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_fwd

gated_delta_recurrence_fwd()

max.nn.state_space.gated_delta_recurrence_fwd(qkv_conv_output, decay_per_token, beta_per_token, recurrent_state, slot_idx, input_row_offsets)

source

Applies the gated delta recurrence pass, mutating a slot-indexed state pool in place.

recurrent_state is a mutable pool of shape [max_slots, nv, KD, VD] and the kernel reads/writes slot slot_idx[batch_item] directly. There is no recurrent_state_out graph output: the pool is mutated in place.

Parameters:

• qkv_conv_output (TensorValue) – The [total_seq_len, conv_dim] output of gated_delta_conv1d_fwd().
• decay_per_token (TensorValue) – The [total_seq_len, num_value_heads] decays.
• beta_per_token (TensorValue) – The [total_seq_len, num_value_heads] beta gates.
• recurrent_state (BufferValue) – The [max_slots, nv, KD, VD] mutable pool.
• slot_idx (TensorValue) – The [batch_size] uint32 slot indices into the pool.
• input_row_offsets (TensorValue) – The [batch_size + 1] uint32 ragged offsets.

Returns:

The recurrence output, [total_seq_len, value_dim].

Return type:

TensorValue