Struct DFlashLevers 

pub struct DFlashLevers {

    pub any_diagnostic_armed: bool,
    pub debug_dump: bool,
    pub debug_dump_full: bool,
    pub log_drafts: bool,
    pub block_dump: bool,
    pub block_dump_at_pos: usize,
    pub option_b_diag: bool,
    pub force_pattern: bool,
    pub force_noise_pattern: bool,
    pub force_no_ctx: bool,
    pub force_ctx_used: Option<usize>,
    pub precompute: bool,
    pub precompute_commit: bool,
    pub propose_warmup_n: usize,
    pub option_b: bool,
    pub option_b_no_ctx: bool,
    pub dflash2: bool,
    pub batch_propose_width: usize,
    pub draft_cap: Option<usize>,
    pub verify_trace: bool,
    pub precompute_dump: bool,
    pub ctx_parity_dump: bool,
    pub no_decode_append: bool,
    pub full_precompute: bool,
    pub ctxlen_probe: bool,
    pub dspark_markov: bool,
    pub conf_tau: f32,
    pub dspark_shift: Option<bool>,
    pub dspark_anchor_bias: bool,
    pub dspark_conf_trace: bool,
}

2026-09-25: Diagnostic and A/B levers for one loaded DFlash drafter.

Fields

any_diagnostic_armed: bool

2026-09-25: A variable in GRAPH_SUPPRESSING_DIAGNOSTICS is set, at any value. forward_block then runs eagerly, without CUDA graphs. This is presence, not truth: METRALE_DFLASH_BLOCK_DUMP=0 suppresses graphs and arms no dump.

debug_dump: bool

2026-09-25: METRALE_DFLASH_DEBUG_DUMP=1: log the first 10 BF16 values of key forward_block intermediates.

debug_dump_full: bool

2026-09-25: METRALE_DFLASH_DEBUG_DUMP_FULL=1: one-shot whole-tensor dumps to /tmp, and the first propose’s drafts to the log.

log_drafts: bool

2026-09-25: METRALE_DFLASH_LOG_DRAFTS=1: log the first propose’s drafts (one-shot per model).

block_dump: bool

2026-09-25: METRALE_DFLASH_BLOCK_DUMP=1: one-shot /tmp dumps of the block logits and per-layer intermediates, gated by block_dump_armed_at.

block_dump_at_pos: usize

2026-09-25: METRALE_DFLASH_BLOCK_DUMP_AT_POS=<n>: arm the block dump only at decode position >= n. Unset or unparseable is 0.

option_b_diag: bool

2026-09-25: METRALE_DFLASH_OPTION_B_DIAG=1: one-shot read-back of the K row layer 0 just wrote to the paged drafter cache, compared with its source.

force_pattern: bool

2026-09-25: METRALE_DFLASH_DEBUG_FORCE_PATTERN=1: overwrite the first ctx slot forward_block reads with a deterministic pattern.

force_noise_pattern: bool

2026-09-25: METRALE_DFLASH_DEBUG_FORCE_NOISE_PATTERN=1: the same, for the noise rows.

force_no_ctx: bool

2026-09-25: METRALE_DFLASH_DEBUG_CTX_OFF=1: drop ctx conditioning (eff_ctx = 0).

force_ctx_used: Option<usize>

2026-09-25: METRALE_DFLASH_DEBUG_CTX_USED=<n>: cap eff_ctx at n (it stays bounded by the ctx rows present and ctx_window).

precompute: bool

2026-09-25: METRALE_DFLASH_PRECOMPUTE=1: also run the ctx K/V precompute from forward_block; propose_drafts runs it either way.

precompute_commit: bool

2026-09-25: METRALE_DFLASH_PRECOMPUTE_COMMIT=1: let that diagnostic run write to the paged cache (its commit argument).

propose_warmup_n: usize

2026-09-25: METRALE_DFLASH_PROPOSE_WARMUP_N=<n>: eager passes per block width before graph capture. Unset or unparseable is 2.

