app.py

import gradio as gr
import onnxruntime as ort
import json
import torch

# Load vocab
with open("vocab.json", "r", encoding="utf-8") as f:
    vocab = json.load(f)
inv_vocab = {i:tok for tok,i in vocab.items()}

pad_idx = vocab.get("<pad>", 0)
unk_idx = vocab.get("<unk>", 1)

# Load ONNX model
session = ort.InferenceSession("chat_model.onnx")

# Tokenizer helper
def tokenize(text):
    return [vocab.get(tok, unk_idx) for tok in text.split(" ")]

def pad_sequence(seq, max_len=20):
    seq = seq + [pad_idx]*(max_len - len(seq))
    return seq[:max_len]

# Fungsi chat mini
def chat_onnx(input_text, max_len=20):
    input_ids = pad_sequence(tokenize(input_text), max_len)
    input_tensor = np.array([input_ids], dtype=np.int64)
    
    output_ids = []
    h = None  # ONNX simple RNN ini biasanya stateless
    for _ in range(max_len):
        ort_inputs = {"input": input_tensor}
        ort_outs = session.run(None, ort_inputs)
        next_token = int(ort_outs[0][0, -1].argmax())
        output_ids.append(next_token)
        input_tensor = np.array([[next_token]], dtype=np.int64)
    
    return " ".join([inv_vocab.get(i, "<unk>") for i in output_ids])

# Gradio interface
iface = gr.Interface(fn=chat_onnx, inputs="text", outputs="text")
iface.launch()