The city plans citywide RoCEv2 cutover for GPU trainers. What deployment sequencing best reduces risk?

Citywide RoCEv2 cutover for GPU trainers should be sequenced as a staged enablement: first a lab leaf pair or limited pilot that validates lossless priorities, markings, and congestion behavior, then wider production rollout. Piloting reduces blast radius when Cos maps, PFC watchdogs, or NIC settings are wrong on the municipal fabric. Staging RoCEv2 enablement in a lab leaf pair or limited pilot to validate only ECMP hashing efficiency before citywide cutover leaves the lossless QoS behavior unproven, since balanced hashing does not confirm PFC, ECN marking, or priority mappings that GPU trainers require. Staging RoCEv2 enablement in a lab leaf pair or limited pilot to validate only BGP EVPN route exchange before citywide cutover proves control-plane reachability but not lossless forwarding, leaving PFC thresholds, ECN behavior, and congestion response untested. Staging RoCEv2 enablement in a lab leaf pair or limited pilot to validate PFC pause on best-effort traffic before citywide cutover tests the wrong priority, since pausing best-effort does not validate the lossless RoCEv2 class and risks head-of-line blocking rather than protecting trainer traffic. Exam caveat: a successful lab pilot must exercise representative trainer traffic and the same Cos/DSCP/PFC/ECN templates intended for production—not only ICMP or a single flow. Operational check: stand up a civic lab leaf pair (or constrained pilot rack), enable RoCEv2 with production-like QoS, pass lossless and congestion validation tests, then expand cutover in controlled waves with rollback criteria documented.