File size: 3,030 Bytes
ea3734f |
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 59 60 |
import torch
# https://github.com/NVIDIA/Megatron-LM/blob/0bb597b42c53355a567aba2a1357cc34b9d99ddd/megatron/text_generation/sampling.py
# https://github.com/huggingface/transformers/blob/a44985b41cfa2de48a5e1de7f1f93b7483da25d1/src/transformers/generation/logits_process.py#L231
def modify_logits_for_top_k_filtering(logits, top_k):
"""Set the logits for none top-k values to -inf. Done in-place."""
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits.masked_fill_(indices_to_remove, float("-Inf"))
# https://github.com/NVIDIA/Megatron-LM/blob/0bb597b42c53355a567aba2a1357cc34b9d99ddd/megatron/text_generation/sampling.py
# https://github.com/huggingface/transformers/blob/a44985b41cfa2de48a5e1de7f1f93b7483da25d1/src/transformers/generation/logits_process.py#L170
def modify_logits_for_top_p_filtering(logits, top_p):
"""Set the logits for none top-p values to -inf. Done in-place."""
if top_p <= 0.0 or top_p >= 1.0:
return
# First sort and calculate cumulative sum of probabilities.
sorted_logits, sorted_indices = torch.sort(logits, descending=False)
cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
# Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
sorted_indices_to_remove = cumulative_probs <= (1 - top_p)
# scatter sorted tensors to original indexing
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
logits.masked_fill_(indices_to_remove, float("-inf"))
# https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py
def sample(logits, top_k=1, top_p=0.0, temperature=1.0):
"""Sample from top-k logits.
Arguments:
logits: Tensor of shape (batch_size, vocab_size)
"""
logits = torch.nan_to_num(logits)
logits = torch.where(logits == float("-inf"), 0, logits)
logits = torch.where(logits == float("inf"), 0, logits)
if top_k == 1: # Short-circuit for greedy decoding
return logits.argmax(dim=-1)
else:
if top_p > 0.0:
assert top_p <= 1.0, "top-p should be in (0, 1]."
if top_k > 0:
top_k = min(top_k, logits.size(-1)) # Safety check
logits_top, indices = torch.topk(logits, top_k, dim=-1)
if temperature != 1.0:
logits_top /= temperature
modify_logits_for_top_p_filtering(logits_top, top_p)
return indices[
torch.arange(indices.shape[0], device=indices.device),
torch.multinomial(torch.softmax(logits_top, dim=-1), num_samples=1).squeeze(dim=-1),
]
else:
# Clone so that when we modify for top_p we don't change the original logits
logits_top = logits / temperature if temperature != 1.0 else logits.clone()
modify_logits_for_top_p_filtering(logits_top, top_p)
return torch.multinomial(torch.softmax(logits_top, dim=-1), num_samples=1).squeeze(dim=-1)
|