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

gated_delta_recurrence_verify_ring_fwd

gated_delta_recurrence_verify_ring_fwd()

max.nn.state_space.gated_delta_recurrence_verify_ring_fwd(qkv_conv_output, decay_per_token, beta_per_token, recurrent_state, ring, slot_idx, ring_slot_idx, input_row_offsets, verify_width)

source

Runs the recurrence over a speculative verify window into a ring.

Produces the same readout as gated_delta_recurrence_fwd() without writing recurrent_state. Each token writes its decay, raw key and delta factor to its request’s ring row, which gated_delta_state_fold() applies once acceptance is known.

At a verify width of zero this is gated_delta_recurrence_fwd() on recurrent_state, and no ring is written.

Parameters:

• qkv_conv_output (TensorValue) – The [total_seq_len, conv_dim] conv output.
• decay_per_token (TensorValue) – The [total_seq_len, nv] decays.
• beta_per_token (TensorValue) – The [total_seq_len, nv] beta gates.
• recurrent_state (BufferValue) – The [max_slots, nv, KD, VD] live pool, written only at a verify width of zero.
• ring (BufferValue) – The [ring_rows, nk, RING_LEN, record_stride] ring pool.
• slot_idx (TensorValue) – The [batch_size] live pool rows.
• ring_slot_idx (TensorValue) – The [batch_size] ring rows.
• input_row_offsets (TensorValue) – The [batch_size + 1] uint32 ragged offsets.
• verify_width (TensorValue) – The operand verify_width_operand() returns.

Returns:

The recurrence output, [total_seq_len, value_dim].

Return type:

TensorValue