Deepseek VL2 Web API (Modified based on web_demo.py from the official repo)

web_api.py

# Copyright (c) 2023-2024 DeepSeek.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy of
# this software and associated documentation files (the "Software"), to deal in
# the Software without restriction, including without limitation the rights to
# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
# the Software, and to permit persons to whom the Software is furnished to do so,
# subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

# -*- coding:utf-8 -*-
from argparse import ArgumentParser
import base64
import io
import os
import sys
import time
import json
from typing import List, Dict, Any, Optional, Union, AsyncGenerator
import uuid

import torch
from fastapi import FastAPI, Request, HTTPException
from fastapi.responses import JSONResponse, StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
import uvicorn
from PIL import Image

# Import DeepSeek VL2 components
from deepseek_vl2.serve.app_modules.utils import configure_logger, strip_stop_words
from deepseek_vl2.serve.inference import (
    convert_conversation_to_prompts,
    deepseek_generate,
    load_model,
)

logger = configure_logger()

MODELS = [
    "DeepSeek-VL2-tiny",
    "DeepSeek-VL2-small",
    "DeepSeek-VL2",
    "deepseek-ai/deepseek-vl2-tiny",
    "deepseek-ai/deepseek-vl2-small",
    "deepseek-ai/deepseek-vl2",
]

IMAGE_TOKEN = "<image>"

# Create FastAPI app
app = FastAPI(title="DeepSeek-VL2 API")

# Add CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Global model storage
MODEL_CACHE = {}

def get_model(model_name):
    """Get or load model from cache"""
    global MODEL_CACHE, args
    
    if model_name in MODEL_CACHE:
        return MODEL_CACHE[model_name]
    
    model_path = args.local_path if args.local_path else model_name
    
    print(f"Loading DeepSeek-VL2 model from {model_path}...")
    tokenizer, vl_gpt, vl_chat_processor = load_model(model_path)
    MODEL_CACHE[model_name] = (tokenizer, vl_gpt, vl_chat_processor)
    print(f"Model loaded successfully.")
    
    return tokenizer, vl_gpt, vl_chat_processor

def process_messages(messages):
    """Extract text and images from messages"""
    text = ""
    images = []
    
    # Process all messages, prioritizing the last user message
    user_messages = [msg for msg in messages if msg["role"] == "user"]
    if not user_messages:
        return text, images
    
    last_message = user_messages[-1]
    content = last_message["content"]
    
    if isinstance(content, str):
        text = content
    elif isinstance(content, list):
        for part in content:
            if part.get("type") == "text":
                text += part.get("text", "")
            elif part.get("type") == "image_url":
                image_data = part.get("image_url", {})
                if "url" in image_data:
                    url = image_data["url"]
                    if url.startswith("data:image"):
                        try:
                            base64_data = url.split(",")[1]
                            img_bytes = base64.b64decode(base64_data)
                            img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
                            images.append(img)
                        except Exception as e:
                            logger.error(f"Error processing image: {e}")
    
    return text, images

@app.post("/v1/chat/completions")
async def chat_completion(request: Request):
    """OpenAI-compatible chat completion endpoint"""
    try:
        body = await request.json()
        
        # Extract parameters
        model = body.get("model", args.model_name)
        messages = body.get("messages", [])
        temperature = float(body.get("temperature", 0.1))
        top_p = float(body.get("top_p", 0.9))
        max_tokens = int(body.get("max_tokens", 2048))
        stream = bool(body.get("stream", False))
        repetition_penalty = float(body.get("repetition_penalty", 1.1))
        
        if not model in MODELS:
            raise HTTPException(
                status_code=400, 
                detail=f"Model {model} not found. Available models: {', '.join(MODELS)}"
            )
        
        # Get model
        tokenizer, vl_gpt, vl_chat_processor = get_model(model)
        
        # Process images and text
        text, images = process_messages(messages)
        
        if not text:
            raise HTTPException(status_code=400, detail="No text content provided")
        
        # Format the prompt
        if images and len(images) > 0:
            num_images = len(images)
            image_tokens = "\n".join([IMAGE_TOKEN] * num_images)
            text = image_tokens + "\n" + text
            text_with_images = (text, images)
        else:
            text_with_images = text
        
        # Initialize conversation
        conversation = vl_chat_processor.new_chat_template()
        conversation.append_message(conversation.roles[0], text_with_images)
        conversation.append_message(conversation.roles[1], "")
        
