app.py

from fastapi import FastAPI
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

MODEL = "google/gemma-4-E4B"

# Carregamento otimista para CPU
tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    device_map={"": "cpu"},
    torch_dtype=torch.float32,
    low_cpu_mem_usage=True,
    trust_remote_code=True
)

generator = pipeline("text-generation", model=model, tokenizer=tokenizer, device=-1)

def generate(prompt):
    out = generator(prompt, max_new_tokens=128, do_sample=False)
    return out[0]["generated_text"]

demo = gr.Interface(fn=generate, inputs=gr.Textbox(lines=4, label="Prompt"), outputs="text", title="Gemma-4-E4B (CPU)")

app = FastAPI()
app = gr.mount_gradio_app(app, demo, path="/")