app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

app = FastAPI()

# Define paths
base_model_path = "NousResearch/Hermes-3-Llama-3.2-3B"
adapter_path = "zach9111/llama_startup_adapter"

# Check if GPU is available
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load base model with `device_map="auto"` to handle GPUs automatically
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_path, torch_dtype=torch.float16, device_map="auto"
)

# Load adapter and ensure it is on the correct device
model = PeftModel.from_pretrained(base_model, adapter_path).to(device)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_path)


class GenerateRequest(BaseModel):
    prompt: str

# **Use `model.generate()` instead of `pipeline()`**
def generate_text_from_model(prompt: str):
    try:
        input_ids = tokenizer(f"<s>[INST] {prompt} [/INST]", return_tensors="pt").input_ids.to(device)
        output_ids = model.generate(input_ids, max_length=512)
        generated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
        return generated_text
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

# Root endpoint for testing
@app.get("/")
async def root():
    return {"message": "Model is running! Use /generate/ for text generation."}

# Text generation endpoint
@app.post("/generate/")
async def generate_text(request: GenerateRequest):
    response = generate_text_from_model(request.prompt)
    return {"response": response}