-
Notifications
You must be signed in to change notification settings - Fork 19
/
sliding_window.py
52 lines (37 loc) · 1.54 KB
/
sliding_window.py
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
"""Generates a sliding window attention mask"""
import torch
from torch.nn.attention.flex_attention import _mask_mod_signature, and_masks
from attn_gym.masks import causal_mask
def generate_sliding_window(window_size: int) -> _mask_mod_signature:
"""Generates a sliding window attention mask with a given window size.
Args:
window_size: The size of the sliding window.
Note:
We assume that the window size represents the lookback size and we mask out all future tokens
similar to causal masking.
"""
def sliding_window(b, h, q_idx, kv_idx):
return q_idx - kv_idx <= window_size
sliding_window_mask = and_masks(sliding_window, causal_mask)
sliding_window_mask.__name__ = f"sliding_window_{window_size}"
return sliding_window_mask
def main(device: str = "cpu"):
"""Visualize the attention scores of sliding window mask mod.
Args:
device (str): Device to use for computation. Defaults
"""
from attn_gym import visualize_attention_scores
B, H, SEQ_LEN, HEAD_DIM = 1, 1, 12, 8
def make_tensor():
return torch.ones(B, H, SEQ_LEN, HEAD_DIM, device=device)
query, key = make_tensor(), make_tensor()
sliding_window_mask = generate_sliding_window(3)
visualize_attention_scores(
query, key, mask_mod=sliding_window_mask, device=device, name="sliding_window_mask"
)
if __name__ == "__main__":
try:
from jsonargparse import CLI
except ImportError:
raise ImportError("Be sure to run: pip install -e .'[viz]'")
CLI(main)