Creating a handler.py file to support HF dedicated inference endpoints
Browse filesHi all,
Made a quick handler.py file for people to use. It works for me pretty well. Feel free to test it out :)
Here is an example for when making a request to the endpoint:
import requests
import base64
import json
# Endpoint URL
api_url = 'your API URL'
# Path to the image file you want to test with
image_path = '' # Replace with the path to your image file
# The question to send to the API
question = 'what is going on in this picture?'
# Open the image file in binary mode and encode it in base64
with open(image_path, 'rb') as img:
encoded_image = base64.b64encode(img.read()).decode('utf-8')
# Prepare the JSON payload
payload = {
'inputs': {
'image': encoded_image,
'question': question
}
}
# Set the headers
headers = {
"Accept" : "application/json",
"Authorization": "Bearer HUGGINGFACE_TOKEN",
"Content-Type": "application/json"
}
# Send the POST request
response = requests.post(api_url, headers=headers, data=json.dumps(payload))
# Print the response from the server
print(response.json().get('body', {}))
- handler.py +58 -0
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
2 |
+
from PIL import Image
|
3 |
+
import torch
|
4 |
+
from io import BytesIO
|
5 |
+
import base64
|
6 |
+
|
7 |
+
class EndpointHandler:
|
8 |
+
def __init__(self, model_dir):
|
9 |
+
self.model_id = "zesquirrelnator/moondream2-finetuneV2"
|
10 |
+
self.model = AutoModelForCausalLM.from_pretrained(self.model_id, trust_remote_code=True)
|
11 |
+
self.tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2", trust_remote_code=True)
|
12 |
+
|
13 |
+
# Check if CUDA (GPU support) is available and then set the device to GPU or CPU
|
14 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
15 |
+
self.model.to(self.device)
|
16 |
+
|
17 |
+
def preprocess_image(self, encoded_image):
|
18 |
+
"""Decode and preprocess the input image."""
|
19 |
+
decoded_image = base64.b64decode(encoded_image)
|
20 |
+
img = Image.open(BytesIO(decoded_image)).convert("RGB")
|
21 |
+
return img
|
22 |
+
|
23 |
+
def __call__(self, data):
|
24 |
+
"""Handle the incoming request."""
|
25 |
+
try:
|
26 |
+
# Extract the inputs from the data
|
27 |
+
inputs = data.pop("inputs", data)
|
28 |
+
input_image = inputs['image']
|
29 |
+
question = inputs.get('question', "move to the red ball")
|
30 |
+
|
31 |
+
# Preprocess the image
|
32 |
+
img = self.preprocess_image(input_image)
|
33 |
+
|
34 |
+
# Perform inference
|
35 |
+
enc_image = self.model.encode_image(img).to(self.device)
|
36 |
+
answer = self.model.answer_question(enc_image, question, self.tokenizer)
|
37 |
+
|
38 |
+
# If the output is a tensor, move it back to CPU and convert to list
|
39 |
+
if isinstance(answer, torch.Tensor):
|
40 |
+
answer = answer.cpu().numpy().tolist()
|
41 |
+
|
42 |
+
# Create the response
|
43 |
+
response = {
|
44 |
+
"statusCode": 200,
|
45 |
+
"body": {
|
46 |
+
"answer": answer
|
47 |
+
}
|
48 |
+
}
|
49 |
+
return response
|
50 |
+
except Exception as e:
|
51 |
+
# Handle any errors
|
52 |
+
response = {
|
53 |
+
"statusCode": 500,
|
54 |
+
"body": {
|
55 |
+
"error": str(e)
|
56 |
+
}
|
57 |
+
}
|
58 |
+
return response
|