import base64 from io import BytesIO from fastapi import FastAPI, HTTPException from pydantic import BaseModel from transformers import AutoModelForCausalLM, AutoProcessor import torch from PIL import Image import subprocess # Install flash-attn subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True) app = FastAPI() models = { "microsoft/Phi-3.5-vision-instruct": AutoModelForCausalLM.from_pretrained( "microsoft/Phi-3.5-vision-instruct", trust_remote_code=True, torch_dtype="auto", attn_implementation="flash_attention_2" ).cuda().eval() } processors = { "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained( "microsoft/Phi-3.5-vision-instruct", trust_remote_code=True ) } class InputData(BaseModel): image: str text_input: str model_id: str = "microsoft/Phi-3.5-vision-instruct" @app.post("/run_example") async def run_example(input_data: InputData): try: model = models[input_data.model_id] processor = processors[input_data.model_id] # Decode base64 image image_data = base64.b64decode(input_data.image) image = Image.open(BytesIO(image_data)).convert("RGB") user_prompt = '<|user|>\n' assistant_prompt = '<|assistant|>\n' prompt_suffix = "<|end|>\n" prompt = f"{user_prompt}<|image_1|>\n{input_data.text_input}{prompt_suffix}{assistant_prompt}" inputs = processor(prompt, image, return_tensors="pt").to("cuda:0") generate_ids = model.generate( **inputs, max_new_tokens=1000, eos_token_id=processor.tokenizer.eos_token_id, ) generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:] response = processor.batch_decode( generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0] return {"response": response} except Exception as e: raise HTTPException(status_code=500, detail=str(e))