Spaces:
Running
Running
| from fastapi import FastAPI, Request | |
| from pydantic import BaseModel | |
| from transformers import AutoModelForCausalLM, AutoTokenizer | |
| import torch | |
| # ✅ Initialize FastAPI app | |
| app = FastAPI(title="Milo AI Backend", version="1.0") | |
| # ✅ Load model and tokenizer once at startup | |
| model_name = "milo-phi3" | |
| device = "cuda" if torch.cuda.is_available() else "cpu" | |
| model = AutoModelForCausalLM.from_pretrained(model_name).to(device) | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| # ✅ Define request schema for clarity | |
| class MiloRequest(BaseModel): | |
| message: str | |
| async def milo_generate(request: MiloRequest): | |
| """ | |
| Generate a WhatsApp-style reply using Milo AI. | |
| """ | |
| # Tokenize input | |
| inputs = tokenizer(request.message, return_tensors="pt").to(device) | |
| # Generate response | |
| output = model.generate( | |
| **inputs, | |
| max_new_tokens=200, | |
| do_sample=True, # Enable sampling for natural replies | |
| temperature=0.7, # Controls creativity | |
| top_p=0.9, # Nucleus sampling for variety | |
| pad_token_id=tokenizer.eos_token_id | |
| ) | |
| # Decode and clean response | |
| reply = tokenizer.decode(output[0], skip_special_tokens=True) | |
| return {"reply": reply} | |