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

Python function

short_conv_ring

short_conv_ring()

max.nn.state_space.short_conv_ring(x, weight, ring, input_row_offsets, positions, conv_rows, layer_row, *, commit)

source

Returns x + conv(x) over a ragged batch, optionally committing the chunk to ring.

Parameters:

• x (TensorValue) – [total_seq_len, channels] input.
• weight (TensorValue) – [channels, kernel_size] taps; the last multiplies the current token.
• ring (BufferValue) – [slots, ring_len, channels] conv state; written in place when commit is set.
• input_row_offsets (TensorValue) – [batch + 1] uint32.
• positions (TensorValue) – [total_seq_len] uint32 position per token.
• conv_rows (TensorValue) – [num_layers, batch] uint32 ring slot per sequence.
• layer_row (TensorValue) – Scalar uint32 CPU row of conv_rows this layer reads.
• commit (bool) – Whether to write each sequence’s last ring_len rows of x into its slot of ring.

Returns:

Same shape and dtype as x.

Return type:

TensorValue