Struct DflashProposerState 

pub struct DflashProposerState {

    pub block_table: Vec<u32>,
    pub seq_len: usize,
    pub last_num_drafted: usize,
    pub prefill_done: bool,
    pub ctx_hidden_acc: DevicePtr,
    pub ctx_len: usize,
    pub last_num_accepted: usize,
    pub skip_next_decode_append: bool,
    pub max_ctx_len: usize,
    pub ctx_slot_bytes: usize,
    pub block_table_dev: Option<DevicePtr>,
    pub ctx_count_drafter: usize,
    pub max_ctx_count_drafter: usize,
    pub ctx_committed: usize,
    pub ctx_positions: Vec<i32>,
}

2026-09-25: Per-sequence DFlash drafter state. One block table serves every drafter layer: the paged cache has a pool per layer, addressed by the same slots.

Fields

block_table: Vec<u32>

2026-09-25: Host copy of the drafter block table.

seq_len: usize

last_num_drafted: usize

2026-09-25: Drafts the last propose returned; after_verify and free_state reset it to 0.

prefill_done: bool

ctx_hidden_acc: DevicePtr

2026-09-25: max_ctx_len rows of ctx_slot_bytes: captured target hiddens, one row per ctx position, the first ctx_len valid. Model-engine appends rows and slides the window when it is full; propose_drafts appends while there is room.

ctx_len: usize

2026-09-25: Valid rows in ctx_hidden_acc, at most max_ctx_len.

last_num_accepted: usize

skip_next_decode_append: bool

2026-09-25: Model-engine has already appended the latest capture; the next propose_drafts skips its own append and clears the flag.

max_ctx_len: usize

2026-09-25: Rows ctx_hidden_acc holds (the window of alloc_state_windowed).

ctx_slot_bytes: usize

2026-09-25: Bytes per ctx_hidden_acc row: target_layer_ids.len() * target_hidden_size * 2.

block_table_dev: Option<DevicePtr>

2026-09-25: Device copy of block_table, made once when the first propose allocates the table; None before that and after free_state.

ctx_count_drafter: usize

2026-09-25: Ctx rows of the paged cache the γ block attends to; propose_drafts sets it to ctx_len.

max_ctx_count_drafter: usize

2026-09-25: Slots the allocated block table covers (block_table.len() * BLOCK_SIZE).

ctx_committed: usize

2026-09-25: Ctx rows [0, ctx_committed) whose K/V is already in the paged cache; the next precompute starts here. Model-engine’s window slides reset it to 0.

ctx_positions: Vec<i32>

2026-09-25: One RoPE position per ctx row (len() == ctx_len), stamped when the row is appended and kept when the window slides. Model-engine’s update_dflash_ctx_len_after_prefill seeds prefill row i with i.

Trait Implementations

impl ProposerState for DflashProposerState

fn as_any(&self) -> &dyn Any

fn as_any_mut(&mut self) -> &mut dyn Any