app.py

from fastapi import FastAPI, UploadFile, File
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import torch
import io
import asyncio
import time
from contextlib import asynccontextmanager
from typing import List, Tuple

# Configuration
MODEL_ID = "openai/clip-vit-large-patch14"
BATCH_SIZE = 32
BATCH_TIMEOUT = 0.05  # 50ms wait to fill batch
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32

# Global State
model = None
processor = None
request_queue = asyncio.Queue()

class SmartBatcher:
    """
    Collects individual inference requests and processes them in optimal batches.
    """
    def __init__(self):
        self.loop = asyncio.get_event_loop()
        self.processing_task = None

    def start(self):
        self.processing_task = self.loop.create_task(self.process_batches())
        print("🚀 Smart Batcher started.")

    async def process_batches(self):
        while True:
            # 1. Collect Requests
            batch = []
            
            # Wait for first item
            item = await request_queue.get()
            batch.append(item)
            
            # Try to fill batch within timeout window
            start_wait = time.time()
            while len(batch) < BATCH_SIZE:
                # Calculate remaining time in timeout window
                remaining = BATCH_TIMEOUT - (time.time() - start_wait)
                if remaining <= 0:
                    break
                
                try:
                    # Non-blocking check for more items
                    # We use wait_for to respect the timeout window
                    additional_item = await asyncio.wait_for(request_queue.get(), timeout=remaining)
                    batch.append(additional_item)
                except asyncio.TimeoutError:
                    break
                except Exception:
                    break
            
            # 2. Process Batch
            if batch:
                await self.run_inference(batch)

    async def run_inference(self, batch: List[Tuple]):
        # Unpack batch: [(input_data, type, future), ...]
        text_inputs = []
        image_inputs = []
        
        # Sort indices to maintain order mapping
        # batch structure: (data, 'text'|'image', future)
        
        for i, (data, kind, fut) in enumerate(batch):
            if kind == 'text':
                text_inputs.append((i, data, fut))
            elif kind == 'image':
                image_inputs.append((i, data, fut))

        # Run Text Batch
        if text_inputs:
            texts = [t[1] for t in text_inputs]
            try:
                # Prepare Inputs
                inputs = processor(
                    text=texts, 
                    padding=True, 
                    truncation=True,
                    return_tensors="pt"
                ).to(DEVICE)

                # Inference
                with torch.inference_mode():
                    outputs = model.get_text_features(**inputs)
                    outputs = outputs / outputs.norm(dim=-1, keepdim=True)
                    vectors = outputs.cpu().tolist()

                # Distribute Results
                for j, vector in enumerate(vectors):
                    original_idx, _, fut = text_inputs[j]
                    if not fut.done():
                        fut.set_result(vector)
            except Exception as e:
                for _, _, fut in text_inputs:
                    if not fut.done():
                        fut.set_exception(e)

        # Run Image Batch
        if image_inputs:
            images = [t[1] for t in image_inputs]
            try:
                # Prepare Inputs
                inputs = processor(images=images, return_tensors="pt").to(DEVICE)

                # Inference
                with torch.inference_mode():
                    outputs = model.get_image_features(**inputs)
                    outputs = outputs / outputs.norm(dim=-1, keepdim=True)
                    vectors = outputs.cpu().tolist()

                # Distribute Results
                for j, vector in enumerate(vectors):
                    original_idx, _, fut = image_inputs[j]
                    if not fut.done():
                        fut.set_result(vector)
            except Exception as e:
                for _, _, fut in image_inputs:
                    if not fut.done():
                        fut.set_exception(e)


@asynccontextmanager
async def lifespan(app: FastAPI):
    global model, processor
    print("🧠 Loading CLIP Model...")
    
    # Load Model
    model = CLIPModel.from_pretrained(
        MODEL_ID, 
        torch_dtype=DTYPE, 
        low_cpu_mem_usage=True
    ).to(DEVICE).eval()
    
    # Compile model for faster inference (Linux/CUDA mostly, graceful fallback)
    try:
        model = torch.compile(model)
        print("⚡ Torch Compile enabled.")
    except Exception:
        print("⚠️ Torch Compile skipped (not supported).")

    processor = CLIPProcessor.from_pretrained(MODEL_ID)
    
    # Start Batcher
    batcher = SmartBatcher()
    batcher.start()
    
    yield
    print("🛑 Shutting down.")

app = FastAPI(lifespan=lifespan)

@app.post("/embed-text")
async def embed_text(text: str):
    loop = asyncio.get_running_loop()
    fut = loop.create_future()
    
    await request_queue.put((text, 'text', fut))
    
    # Wait for batch processor to set result
    result = await fut
    return {"vector": result}

@app.post("/embed-image")
async def embed_image(file: UploadFile = File(...)):
    # Read image immediately to avoid holding file handle in queue too long
    content = await file.read()
    image = Image.open(io.BytesIO(content)).convert("RGB")
    
    loop = asyncio.get_running_loop()
    fut = loop.create_future()
    
    await request_queue.put((image, 'image', fut))
    
    result = await fut
    return {"vector": result}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8001)