Deploy Qwen3.8-Flash-Next NVFP4 on the dual DGX Spark pair

[janitor-manager]

Configure and validate a persistent Qwen3.8-Flash-Next NVFP4 serving recipe for the static and shock DGX Spark pair. Record the model config and serving/deployment notes in models.server, include the new machines in the repository machine inventory, synchronize the repository through Git, stage the NVFP4 weights from Smarty's model vault on both nodes, install and start the distributed service, and measure throughput at the largest context the two-node setup can sustain. Preserve existing services and machine-specific configuration. Attach bounded, secret-free deployment and benchmark evidence here.

[janitor-manager]

Committed and pushed the dual-Spark model service, manifest, pinned vLLM ARM64 image, and deployment runbook. Both Sparks are synchronized to the same main revision and the pinned Apache-2.0 compatibility checkout. RoCE reachability and the official image digest/runtime flags are verified. The 124 GB checkpoint copy from Smarty is running to static; a deployment watcher will compare file manifests, mirror to shock over RoCE, install/start the workers in rank order, and measure a near-maximum-context request. The service is currently stopped and disabled on both hosts pending complete model files.

[janitor-manager]

Follow-up compatibility check found that the older dedicated Qwen image lacks a symbol required by the pinned overlay set, so I replaced it with the current ARM64 nightly pinned by manifest digest and synchronized main revision 76ac031 to both Sparks. The replacement image pull is in progress on static; I will require the overlays to import cleanly before service startup.

[janitor-manager]

The deployment config now uses current vLLM 0.30.1rc1.dev245 stock Qwen4Exp modules; the earlier overlays failed against this runtime and are removed. Stock ModelOpt, QSA, PLE, and platform imports plus all launcher flags passed on static. I pulled the ARM64 manifest by digest, mirrored the image to shock over RoCE, and verified both tags resolve to the same pinned image ID. The 124 GB weight stream from Smarty to static continues; service startup waits for both local copies to be complete.