File size: 2,300 Bytes
b66f552 | 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 61 62 63 | # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import warnings
import torch
from fla.ops.attn.parallel import parallel_attn
def parallel_forgetting_attn(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
scale: float | None = None,
cu_seqlens: torch.LongTensor | None = None,
head_first: bool = False,
) -> torch.Tensor:
r"""
Args:
q (torch.Tensor):
queries of shape `[B, T, HQ, K]`.
k (torch.Tensor):
keys of shape `[B, T, H, K]`.
GQA will be applied if HQ is divisible by H.
v (torch.Tensor):
values of shape `[B, T, H, V]`.
g (torch.Tensor):
Log decay at rach time step (in **log space**) of shape `[B, T, HQ]` if `head_first=False` else `[B, HQ, T]`.
scale (Optional[float]):
Scale factor for attention scores.
If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
cu_seqlens (torch.LongTensor):
Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
consistent with the FlashAttention API.
head_first (Optional[bool]):
Whether the inputs are in the head-first format. Default: `False`.
This argument has been deprecated.
Returns:
o (torch.Tensor):
Outputs of shape `[B, T, HQ, V]`.
"""
if scale is None:
scale = k.shape[-1] ** -0.5
if cu_seqlens is not None:
assert q.shape[0] == 1, "batch size must be 1 when cu_seqlens are provided"
if head_first:
raise DeprecationWarning(
"head_first is deprecated and will be removed in a future version. "
"Please use head_first=False for now instead.",
)
if not head_first and q.shape[1] < q.shape[2]:
warnings.warn(
f"Input tensor shape suggests potential format mismatch: seq_len ({q.shape[1]}) < num_heads ({q.shape[2]}). "
"This may indicate the inputs were passed in head-first format [B, H, T, ...] "
"when head_first=False was specified. "
"Please verify your input tensor format matches the expected shape [B, T, H, ...].",
)
o = parallel_attn(q, k, v, g, scale, cu_seqlens)
return o
|