For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

causal_conv1d_varlen_states_shape

def causal_conv1d_varlen_states_shape[dtype: DType](x: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=x.static_spec], cu_seqlens: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=cu_seqlens.static_spec]) -> IndexList[Int(3)]

Returns the output shape for the causal_conv1d_varlen_states op.

The output is a state buffer with one entry per sequence: shape (batch, dim, state_len). The state_len dimension is determined by the output allocation at runtime; this function returns 0 for that dimension as a placeholder.

Parameters:

dtype (DType): Element type of the packed input tensor x.

Args:

x (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=x.static_spec]): Packed input tensor with shape (total_tokens, dim).
cu_seqlens (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=cu_seqlens.static_spec]): Cumulative sequence lengths with shape (batch + 1,).

Returns:

IndexList[Int(3)]: The output state shape (batch, dim, 0) where batch is inferred from cu_seqlens and state_len is filled at runtime.