Struct PrecisionSchedule 

pub struct PrecisionSchedule { /* private fields */ }

Per-(layer, role) precision schedule built from MODEL.toml’s [precision] block. Lookups are O(1) for the common tables; per-layer overrides hit a small sorted set.

Implementations

impl PrecisionSchedule

pub fn build( router_dtype: Dtype, lm_head_dtype: Dtype, sensitive_block_indices: &[u16], sensitive_block_dtype: Dtype, default_dtype: Dtype, ) -> Self

Build from the four documented MODEL.toml fields:

• router_dtype: dtype for the MoE gate
• lm_head_dtype: dtype for the final unembedding
• sensitive_block_dtype + sensitive_block_indices: the “extra precision” tier for the first/last few blocks
• default_dtype: bulk fallback (typically Inherit)

For now, the simpler [precision] schema only exposes these four; per-tensor / per-layer YAML can extend later.

pub fn dtype_for(&self, layer_idx: Option<u16>, role: Role) -> Dtype

Resolve the target dtype for a tensor. layer_idx = None is used for non-layer tensors (embedding, lm_head, final norm). Lookup order:

1. Sensitive-layer override (only for Attention/Expert)
2. Per-role default
3. Global default

pub fn has_any_override(&self) -> bool

True iff the schedule will produce any non-Inherit overrides. Loaders can use this to skip the per-tensor lookups entirely when no overrides are configured (default case).

Trait Implementations

impl Clone for PrecisionSchedule

fn clone(&self) -> PrecisionSchedule

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for PrecisionSchedule

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

Formats the value using the given formatter.

impl Default for PrecisionSchedule

fn default() -> Self

Empty schedule — every lookup returns Inherit. Bit-exact equivalent to the pre-C.3 behaviour. MODEL.toml omits the [precision] block to opt into this default.