app.py

import gradio as gr
import torch

import bdh
from bdh import BDHConfig, BDH

# Device (Spaces usually run on CPU)
device = torch.device("cpu")

# Small config for live demo (important!)
config = BDHConfig(
    n_layer=2,
    n_embd=128,
    n_head=4,
    mlp_internal_dim_multiplier=32,
    vocab_size=256,
)

model = BDH(config).to(device)
model.eval()

@torch.no_grad()
def generate_text(prompt: str, max_tokens: int = 50):
    if len(prompt) == 0:
        return "Please enter a prompt."

    idx = torch.tensor(
        bytearray(prompt, "utf-8"),
        dtype=torch.long,
        device=device
    ).unsqueeze(0)

    out = model.generate(
        idx,
        max_new_tokens=max_tokens,
        top_k=5
    )

    return bytes(
        out.squeeze(0).to(torch.uint8).cpu()
    ).decode(errors="replace")


gr.Interface(
    fn=generate_text,
    inputs=[
        gr.Textbox(label="Prompt"),
        gr.Slider(10, 100, value=50, step=10, label="Max tokens")
    ],
    outputs=gr.Textbox(label="BDH Output"),
    title="BDH – Beyond Transformer (Live Demo)",
    description="Inference-only demo of the BDH architecture"
).launch()