pub struct Fixed {Show 16 fields
pub hidden: DevicePtr,
pub residual: DevicePtr,
pub logits: DevicePtr,
pub tokens: DevicePtr,
pub meta: AttnMetadataDev,
pub batch_meta: AttnMetadataDev,
pub verify_meta: AttnMetadataDev,
pub draft: Option<DraftFixed>,
pub verify_batch_meta: AttnMetadataDev,
pub verify_wy_tables: DevicePtr,
pub verify_batch_tokens: DevicePtr,
pub ffn_act_q8: DevicePtr,
pub k_pools: Vec<DevicePtr>,
pub v_pools: Vec<DevicePtr>,
pub block_size: u32,
pub cache_stride: u64,
}Expand description
2026-09-28: Device addresses that never move after boot.
Fields§
2026-09-28: The residual stream (BufferArena::hidden_states).
residual: DevicePtr2026-09-28: The copy of the stream the norms write (BufferArena::residual).
logits: DevicePtr2026-09-28: The logits buffer the step returns.
tokens: DevicePtr2026-09-29: Where a verify’s argmax writes each row’s token (i32 per row): the
scratch buffer’s start, which the verify reads back.
meta: AttnMetadataDev2026-09-28: The single-sequence step’s attention metadata, at its fixed upload
address. max_blocks_per_seq is ignored: a step supplies it.
batch_meta: AttnMetadataDev2026-09-28: The multi-sequence step’s attention metadata (one row per sequence), at its
fixed upload address; max_blocks_per_seq is ignored.
verify_meta: AttnMetadataDev2026-09-29: The MTP verify step’s attention metadata (one row per verified token), at
its fixed upload address; max_blocks_per_seq is ignored.
draft: Option<DraftFixed>2026-09-29: The MTP draft head’s buffers; None without a bound draft head.
verify_batch_meta: AttnMetadataDev2026-09-30: The batched MTP verify’s attention metadata (one row per verified row,
stage_verify_metadata at scratch + 32768); max_blocks_per_seq is ignored.
verify_wy_tables: DevicePtr2026-09-30: The batched verify’s per-GDN-layer WY pointer tables
(upload_verify_wy_tables, VERIFY_WY_LAYER_STRIDE_BYTES apart); NULL without MTP.
verify_batch_tokens: DevicePtr2026-09-30: Where the batched verify’s argmax writes each row’s token: the mapped host
blob’s device alias, or scratch when there is none (verify_rows_argmax).
ffn_act_q8: DevicePtr2026-09-28: The quantized-activation scratch the NVFP4 MMQ GEMMs read
(BufferArena::ffn_act_q8), sized for the widest batch.
k_pools: Vec<DevicePtr>2026-09-28: K pool per attention layer (PagedKvCache::k_pool_ptr).
v_pools: Vec<DevicePtr>2026-09-28: V pool per attention layer.
block_size: u322026-09-28: Tokens per KV block.
cache_stride: u642026-09-28: PagedKvCache::cache_stride.