Skip to content

Current compatible Flare baseline — reconstruction

This is a public-safe description, not an executable operator configuration. Use managed recipe load/status/logs/unload and the owner-controlled restoration plan. No credential value, private endpoint, device UUID or operator path is required in this document; supply protected secret references through the existing supported resolver.

Pin / control Executed baseline
Upstream NVIDIA/flare-local-inference d5fd9824593c2a70f6bdc001b39f16b4c5b3b0b4
Compatibility source 0b21d7f4883ac6f2b099df55813e8be06777b9df; minimal exact-byte legacy GLM-template upgrade
Image sha256:b3c33bcf858d3764dbe155cbda5a006d364fa3a4309922bcb92e51e925c05804
Checkpoint/package model identity dec48abd33efa73c3bb7c95b74eee10cad34f9be; retained signed package, ordinary authorization
Topology TP2, configured concurrency four, fixed two draft tokens, adaptive verification disabled
Host cache Unchanged upstream default allocation, 32 GiB host budget; no 75% candidate
Driver 615.71.10; hardware/runtime comparison boundary relative to older observations
Identity contract Explicit advertised alias in catalog id; exact canonical package ID in catalog root and model-bearing responses; verified managed incarnation
Workload Frozen prior body bytes and prospective action order; fixtures, three fresh fixed1 repetitions, then small12 and pageheavy

The exact frozen action, adapter, bodies and native per-phase evidence are retained in the bundle archives. Do not execute the private action as a generic public workflow. A schema fixture precedes measurements; failed predecessors remain separate attempted windows. Current-native host decision metrics are retained as emitted, including their reason labels. Do not synthesize retired row/page-pressure counters or infer pressure from missing metrics.

Measured startup host geometry was 52 rows and 7,314 pages under the 32 GiB budget. The canonical run completed 391 adapter requests plus one smoke and restored the incumbent; performance eligibility remains false. Exact package file/digest and signed verification receipts belong in the final sanitized identity artifacts; the checkpoint ID alone is not a file-integrity proof. This baseline does not clear the earlier full numerical gate or authorize a serving change.

Measured hardware: two RTX PRO 6000 Blackwell Max-Q GPUs, TP2, historical configured 300 W limit each (not reverified in this run). Direct online requests were serial within each phase; configured concurrency four is not a C4 measurement. Prime and reverse-replay populations are separated; small12 and pageheavy measured lanes were warm GPU hits. The canonical alias is a request/catalog role and the signed package ID is a response role.