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_conv1d_verify_fwd

gated_delta_conv1d_verify_fwd()

max.nn.state_space.gated_delta_conv1d_verify_fwd(qkv_input_ragged, conv_weight, conv_state, slot_idx, input_row_offsets, verify_width, *, rollback)

source

Runs gated_delta_conv1d_fwd() in a speculative verify graph.

At a verify width of zero the forward launch writes the window and the rollback launch does nothing, leaving its output unwritten. Otherwise the forward launch only reads the window and the rollback launch writes it.

The forward launch with drafts also copies its input for the rollback to replay. Every row is then one token and its drafts, so the copy is the whole input. Every other launch leaves the copy unwritten.

Parameters:

• qkv_input_ragged (TensorValue) – The [total_seq_len, conv_dim] conv input.
• conv_weight (TensorValue) – The [conv_dim, kernel_size] depthwise weights.
• conv_state (BufferValue) – The [max_slots, conv_dim, kernel_size - 1] pool.
• slot_idx (TensorValue) – The [batch_size] uint32 pool rows.
• input_row_offsets (TensorValue) – The [batch_size + 1] uint32 ragged offsets.
• verify_width (TensorValue) – The operand verify_width_operand() returns.
• rollback (bool) – Whether this is the rollback’s launch.

Returns:

The conv output, [total_seq_len, conv_dim], and the copy of the input, [batch_size * (1 + K), conv_dim].

Return type:

tuple[TensorValue, TensorValue]