Module default_tier_boot 

2026-10-01: Default-Tier Boot: a boot-only check of the serve’s memory plan. Once the serve is up it sends one short warmup wave, then reads GET /memory and reports the KV pool’s block count and whether the serve’s device footprint stays inside its --gpu-memory-utilization budget.

Every other gate pins a reduced-precision serve; this one exists for the serve’s default precision tier, whose KV pool can shrink without any of them noticing. It lands informational: the metrics are recorded and the verdict is info until bounds are declared for it in BENCH.toml.

The pins (the benchmark’s definition, not parameters):

• Warmup: WARMUP_REQUESTS concurrent chat requests, each a distinct prompt of about WARMUP_PROMPT_TOKENS tokens with max_tokens = WARMUP_OUTPUT_TOKENS, temperature 0, seed 0. It brings the serve to the state a first real wave leaves it in.
• Read: one GET /memory after the wave has finished.

Owner: bench, default-tier-boot. Invariants: the run verdict is never a pass (memory::verdict_for).

Re-exports

pub use memory::Footprint;

pub use memory::MemoryReport;

pub use memory::footprint;

pub use memory::verdict_for;

Modules

memory

2026-10-01: What a server reports about its memory (GET /memory), and the pure arithmetic the default-tier boot check derives from it. The server fills MemoryReport; the driver reads it back, so the one type is the wire contract on both sides.

Structs

DefaultTierBoot

Constants

DESCRIPTOR

MEMORY_PATH

2026-10-01: The server’s memory endpoint.

METADATA