option_b: bool

2026-09-25: The paged drafter cache for ctx K/V (“Option B”). On unless METRALE_DFLASH_OPTION_B=0; resolved by option_b_from in dflash_head.rs.

option_b_no_ctx: bool

2026-09-25: METRALE_DFLASH_OPTION_B_NO_CTX=1: pass ctx_count = 0, so paged attention sees only the γ K/V written in-layer.

dflash2: bool

2026-09-25: The DFlash2 conv and selector path. Runs when the checkpoint carries its components, unless METRALE_DFLASH2=0.

batch_propose_width: usize

2026-09-25: METRALE_DFLASH_BATCH_PROPOSE=<width> caps the cross-sequence batch. Unset is usize::MAX (as wide as the scratch bands allow); 0 or 1 makes propose_batch_max return 1.

draft_cap: Option<usize>

2026-09-25: METRALE_DFLASH_DRAFT_CAP=<n>: return at most n drafts per propose. None caps at the block width in flight (block_g).

verify_trace: bool

2026-09-25: METRALE_DFLASH_VERIFY_TRACE=1: log every propose’s drafts before the row-0 drop and the cap.

precompute_dump: bool

2026-09-25: METRALE_DFLASH_PRECOMPUTE_DUMP=1: one-shot /tmp dumps of the ctx K/V precompute intermediates.

ctx_parity_dump: bool

2026-09-25: METRALE_DFLASH_CTX_PARITY_DUMP=1: one-shot /tmp dump of the accumulated ctx hidden rows.

no_decode_append: bool

2026-09-25: METRALE_DFLASH_DEBUG_NO_DECODE_APPEND=1: skip the ctx append at the start of propose_drafts.

full_precompute: bool

2026-09-25: METRALE_DFLASH_DEBUG_FULL_PRECOMPUTE=1: the non-batched propose recomputes the whole ctx prefix (committed = 0) each step.

ctxlen_probe: bool

2026-09-25: METRALE_DFLASH_CTXLEN_PROBE=1: warn when ctx_positions is not strictly increasing, and log ctx_len against the position when the position is a multiple of 16.

dspark_markov: bool

2026-09-25: The sequential Markov fixup. Runs when the drafter carries the head, unless METRALE_DSPARK_MARKOV=0, which takes the plain per-row argmax loop.

conf_tau: f32

2026-09-25: METRALE_DSPARK_CONF_TAU=<t>: sigmoid-space acceptance threshold for the confidence head. Unset or unparseable is 0.0, which leaves the head off.

dspark_shift: Option<bool>

2026-09-25: METRALE_DSPARK_SHIFT=1|0 forces the shifted-row convention on or off; any other value (None) defers to the drafter config.

dspark_anchor_bias: bool

2026-09-25: Row 0 carries the Markov bias. On unless METRALE_DSPARK_ANCHOR_BIAS=0. Confidence truncation also requires it: an unbiased row never writes its confidence slot.

dspark_conf_trace: bool

2026-09-25: METRALE_DSPARK_CONF_TRACE=1: log the confidence logits and sigmoids.

Implementations

impl DFlashLevers

pub fn from_env() -> Self

2026-09-25: Resolve from the environment. The head calls it once, in from_weights.rs; per-step code reads the head’s levers field.

pub fn defaults() -> Self

2026-09-25: What from_env resolves to with none of these variables set (checked by nothing_set_resolves_to_defaults).

pub fn block_dump_armed_at(&self, position: usize) -> bool

2026-09-25: The block dump is armed for this decode position.

Trait Implementations

impl Clone for DFlashLevers

fn clone(&self) -> DFlashLevers

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DFlashLevers

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for DFlashLevers

fn default() -> DFlashLevers

Returns the “default value” for a type.

impl PartialEq for DFlashLevers

fn eq(&self, other: &DFlashLevers) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for DFlashLevers

impl StructuralPartialEq for DFlashLevers