| |
| |
| |
| |
| |
| |
|
|
| import torch |
|
|
| from ...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, |
| window_size: int | None = None, |
| cu_seqlens: torch.LongTensor | None = None, |
| **kwargs |
| ) -> 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 factors of shape `[B, T, HQ]`. |
| scale (Optional[float]): |
| Scale factor for attention scores. |
| If not provided, it will default to `1 / sqrt(K)`. Default: `None`. |
| window_size (Optional[int]): |
| Sliding window size. If provided, each query at position i only attends to |
| keys in `[i - window_size + 1, i]`. If `None`, full causal attention is used. |
| Default: `None`. |
| cu_seqlens (torch.LongTensor): |
| Cumulative sequence lengths of shape `[N+1]` used for variable-length training, |
| consistent with the FlashAttention API. |
| |
| Returns: |
| o (torch.Tensor): |
| Outputs of shape `[B, T, HQ, V]`. |
| """ |
| if 'head_first' in kwargs: |
| raise DeprecationWarning( |
| "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.", |
| ) |
| if scale is None: |
| scale = k.shape[-1] ** -0.5 |
| if cu_seqlens is not None and q.shape[0] != 1: |
| raise ValueError( |
| f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`. " |
| f"Please flatten variable-length inputs before processing.", |
| ) |
|
|
| o = parallel_attn(q, k, v, g, scale, window_size=window_size, cu_seqlens=cu_seqlens) |
| return o |
|
|
