"""Chat with Têtu Code 2 in the terminal (no tools). For the agent in OpenCode: python -m tetu_code.serve python chat.py (GPU) python chat.py --cpu """ import argparse from pathlib import Path import re import torch from tetu_code import fmt from tetu_code.generate import Decoder, load ROOT = Path(__file__).resolve().parent def main(): p = argparse.ArgumentParser() p.add_argument("--cpu", action="store_true") p.add_argument("--temp", type=float, default=0.2) p.add_argument("--max_new", type=int, default=1024) a = p.parse_args() dev = "cpu" if a.cpu or not torch.cuda.is_available() else "cuda" model, tok = load(str(ROOT), device=dev, dtype=torch.float32 if dev == "cpu" else torch.bfloat16) dec = Decoder(model, tok, max_len=16384, graph=dev == "cuda") stop = {tok.token_to_id(t) for t in ("<|end|>", "<|eos|>", "<|user|>", "<|/call|>")} text = f"<|system|>{fmt.system_text()}" print("Têtu Code 2 — /reset to clear, Ctrl+C to quit") while True: q = input("\n> ").strip() if q == "/reset": text = f"<|system|>{fmt.system_text()}" continue text += f"<|user|>{q}<|assistant|>" ids = tok.encode(text, add_special_tokens=False).ids out = [] for t in dec.generate(ids, max_new=a.max_new, temp=a.temp, stop=stop): if t in stop: break out.append(t) print(tok.decode([t], skip_special_tokens=False), end="", flush=True) reply = tok.decode(out, skip_special_tokens=False) text += reply + "<|end|>" print() if __name__ == "__main__": main()