        # Convert to format expected by DeepSeek generator
        all_conv, last_image = convert_conversation_to_prompts(conversation)
        stop_words = conversation.stop_str
        
        response_id = f"chatcmpl-{uuid.uuid4().hex}"
        created_time = int(time.time())
        
        if stream:
            # Handle streaming response
            async def generate_stream():
                # Start event - empty role delta
                yield f"data: {json.dumps({'id': response_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model, 'choices': [{'index': 0, 'delta': {'role': 'assistant'}, 'finish_reason': None}]})}\n\n"
                
                full_response = ""
                with torch.no_grad():
                    for chunk in deepseek_generate(
                        conversations=all_conv,
                        vl_gpt=vl_gpt,
                        vl_chat_processor=vl_chat_processor,
                        tokenizer=tokenizer,
                        stop_words=stop_words,
                        max_length=max_tokens,
                        temperature=temperature,
                        repetition_penalty=repetition_penalty,
                        top_p=top_p,
                        chunk_size=args.chunk_size
                    ):
                        full_response += chunk
                        
                        # Send chunk
                        chunk_data = {
                            'id': response_id,
                            'object': 'chat.completion.chunk',
                            'created': created_time,
                            'model': model,
                            'choices': [{'index': 0, 'delta': {'content': chunk}, 'finish_reason': None}]
                        }
                        yield f"data: {json.dumps(chunk_data)}\n\n"
                
                # End event - empty delta with finish reason
                yield f"data: {json.dumps({'id': response_id, 'object': 'chat.completion.chunk', 'created': created_time, 'model': model, 'choices': [{'index': 0, 'delta': {}, 'finish_reason': 'stop'}]})}\n\n"
                yield "data: [DONE]\n\n"
                
                # Clean up
                torch.cuda.empty_cache()
            
            return StreamingResponse(generate_stream(), media_type="text/event-stream")
        else:
            # Non-streaming response
            full_response = ""
            with torch.no_grad():
                for chunk in deepseek_generate(
                    conversations=all_conv,
                    vl_gpt=vl_gpt,
                    vl_chat_processor=vl_chat_processor,
                    tokenizer=tokenizer,
                    stop_words=stop_words,
                    max_length=max_tokens,
                    temperature=temperature,
                    repetition_penalty=repetition_penalty,
                    top_p=top_p,
                    chunk_size=args.chunk_size
                ):
                    full_response += chunk
            
            response = strip_stop_words(full_response, stop_words)
            conversation.update_last_message(response)
            
            # Simple token counting
            prompt_tokens = len(text) // 4
            completion_tokens = len(response) // 4
            
            # Clean up
            torch.cuda.empty_cache()
            
            # Return OpenAI-style response
            return {
                "id": response_id,
                "object": "chat.completion",
                "created": created_time,
                "model": model,
                "choices": [
                    {
                        "index": 0,
                        "message": {
                            "role": "assistant",
                            "content": response
                        },
                        "finish_reason": "stop"
                    }
                ],
                "usage": {
                    "prompt_tokens": prompt_tokens,
                    "completion_tokens": completion_tokens,
                    "total_tokens": prompt_tokens + completion_tokens
                }
            }
    
    except Exception as e:
        logger.error(f"Error in chat_completion: {str(e)}")
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/v1/models")
async def list_models():
    """List available models endpoint"""
    models_list = [{"id": model, "object": "model"} for model in MODELS]
    return {"object": "list", "data": models_list}

@app.get("/health")
async def health_check():
    """Health check endpoint"""
    return {"status": "ok"}

if __name__ == "__main__":
    parser = ArgumentParser()
    parser.add_argument("--model_name", type=str, default="deepseek-ai/deepseek-vl2-tiny", required=False, choices=MODELS, help="model name")
    parser.add_argument("--local_path", type=str, default="", help="huggingface ckpt, optional")
    parser.add_argument("--host", type=str, default="0.0.0.0", help="host address")
    parser.add_argument("--port", type=int, default=37913, help="port number")
    parser.add_argument("--chunk_size", type=int, default=-1,
                        help="chunk size for the model for prefilling")
    args = parser.parse_args()
    
    # Preload default model if desired
    if not hasattr(args, 'lazy_load') or not args.lazy_load:
        get_model(args.model_name)
    
    # Start the server
    uvicorn.run(app, host=args.host, port=args.port)