Spaces:
Runtime error
Runtime error
Upload main.py
Browse files
main.py
ADDED
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import pipeline, BitsAndBytesConfig
|
3 |
+
from fastapi import FastAPI, HTTPException
|
4 |
+
from pydantic import BaseModel
|
5 |
+
import requests
|
6 |
+
from PIL import Image
|
7 |
+
from io import BytesIO
|
8 |
+
|
9 |
+
# Set up device (CPU or GPU)
|
10 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
11 |
+
|
12 |
+
# Configure quantization if using GPU
|
13 |
+
if device == "cuda":
|
14 |
+
print("GPU found. Using 4-bit quantization.")
|
15 |
+
quantization_config = BitsAndBytesConfig(
|
16 |
+
load_in_4bit=True,
|
17 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
18 |
+
)
|
19 |
+
else:
|
20 |
+
print("GPU not found. Using CPU with default settings.")
|
21 |
+
quantization_config = None
|
22 |
+
|
23 |
+
# Load model pipeline
|
24 |
+
model_id = "bczhou/tiny-llava-v1-hf"
|
25 |
+
pipe = pipeline("image-to-text", model=model_id, device=device)
|
26 |
+
|
27 |
+
print(f"Using device: {device}")
|
28 |
+
|
29 |
+
# Initialize FastAPI application
|
30 |
+
app = FastAPI()
|
31 |
+
|
32 |
+
# Health check endpoint to ensure API is running
|
33 |
+
@app.get("/")
|
34 |
+
async def root():
|
35 |
+
return {"message": "API is running fine."}
|
36 |
+
|
37 |
+
# Define Pydantic model for request input
|
38 |
+
class ImagePromptInput(BaseModel):
|
39 |
+
image_url: str
|
40 |
+
prompt: str
|
41 |
+
|
42 |
+
# FastAPI route for generating text from an image
|
43 |
+
@app.post("/generate")
|
44 |
+
async def generate_text(input_data: ImagePromptInput):
|
45 |
+
try:
|
46 |
+
# Download and process the image
|
47 |
+
response = requests.get(input_data.image_url)
|
48 |
+
image = Image.open(BytesIO(response.content)).convert("RGB")
|
49 |
+
image = image.resize((750, 500)) # Resize image to fixed dimensions
|
50 |
+
|
51 |
+
# Create a full prompt to pass to the model
|
52 |
+
full_prompt = f"USER: <image>\n{input_data.prompt}\nASSISTANT: "
|
53 |
+
|
54 |
+
# Generate response using the model pipeline
|
55 |
+
outputs = pipe(image, prompt=full_prompt, generate_kwargs={"max_new_tokens": 200})
|
56 |
+
|
57 |
+
# Return generated text
|
58 |
+
generated_text = outputs[0]['generated_text'] #type: ignore
|
59 |
+
return {"response": generated_text}
|
60 |
+
|
61 |
+
except Exception as e:
|
62 |
+
# Return error if something goes wrong
|
63 |
+
raise HTTPException(status_code=500, detail=str(e))
|