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_update_cpu_no_bias

def causal_conv1d_update_cpu_no_bias[x_dtype: DType, conv_state_dtype: DType, weight_dtype: DType, output_dtype: DType](batch: Int, dim: Int, seqlen: Int, width: Int, state_len: Int, x: TileTensor[x_dtype, Storage=x.Storage, address_space=x.address_space, linear_idx_type=x.linear_idx_type], conv_state: TileTensor[conv_state_dtype, Storage=conv_state.Storage, address_space=conv_state.address_space, linear_idx_type=conv_state.linear_idx_type], weight: TileTensor[weight_dtype, Storage=weight.Storage, address_space=weight.address_space, linear_idx_type=weight.linear_idx_type], output: TileTensor[output_dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], x_batch_stride: UInt32, x_c_stride: UInt32, x_l_stride: UInt32, conv_state_batch_stride: UInt32, conv_state_c_stride: UInt32, conv_state_l_stride: UInt32, weight_c_stride: UInt32, weight_width_stride: UInt32, out_batch_stride: UInt32, out_c_stride: UInt32, out_l_stride: UInt32, silu_activation: Bool)

CPU implementation of causal conv1d update without bias.

Performs incremental convolution for autoregressive decode by treating conv_state followed by x as a virtual sliding window, computing the output for the new positions, then shifting the newest state_len values back into conv_state in place.

Parameters:

x_dtype (DType): Element type of the input tensor x.
conv_state_dtype (DType): Element type of the convolution state tensor conv_state.
weight_dtype (DType): Element type of the weight tensor weight.
output_dtype (DType): Element type of the output tensor output.

Args:

batch (Int): Number of sequences processed in parallel.
dim (Int): Number of channels per sequence position.
seqlen (Int): Number of new input positions in x (1 for autoregressive decode).
width (Int): Convolution kernel width in positions.
state_len (Int): Length of the rolling buffer stored in conv_state; must be at least width - 1.
x (TileTensor[x_dtype, Storage=x.Storage, address_space=x.address_space, linear_idx_type=x.linear_idx_type]): Input tensor of shape (B, C, L) holding the new positions to convolve.
conv_state (TileTensor[conv_state_dtype, Storage=conv_state.Storage, address_space=conv_state.address_space, linear_idx_type=conv_state.linear_idx_type]): Rolling convolution state of shape (B, C, S) that holds the last state_len values; updated in place.
weight (TileTensor[weight_dtype, Storage=weight.Storage, address_space=weight.address_space, linear_idx_type=weight.linear_idx_type]): Convolution weights of shape (C, W).
output (TileTensor[output_dtype, Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output tensor of shape (B, C, L) receiving the convolved values for the new positions.
x_batch_stride (UInt32): Stride in elements between batches of x.
x_c_stride (UInt32): Stride in elements between channels of x.
x_l_stride (UInt32): Stride in elements between sequence positions of x.
conv_state_batch_stride (UInt32): Stride in elements between batches of conv_state.
conv_state_c_stride (UInt32): Stride in elements between channels of conv_state.
conv_state_l_stride (UInt32): Stride in elements between state positions of conv_state.
weight_c_stride (UInt32): Stride in elements between channels of weight.
weight_width_stride (UInt32): Stride in elements between taps of weight along the kernel-width dimension.
out_batch_stride (UInt32): Stride in elements between batches of output.
out_c_stride (UInt32): Stride in elements between channels of output.
out_l_stride (UInt32): Stride in elements between sequence positions of output.
silu_activation (Bool): Whether to apply the SiLU activation to the output values before storing.