Struct LoadTimings 

#[non_exhaustive]
pub struct LoadTimings {
    pub tokenizer_load: Duration,
    pub token_table_read: Duration,
    pub probe: Option<Duration>,
    pub parity_check_total: Option<Duration>,
    pub parity_cpu_reference: Option<Duration>,
    pub parity_accelerator: Option<Duration>,
    pub parity_cache: ParityCacheOutcome,
    pub session_creates: Vec<SessionCreate>,
    pub prewarm: Option<Duration>,
    pub prewarm_per_length: Vec<(usize, Duration)>,
    pub total: Duration,
}

Where an E40 load spent its time; see Embedder::load_timings. Phases are wall-clock durations of the work as Embedder::load ran it; they nest (session_creates happen inside the probe and the parity check) and do not add up to total exactly.

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

tokenizer_load: Duration

Reading tokenizer.json.

token_table_read: Duration

Reading and validating embed_tokens.f16.npy.

probe: Option<Duration>

Probe inference on the accelerator (Auto only), including its graph build.

parity_check_total: Option<Duration>

The whole parity check, from the cache lookup to the stored verdict. None when no check ran (CPU provider).

parity_cpu_reference: Option<Duration>

Inside the check: CPU reference sessions, tokenizer reload and CPU embeddings. None on a cache hit.

parity_accelerator: Option<Duration>

Inside the check: the accelerator’s embeddings of the parity prompts, including its graph builds. None on a cache hit.

parity_cache: ParityCacheOutcome

What the verdict cache did.

session_creates: Vec<SessionCreate>

Every graph built during the load, in build order, on both providers.

prewarm: Option<Duration>

Pre-warm after admission (EmbedderConfig::with_prewarm): the graph builds still missing plus one dummy run per graph. Its graph builds are also in session_creates. None when no pre-warm ran (disabled, CPU provider, or an accelerator not admitted).

prewarm_per_length: Vec<(usize, Duration)>

Core ML only: the pre-warm time of each function as (length in tokens, time), shortest first. The cpu_and_gpu lengths (256, 512) carry the shader build, up to 51 s cold. Empty for every other backend and when no pre-warm ran.

total: Duration

Embedder::load from entry to return.

Trait Implementations

impl Clone for LoadTimings

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LoadTimings

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

Formats the value using the given formatter.

impl Default for LoadTimings

fn default() -> Self

Returns the “default value” for a type.

impl Eq for LoadTimings

impl PartialEq for LoadTimings

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for LoadTimings