Struct PrewarmStats 

pub(super) struct PrewarmStats {
    pub warmed: usize,
    pub total_compile_ms: u64,
    pub total_fsync_ms: u64,
    pub fully_cached: bool,
    pub skipped: usize,
    pub fresh_compiles: usize,
    pub engine_count_delta: i64,
    pub engine_count_before: usize,
    pub engine_count_after: usize,
    pub max_warmed_seq: usize,
}

Aggregate per-worker statistics returned by trt_prewarm.

total_compile_ms and total_fsync_ms sum across only the shapes that completed successfully on this worker’s shard, whether they were cache hits or fresh compiles. They are intended for the "TensorRT pre-warm complete" summary log emitted by the worker.

Fields

warmed: usize

total_compile_ms: u64

total_fsync_ms: u64

fully_cached: bool

true when the dimensional-extreme coverage check determined the entire shard was already cached and the remaining shapes were skipped. false on cold cache, on a fresh compile, or when the check phase detected at least one slow (≥ CACHE_HIT_THRESHOLD_MS) shape.

skipped: usize

Number of shapes in the shard that were skipped because fully_cached was determined to be true. Zero on cold cache or when all shapes were run.

fresh_compiles: usize

Number of shapes that reported a fresh compile (!cache_hit and Ok(_) from session.run()). Used together with engine_count_delta by the worker to decide whether the on-disk artifacts match what the per-shape logs claimed.

engine_count_delta: i64

Net .engine file count change across this worker’s prewarm sweep (count_after_last_shape - count_before_first_shape). SM-filtered: reflects only plans matching the worker’s GPU compute capability, so a stale _sm89.engine next to a fresh _sm120.engine does not silently zero out the delta on a Blackwell worker. Compared against fresh_compiles to detect compile-success-without-persistence.

engine_count_before: usize

.engine file count observed in engine_cache_dir before the worker ran any of its shard’s shapes. SM-filtered: counts only plans matching the worker’s GPU compute capability (see the sm parameter on trt_prewarm). When SM detection failed and sm == None, falls back to the legacy unfiltered count.

engine_count_after: usize

.engine file count observed in engine_cache_dir after the worker finished its shard (post final fsync). SM-filtered with the same semantics as engine_count_before.

max_warmed_seq: usize

Largest sequence length among the shapes this worker successfully warmed (fresh compile or warm-cache hit). Zero when no shape succeeded (e.g. every compile failed, the worker-3 seq=8192 failure mode). Folded into the pool-wide warmed_seq_ceiling atomic by the worker so super::jit_guard::TrtJitGuard knows the highest sequence tier with a persisted engine plan. See worker.rs.