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_state_fold

gated_delta_state_fold()

max.nn.state_space.gated_delta_state_fold(recurrent_state, ring, row_ids, ring_row_ids, num_accepted, verify_width)

source

Applies a verify’s accepted ring records to the state pool.

Advances recurrent_state over the first num_accepted[b] records written by gated_delta_recurrence_verify_ring_fwd(), for every layer in one launch. A row with num_accepted[b] == 0 is left unchanged. Nothing runs at a verify width of zero, where the verify wrote no records.

Parameters:

• recurrent_state (BufferValue) – The [max_slots, nv, KD, VD] live pool, folded in place.
• ring (BufferValue) – The [ring_rows, nk, RING_LEN, record_stride] ring pool.
• row_ids (TensorValue) – The [num_layers, batch_size] live pool rows.
• ring_row_ids (TensorValue) – The [num_layers, batch_size] ring rows.
• num_accepted (TensorValue) – The [batch_size] records to fold. A count past RING_LEN names a row the verify was too long to record and committed itself, and is skipped.
• verify_width (TensorValue) – The operand verify_width_operand() returns.

Return type:

None