teragron/TinyStories

1"""2Sample from the trained model with PyTorch3"""4import os5import pickle6from contextlib import nullcontext7import torch8from model import ModelArgs, Transformer9from tokenizer import Tokenizer10 11from tinystories import get_tokenizer_model_path12 13# -----------------------------------------------------------------------------14checkpoint = 'out/ckpt.pt'15start = "" # or "<|endoftext|>" or etc. Can also specify a file, use as: "FILE:prompt.txt"16num_samples = 1 # number of samples to draw17max_new_tokens = 100 # number of tokens generated in each sample18temperature = 1.0 # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions19top_k = 300 # retain only the top_k most likely tokens, clamp others to have 0 probability20tokenizer = "" # override the tokenizer model path21seed = 133722device = 'cuda' if torch.cuda.is_available() else 'cpu' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.23#dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32' or 'bfloat16' or 'float16'24dtype = "float32"25compile = False # use PyTorch 2.0 to compile the model to be faster26exec(open('configurator.py').read()) # overrides from command line or config file27# -----------------------------------------------------------------------------28 29torch.manual_seed(seed)30torch.cuda.manual_seed(seed)31torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul32torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn33device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast34ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]35ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)36 37# init from a model saved in a specific directory38checkpoint_dict = torch.load(checkpoint, map_location=device)39gptconf = ModelArgs(**checkpoint_dict['model_args'])40model = Transformer(gptconf)41state_dict = checkpoint_dict['model']42unwanted_prefix = '_orig_mod.'43for k,v in list(state_dict.items()):44    if k.startswith(unwanted_prefix):45        state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)46model.load_state_dict(state_dict, strict=False)47 48model.eval()49model.to(device)50if compile:51    print("Compiling the model...")52    model = torch.compile(model) # requires PyTorch 2.0 (optional)53 54# load the tokenizer55vocab_source = checkpoint_dict["config"].get("vocab_source", "llama2")56vocab_size = gptconf.vocab_size57if tokenizer:58    # a specific tokenizer is provided, use it59    tokenizer_model = tokenizer60else:61    # let's try to find the tokenizer model automatically. bit gross here...62    query_vocab_size = 0 if vocab_source == "llama2" else vocab_size63    tokenizer_model = get_tokenizer_model_path(vocab_size=query_vocab_size)64enc = Tokenizer(tokenizer_model=tokenizer_model)65 66# encode the beginning of the prompt67if start.startswith('FILE:'):68    with open(start[5:], 'r', encoding='utf-8') as f:69        start = f.read()70start_ids = enc.encode(start, bos=True, eos=False)71x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])72 73# run generation74with torch.no_grad():75    with ctx:76        for k in range(num_samples):77            y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)78            print(enc.decode(y[0].tolist()))79            print('---------------')80