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

gated_delta_conv1d_fwd_shape

def gated_delta_conv1d_fwd_shape[work_dtype: DType, state_dtype: DType](qkv_input_ragged: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=qkv_input_ragged.static_spec], conv_weight: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=conv_weight.static_spec], conv_state: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=conv_state.static_spec], slot_idx: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=slot_idx.static_spec], input_row_offsets: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input_row_offsets.static_spec]) -> IndexList[Int(2)]

Computes the output shape for the gated_delta_conv1d_fwd graph op.

Parameters:

work_dtype (DType): Element type of the per-token input and weight tensors (inferred).
state_dtype (DType): Element type of the persistent conv-state pool, independent of work_dtype (inferred).

Args:

qkv_input_ragged (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=qkv_input_ragged.static_spec]): Ragged QKV input tensor of shape [total_seq_len, conv_dim].
conv_weight (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=conv_weight.static_spec]): Convolution filter of shape [conv_dim, kernel_size].
conv_state (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=conv_state.static_spec]): Mutable conv-state pool of shape [max_slots, conv_dim, kernel_size-1].
slot_idx (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=slot_idx.static_spec]): Per-batch slot indices into the conv-state pool, shape [batch_size].
input_row_offsets (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input_row_offsets.static_spec]): Cumulative row offsets per batch, shape [batch_size + 1].

Returns:

IndexList[Int(2)]