| |
|
|
| import warnings |
|
|
| import torch |
|
|
| from fla.ops.simple_gla.chunk import chunk_simple_gla |
|
|
|
|
| @torch.compiler.disable |
| def chunk_lightning_attn( |
| q: torch.Tensor, |
| k: torch.Tensor, |
| v: torch.Tensor, |
| layer_idx: int, |
| num_layers: int, |
| scale: float | None = None, |
| initial_state: torch.Tensor | None = None, |
| output_final_state: bool = False, |
| cu_seqlens: torch.LongTensor | None = None, |
| head_first: bool = False, |
| ) -> tuple[torch.Tensor, torch.Tensor]: |
| r""" |
| Args: |
| q (torch.Tensor): |
| queries of shape `[B, T, H, K]`. |
| k (torch.Tensor): |
| keys of shape `[B, T, H, K]`. |
| v (torch.Tensor): |
| values of shape `[B, T, H, V]`. |
| layer_idx (int): |
| The index of the current layer. |
| num_layers (int): |
| The total number of layers. Both `layer_idx` and `num_layers` are used to compute the decay factor. |
| scale (Optional[float]): |
| Scale factor for the attention scores. |
| If not provided, it will default to `1 / sqrt(K)`. Default: `None`. |
| initial_state (Optional[torch.Tensor]): |
| Initial state of shape `[N, H, K, V]` for `N` input sequences. |
| For equal-length input sequences, `N` equals the batch size `B`. |
| Default: `None`. |
| output_final_state (Optional[bool]): |
| Whether to output the final state of shape `[N, H, K, V]`. Default: `False`. |
| 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, H, V]`. |
| final_state (torch.Tensor): |
| Final state of shape `[N, H, K, V]` if `output_final_state=True` else `None`. |
| """ |
| 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, ...].", |
| ) |
|
|
| H = q.shape[2] |
| g_gamma = -(8 / H * (1 - layer_idx / num_layers)) * q.new_tensor(range(H), dtype=torch.float) |
| return chunk_simple_gla( |
| q=q, |
| k=k, |
| v=v, |
| scale=scale, |
| g_gamma=g_gamma, |
| initial_state=initial_state, |
| output_final_state=output_final_state, |
| head_first=head_first, |
| cu_seqlens=cu_seqlens, |
| ) |
|
|
