1  """
 2  Sample from a trained model
 3  """
 4  import os
 5  import pickle
 6  from contextlib import nullcontext
 7  import torch
 8  import tiktoken
 9  from model import GPTConfig, GPT
10  
11  # -----------------------------------------------------------------------------
12  init_from = 'resume' # either 'resume' (from an out_dir) or a gpt2 variant (e.g. 'gpt2-xl')
13  out_dir = 'out' # ignored if init_from is not 'resume'
14  start = "\n" # or "<|endoftext|>" or etc. Can also specify a file, use as: "FILE:prompt.txt"
15  num_samples = 10 # number of samples to draw
16  max_new_tokens = 500 # number of tokens generated in each sample
17  temperature = 0.8 # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions
18  top_k = 200 # retain only the top_k most likely tokens, clamp others to have 0 probability
19  seed = 1337
20  device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.
21  dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16' # 'float32' or 'bfloat16' or 'float16'
22  compile = False # use PyTorch 2.0 to compile the model to be faster
23  exec(open('configurator.py').read()) # overrides from command line or config file
24  # -----------------------------------------------------------------------------
25  
26  torch.manual_seed(seed)
27  torch.cuda.manual_seed(seed)
28  torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
29  torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
30  device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
31  ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
32  ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
33  
34  # model
35  if init_from == 'resume':
36      # init from a model saved in a specific directory
37      ckpt_path = os.path.join(out_dir, 'ckpt.pt')
38      checkpoint = torch.load(ckpt_path, map_location=device)
39      gptconf = GPTConfig(**checkpoint['model_args'])
40      model = GPT(gptconf)
41      state_dict = checkpoint['model']
42      unwanted_prefix = '_orig_mod.'
43      for k,v in list(state_dict.items()):
44          if k.startswith(unwanted_prefix):
45              state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
46      model.load_state_dict(state_dict)
47  elif init_from.startswith('gpt2'):
48      # init from a given GPT-2 model
49      model = GPT.from_pretrained(init_from, dict(dropout=0.0))
50  
51  model.eval()
52  model.to(device)
53  if compile:
54      model = torch.compile(model) # requires PyTorch 2.0 (optional)
55  
56  # look for the meta pickle in case it is available in the dataset folder
57  load_meta = False
58  if init_from == 'resume' and 'config' in checkpoint and 'dataset' in checkpoint['config']: # older checkpoints might not have these...
59      meta_path = os.path.join('data', checkpoint['config']['dataset'], 'meta.pkl')
60      load_meta = os.path.exists(meta_path)
61  if load_meta:
62      print(f"Loading meta from {meta_path}...")
63      with open(meta_path, 'rb') as f:
64          meta = pickle.load(f)
65      # TODO want to make this more general to arbitrary encoder/decoder schemes
66      stoi, itos = meta['stoi'], meta['itos']
67      encode = lambda s: [stoi[c] for c in s]
68      decode = lambda l: ''.join([itos[i] for i in l])
69  else:
70      # ok let's assume gpt-2 encodings by default
71      print("No meta.pkl found, assuming GPT-2 encodings...")
72      enc = tiktoken.get_encoding("gpt2")
73      encode = lambda s: enc.encode(s, allowed_special={"<|endoftext|>"})
74      decode = lambda l: enc.decode(l)
75  
76  # encode the beginning of the prompt
77  if start.startswith('FILE:'):
78      with open(start[5:], 'r', encoding='utf-8') as f:
79          start = f.read()
80  start_ids = encode(start)
81  x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])
82  
83  # run generation
84  with torch.no_grad():
85      with ctx:
86          for k in range(num_samples):
87              y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
88              print(decode(y[0].tolist()))
89              print('---------------')