Module run_metrics 

Run mailboxes — the observability surfaces that stay process-global on purpose, and the one call that keeps them honest across a model swap.

Most model-derived state in Atlas is carried: SchedCtx, ForwardContext, ModelLevers, OpCache. A handful of counters cannot be, because their readers cannot be handed a carrier — /metrics answers from an HTTP handler thread and the dashboard polls from the TUI thread, both while the scheduler is mid-step and holding its own context. A process-global address is what an observability surface is for.

That leaves the scoping problem: after a swap the counters would describe two models at once. reset_for_new_run solves it from the other end — the values stay reachable at a fixed address, but they start clean when a run does, so a reader asking “what is the prefix-cache hit rate” gets the rate for the model now running. Prometheus reads the reset as a counter restart, which it already handles.

Called from AtlasCudaBackend::new, which is where a model’s GPU state begins. That is deliberately upstream of the first kernel lookup, so the kernel audit records only this model’s modules.

Structs

RunMetrics

The process’s single run mailbox.

Functions

cache_counts_this_run

Prefix-cache activity since the CURRENT model was loaded.

metrics

Read the mailbox.

reset_for_new_run

Begin a new run’s accounting. Called when a new model’s backend is built.