Fast-Tier Bakeoff Runtime Restoration Evidence¶
Date: 2026-07-08
This note records the Heavy health observations and final runtime state after the disruptive Fast-tier candidate matrix.
Heavy continuity evidence:
docker inspect vllm-gptoss120:
name=/vllm-gptoss120 started=2026-07-07T08:08:16.490006854Z running=true status=running restart_count=0
Selected /health 200 log observations during the matrix window:
2026-07-08T16:59:42.255817263Z GET /health HTTP/1.1 -> 200 OK
2026-07-08T17:03:33.924299374Z GET /health HTTP/1.1 -> 200 OK
2026-07-08T17:10:43.597347407Z GET /health HTTP/1.1 -> 200 OK
2026-07-08T18:40:48.536277436Z GET /health HTTP/1.1 -> 200 OK
2026-07-08T18:43:56.317369855Z GET /health HTTP/1.1 -> 200 OK
Final serve status command:
Final observed status:
SERVE CONTAINER PORT DOCKER HEALTH
heavy vllm-gptoss120 30002 running 200
fast vllm-qwen36 30003 running 200
fast-qwen36-35b-a3b vllm-fast-qwen36-35b-a3b 39010 exited -
fast-gemma4-31b vllm-fast-gemma4-31b 39011 exited -
fast-glm47-flash-sglang sglang-fast-glm47-flash 39012 exited -
fast-glm47-flash-llamacpp llamacpp-fast-glm47-flash 39013 exited -
fast-devstral-small2 vllm-fast-devstral-small2 39014 exited -
GPU memory (index, used MiB, total MiB):
0, 27056, 32607
1, 89386, 97887
Direct health probes:
Conclusion: Heavy stayed on the same running container with restart count 0
and health 200 observations during the matrix window and final handoff.
Production Fast was restored to vllm-qwen36 on port 30003, and all
experimental Fast candidate serves were stopped or absent at final handoff.