pub const VERIFY_WY_TABLE_SEQS: usize = 128;Expand description
2026-09-25: Batched-verify WY pointer-table layout. TransformerModel::upload_verify_wy_tables
(model-engine verify_e2.rs) writes it; the qwen3_ssm GdnStates::Multi arm and the
write-on-accept fold read it.
Per GDN layer, VERIFY_WY_TABLES_PER_LAYER tables of VERIFY_WY_TABLE_SEQS u64
entries lie back to back: table 0 holds each sequence’s h_state, table t + 1 its
intermediate t. A k-row verify writes at most the first k tables and leaves every
other entry zero. The strides do not depend on k.
This constant is the most sequences one batched verify holds: it sizes the tables and
the verify hidden and catch-up stashes (impl_a1/spec_buffers.rs), the GDN carry and
write-on-accept tables, and can_batch_verify_dispatch (verify_e.rs) admits at most
this many. 2026-09-29: 128, the serve’s widest batch (DECODE_META_MAX_ROWS), so one
verify forward covers every speculating sequence: a verify split in chunks reads the MoE
expert weights once per chunk. The tables cost 3,840 B per sequence; the SSM verify pools
are sized by the dispatch cap, not by this (ssm_reserve::mtp_state_slots).