Skip to content

4090/3090 native serve status never populates engine telemetry; pinned OMP client rejects sessions at init #28

Description

@alphastorm

Found during the 2026-08-31 fleet smoke. The released native Windows builds (4090 v0.3.1-qwen38-4090-mtp3.2, 3090 v0.3.0-qwen38-3090.1) return null for the engine telemetry block in /v1/ninfer/status (telemetry_available: false, materializing: null, ...) in every state, warm or idle. The pinned OMP 18.0.9 client validates status strictly (NInfer status field materializing is invalid) and refuses the session at provider init, so the heavy fleet lane cannot serve the pinned client today - the wire protocol itself works (responses complete; observed live).

Note: the 3090 lane passed the same smoke - its status shape apparently satisfies the validator; the failure reproduces only on the 4090 build. The 5090 container fixed this class in v0.4.0 (format_status_json emits the exact hierarchy with concrete zeros). Fix: adopt the v0.4.0 status emission in the durable-4090 rebase campaign (#33 upstream pulls fold in there too). Until then the heavy lane serves raw protocol clients but not the pinned OMP client from a cold start.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions