vllm.models.glm5next.nvidia.ops.third_party.kda.fused_recurrent ¶

def fused_recurrent_gated_delta_rule(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor = None,
    scale: float = None,
    initial_state: torch.Tensor = None,
    inplace_final_state: bool = True,
    cu_seqlens: torch.Tensor | None = None,
    ssm_state_indices: torch.Tensor | None = None,
    num_accepted_tokens: torch.Tensor | None = None,
    use_qk_l2norm_in_kernel: 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, HV, V]`.
            GVA is applied if `HV > H`.
        g (torch.Tensor):
            g (decays) of shape `[B, T, HV]`.
        beta (torch.Tensor):
            betas of shape `[B, T, HV]`.
        scale (Optional[int]):
            Scale factor for the RetNet attention scores.
            If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
        initial_state (Optional[torch.Tensor]):
            Initial state of shape `[N, HV, V, K]` for `N` input sequences.
            For equal-length input sequences, `N` equals the batch size `B`.
            Default: `None`.
        inplace_final_state: bool:
            Whether to store the final state in-place to save memory.
            Default: `True`.
        cu_seqlens (torch.Tensor):
            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
            consistent with the FlashAttention API.
        ssm_state_indices (Optional[torch.Tensor]):
            Indices to map the input sequences to the initial/final states.
        num_accepted_tokens (Optional[torch.Tensor]):
            Number of accepted tokens for each sequence during decoding.

    Returns:
        o (torch.Tensor):
            Outputs of shape `[B, T, HV, V]`.
        final_state (torch.Tensor):
            Final state of shape `[N, HV, V, K]`.

    Examples::
        >>> import torch
        >>> import torch.nn.functional as F
        >>> from einops import rearrange
        >>> from fla.ops.gated_delta_rule import fused_recurrent_gated_delta_rule
        # inputs with equal lengths
        >>> B, T, H, HV, K, V = 4, 2048, 4, 8, 512, 512
        >>> q = torch.randn(B, T, H, K, device='cuda')
        >>> k = F.normalize(torch.randn(B, T, H, K, device='cuda'), p=2, dim=-1)
        >>> v = torch.randn(B, T, HV, V, device='cuda')
        >>> g = F.logsigmoid(torch.rand(B, T, HV, device='cuda'))
        >>> beta = torch.rand(B, T, HV, device='cuda').sigmoid()
        >>> h0 = torch.randn(B, HV, V, K, device='cuda')
        >>> o, ht = fused_gated_recurrent_delta_rule(
            q, k, v, g, beta,
            initial_state=h0,
        )
        # for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
        >>> q, k, v, g, beta = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, g, beta))
        # for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
        >>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.int32)
        >>> o_var, ht_var = fused_gated_recurrent_delta_rule(
            q, k, v, g, beta,
            initial_state=h0,
            cu_seqlens=cu_seqlens
        )
    """
    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."
        )
    if scale is None:
        scale = k.shape[-1] ** -0.5
    else:
        assert scale > 0, "scale must be positive"
    if beta is None:
        beta = torch.ones_like(q[..., 0])
    o, final_state = FusedRecurrentFunction.apply(
        q,
        k,
        v,
        g,
        beta,
        scale,
        initial_state,
        inplace_final_state,
        cu_seqlens,
        ssm_state_indices,
        num_accepted_tokens,
        use_qk_l2norm_in_kernel,
    )
    return o, final_state

def token_stride(x: torch.Tensor) -> int:
    """Token stride (elements) of a ``[B, T, H, D]`` or ``[B, T, H]`` tensor.

    The recurrent kernel walks tokens with this stride and addresses heads
    densely inside a token, so each token's ``[H, D]`` (or ``[H]``) block must
    be contiguous, tokens must not overlap, and with ``B > 1`` sequence ``n``
    must start at token ``n * T`` (dense batch). Column slices of a wider
    per-token projection buffer satisfy this and are consumed in place.
    """
    st = x.stride()
    assert x.dim() in (3, 4) and st[-1] == 1, (x.shape, st)
    assert x.dim() == 3 or st[2] == x.shape[3], (x.shape, st)
    assert st[1] >= x.shape[2] * (x.shape[3] if x.dim() == 4 else 1), (x.shape, st)
    assert x.shape[0] == 1 or st[0] == x.shape[1] * st[1], (x.shape, st)
    return st[1]