")
|
21 |
+
text_prompt = data.get("text", "")
|
22 |
+
|
23 |
+
# Download and process the image
|
24 |
+
image = Image.open(BytesIO(requests.get(image_url).content))
|
25 |
+
|
26 |
+
# Prepare the input in the format expected by the model
|
27 |
+
messages = [
|
28 |
+
{
|
29 |
+
"role": "user",
|
30 |
+
"content": [
|
31 |
+
{"type": "image", "image": image},
|
32 |
+
{"type": "text", "text": text_prompt},
|
33 |
+
],
|
34 |
+
}
|
35 |
+
]
|
36 |
+
|
37 |
+
# Process the input
|
38 |
+
text = self.processor.apply_chat_template(
|
39 |
+
messages, tokenize=False, add_generation_prompt=True
|
40 |
+
)
|
41 |
+
image_inputs, video_inputs = process_vision_info(messages)
|
42 |
+
inputs = self.processor(
|
43 |
+
text=[text],
|
44 |
+
images=image_inputs,
|
45 |
+
videos=video_inputs,
|
46 |
+
padding=True,
|
47 |
+
return_tensors="pt",
|
48 |
+
)
|
49 |
+
|
50 |
+
# Move inputs to the appropriate device
|
51 |
+
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
52 |
+
|
53 |
+
# Generate output
|
54 |
+
with torch.no_grad():
|
55 |
+
output_ids = self.model.generate(**inputs, max_new_tokens=128)
|
56 |
+
|
57 |
+
# Decode the output
|
58 |
+
output_text = self.processor.batch_decode(
|
59 |
+
output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
60 |
+
)[0]
|
61 |
+
|
62 |
+
return {"generated_text": output_text}
|