File size: 2,002 Bytes
2e1dc7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
import torch


def multinomial(input: torch.Tensor,
                num_samples: int,
                replacement=False,
                *,
                generator=None):
    """torch.multinomial with arbitrary number of dimensions, and number of candidates on the last dimension.

    Args:
        input (torch.Tensor): The input tensor containing probabilities.
        num_samples (int): Number of samples to draw.
        replacement (bool): Whether to draw with replacement or not.
    Keywords args:
        generator (torch.Generator): A pseudorandom number generator for sampling.
    Returns:
        torch.Tensor: Last dimension contains num_samples indices
            sampled from the multinomial probability distribution
            located in the last dimension of tensor input.
    """
    input_ = input.reshape(-1, input.shape[-1])
    output_ = torch.multinomial(input_,
                                num_samples=num_samples,
                                replacement=replacement,
                                generator=generator)
    output = output_.reshape(*list(input.shape[:-1]), -1)
    return output


def sample_top_k(probs: torch.Tensor, k: int) -> torch.Tensor:
    """Sample next token from top K values along the last dimension of the input probs tensor.

    Args:
        probs (torch.Tensor): Input probabilities with token candidates on the last dimension.
        k (int): The k in “top-k”.
    Returns:
        torch.Tensor: Sampled tokens.
    """
    top_k_value, _ = torch.topk(probs, k, dim=-1)
    min_value_top_k = top_k_value[..., [-1]]
    probs = probs * (probs >= min_value_top_k).float()
    probs.div_(probs.sum(dim=-1, keepdim=True))
    next_token = multinomial(probs, num_samples=1)
    return next_token


def eval_decorator(fn):

    def inner(self, *args, **kwargs):
        was_training = self.training
        self.eval()
        out = fn(self, *args, **kwargs)
        self.train(was_training)
        return out

    return inner