app.py

"""
TERA V2 — Hugging Face Space (Inference Only)
"""

import os
import json
import numpy as np
import gradio as gr

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
import tensorflow as tf

from tokenizer import TeraTokenizer
from model import TeraLM

print("Loading TERA V2...")

with open("model_config.json", "r") as f:
    config = json.load(f)

vocab_size = config.get("vocab_size", config.get("vocab", 1504))
d_model    = config.get("d_model", config.get("dim", 128))
n_heads    = config.get("n_heads", config.get("heads", 4))
n_layers   = config.get("n_layers", config.get("layers", 3))
max_seq    = config.get("max_seq", config.get("seq_len", 32))

print(f"  Config: vocab={vocab_size}, d={d_model}, heads={n_heads}, layers={n_layers}, seq={max_seq}")

tokenizer = TeraTokenizer()
tokenizer.load("tokenizer.json")
print(f"  Tokenizer vocab: {tokenizer.size()}")

model = TeraLM(
    vocab_size=vocab_size,
    d_model=d_model,
    n_heads=n_heads,
    n_layers=n_layers,
    max_seq=max_seq,
    drop_rate=0.0,
)

dummy = tf.zeros((1, max_seq), dtype=tf.int32)
_ = model(dummy)

# ---- LOAD H5 WEIGHTS ----
model.load_weights("model.weights.h5")
print("  Checkpoint restored ✅")

MAX_SEQ = max_seq

def generate(prompt, max_tokens=40, temperature=0.3, top_k=8, top_p=0.85, rep_penalty=1.4):
    formatted = f"Q: {prompt}\nA:"
    ids = tokenizer.encode(formatted, add_special=False)
    input_ids = ids[-MAX_SEQ:]
    generated = list(input_ids)
    seen = set(input_ids)

    for _ in range(int(max_tokens)):
        inp = tf.constant([generated[-MAX_SEQ:]], dtype=tf.int32)
        logits = model(inp, training=False)
        next_logits = logits[0, -1, :].numpy().astype(np.float64)

        for tid in seen:
            if next_logits[tid] > 0:
                next_logits[tid] /= rep_penalty
            else:
                next_logits[tid] *= rep_penalty

        next_logits = next_logits / max(temperature, 0.01)

        top_k_int = int(top_k)
        if top_k_int > 0:
            top_k_idx = np.argsort(next_logits)[-top_k_int:]
            mask = np.full_like(next_logits, -1e9)
            mask[top_k_idx] = next_logits[top_k_idx]
            next_logits = mask

        probs = np.exp(next_logits - np.max(next_logits))
        probs = probs / probs.sum()

        sorted_idx = np.argsort(probs)[::-1]
        cum = np.cumsum(probs[sorted_idx])
        cutoff = np.searchsorted(cum, top_p) + 1
        allowed = sorted_idx[:cutoff]
        filtered = np.zeros_like(probs)
        filtered[allowed] = probs[allowed]
        filtered = filtered / filtered.sum()

        next_id = np.random.choice(len(filtered), p=filtered)

        if next_id == tokenizer.eos_id:
            break
        generated.append(int(next_id))
        seen.add(int(next_id))

    response = tokenizer.decode(generated[len(input_ids):])
    return response.strip()

with gr.Blocks(title="TERA V2", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # ⭐ TERA V2
    **A language model built from scratch — no pretrained weights, no transformers.**

    *Architecture: Time Mix + Token Shift + GroupNorm + Channel Mix + Squared ReLU*
    """)

    with gr.Row():
        with gr.Column(scale=3):
            user_input = gr.Textbox(label="Your message", placeholder="Ask me anything...", lines=2)
            output_box = gr.Textbox(label="TERA V2 says", lines=4, interactive=False)
            send_btn = gr.Button("Send ⭐", variant="primary")

        with gr.Column(scale=1):
            max_tokens = gr.Slider(10, 60, value=40, step=5, label="Max tokens")
            temperature = gr.Slider(0.1, 1.0, value=0.3, step=0.05, label="Temperature")
            top_k = gr.Slider(3, 20, value=8, step=1, label="Top-K")
            top_p = gr.Slider(0.3, 1.0, value=0.85, step=0.05, label="Top-P")
            rep_penalty = gr.Slider(1.0, 2.0, value=1.4, step=0.05, label="Rep penalty")

    send_btn.click(fn=generate, inputs=[user_input, max_tokens, temperature, top_k, top_p, rep_penalty], outputs=output_box)
    user_input.submit(fn=generate, inputs=[user_input, max_tokens, temperature, top_k, top_p, rep_penalty], outputs=output_box)

    gr.Markdown("---\n*TERA V2 by Vedaco • ~929K parameters • Trained from scratch*")

demo.launch()