tiny_english/generate.py

import argparse
import json
import math
import time
from pathlib import Path

import torch
from tokenizers import Tokenizer

from .model import LanguageModel, ModelConfig
from .artifacts import digest_file


def load_model(checkpoint_path, tokenizer_path, device="cpu"):
    checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
    if digest_file(tokenizer_path) != checkpoint["tokenizer_sha256"]:
        raise ValueError("Tokenizer does not match checkpoint")
    tokenizer = Tokenizer.from_file(str(tokenizer_path))
    model = LanguageModel(ModelConfig(**checkpoint["config"]))
    if tokenizer.get_vocab_size() != model.config.vocab_size:
        raise ValueError("Tokenizer vocabulary does not match model")
    model.load_state_dict(checkpoint["model"])
    step = checkpoint["step"]
    del checkpoint
    return model.to(device).eval(), tokenizer, step


def load_export(directory, device="cpu"):
    from safetensors.torch import load_file
    directory = Path(directory)
    metadata = json.loads((directory / "config.json").read_text(encoding="utf-8"))
    if metadata.get("format") != "tiny-english-safetensors-v1":
        raise ValueError("Unsupported exported model format")
    if metadata.get("dtype") != "bfloat16":
        raise ValueError("This export requires BF16 weights")
    if device != "cuda":
        raise ValueError("This BF16 export requires a CUDA GPU; use --device cuda")
    tokenizer_path = directory / "tokenizer.json"
    weights_path = directory / "model.safetensors"
    if digest_file(tokenizer_path) != metadata["tokenizer_sha256"]:
        raise ValueError("Export tokenizer checksum does not match")
    if digest_file(weights_path) != metadata["weights_sha256"]:
        raise ValueError("Export weights checksum does not match")
    tokenizer = Tokenizer.from_file(str(tokenizer_path))
    model = LanguageModel(ModelConfig(**metadata["model_config"]))
    if tokenizer.get_vocab_size() != model.config.vocab_size:
        raise ValueError("Export tokenizer vocabulary does not match model")
    model.load_state_dict(load_file(str(weights_path), device="cpu"), strict=True)
    return model.to(device=device, dtype=torch.bfloat16).eval(), tokenizer, metadata["checkpoint_step"]


def complete(model, tokenizer, prompt, max_new_tokens=150, temperature=0.8, top_k=40, seed=42, deadline=None, strict_context=False, stop_at_sentence_end=False, min_new_tokens=0):
    if not isinstance(prompt, str) or not prompt.strip():
        raise ValueError("Prompt must not be empty")
    if not math.isfinite(temperature) or temperature <= 0 or top_k < 1 or max_new_tokens < 1 or min_new_tokens < 0:
        raise ValueError("Generation numeric options must be positive and finite")
    tokens = tokenizer.encode(prompt).ids
    prompt_tokens = len(tokens)
    if not tokens:
        raise ValueError("Prompt must contain tokens")
    if strict_context and prompt_tokens > model.config.context_length:
        raise ValueError(f"Prompt has {prompt_tokens} tokens; the maximum is {model.config.context_length}. Shorten the prompt.")
    device = model.embedding.weight.device
    generator = torch.Generator(device=device).manual_seed(seed)
    finish_reason = "length"
    with torch.inference_mode():
        for _ in range(max_new_tokens):
            if deadline is not None and time.monotonic() >= deadline:
                raise TimeoutError("Generation exceeded its time limit. Try fewer tokens or a shorter prompt.")
            inputs = torch.tensor([tokens[-model.config.context_length:]], device=device)
            logits = model(inputs)[0, -1].float() / temperature
            values, indices = logits.topk(min(top_k, logits.numel()))
            next_token = indices[torch.multinomial(values.softmax(-1), 1, generator=generator)].item()
            if next_token == tokenizer.token_to_id("<eos>"):
                finish_reason = "eos"
                break
            tokens.append(next_token)
            if stop_at_sentence_end and len(tokens) - prompt_tokens >= min_new_tokens:
                ending = tokenizer.decode([next_token]).rstrip()
                completion = tokenizer.decode(tokens[prompt_tokens:])
                if ending.endswith((".", "!", "?")) and completion.count('"') % 2 == 0:
                    finish_reason = "sentence"
                    break
    return {
        "text": tokenizer.decode(tokens),
        "completion": tokenizer.decode(tokens[prompt_tokens:]),
        "prompt_tokens": prompt_tokens, "generated_tokens": len(tokens) - prompt_tokens,
        "context_length": model.config.context_length, "finish_reason": finish_reason,
        "sliding_context": len(tokens) > model.config.context_length,
    }


def main():
    parser = argparse.ArgumentParser(description="Continue a simple English prompt")
    source = parser.add_mutually_exclusive_group(required=True)
    source.add_argument("--checkpoint", help="Local training checkpoint")
    source.add_argument("--model-dir", type=Path, help="Downloaded/exported safetensors model directory")
    parser.add_argument("--tokenizer", help="Required with --checkpoint; inferred for --model-dir")
    parser.add_argument("--prompt", default="Once upon a time, a little rabbit found a blue hat.")
    parser.add_argument("--max-new-tokens", type=int, default=150)
    parser.add_argument("--temperature", type=float, default=0.8)
    parser.add_argument("--top-k", type=int, default=40)
    parser.add_argument("--device", choices=["cpu", "cuda"], default="cuda")
    parser.add_argument("--threads", type=int, default=8)
    parser.add_argument("--seed", type=int, default=42)
    args = parser.parse_args()
    if args.checkpoint and not args.tokenizer:
        parser.error("--tokenizer is required with --checkpoint")
    if args.model_dir and args.tokenizer:
        parser.error("--model-dir uses its own tokenizer.json; omit --tokenizer")
    if not math.isfinite(args.temperature) or args.temperature <= 0 or args.top_k < 1 or args.max_new_tokens < 1 or args.threads < 1:
        parser.error("Generation numeric options must be positive")
    torch.set_num_threads(args.threads)
    model, tokenizer, _ = load_export(args.model_dir, args.device) if args.model_dir else load_model(args.checkpoint, args.tokenizer, args.device)
    try:
        result = complete(model, tokenizer, args.prompt, args.max_new_tokens, args.temperature, args.top_k, args.seed)
    except ValueError as error:
        parser.error(str(error))
    print(result["text"])


if __name__ == "__main__":
    main()