Record final RTMLib deployment status

servers/unraid/tail2-pose-server/DEPLOYMENT.md

@@ -6,7 +6,8 @@ GPU-only, receive-only NDI inference service. It never saves frames/crops and ha
 - Token: `/mnt/user/appdata/tail2-pose-server/secrets/api-token` (mode 0600)
 - Source: `TAIL 2_1621D2 (OBSBOT)` / `192.168.50.207`
 - Coordinate contract: unmodified NDI buffer, no horizontal flip
-- Models: official OpenMMLab RTMDet-tiny and RTMW-m 256x192. RTMDet-m/RTMW-l were tested first but failed the 15 FPS requirement. Model hashes are in `models/SHA256SUMS`.
+- Default backend: RTMLib 0.0.15, official YOLOX-m HumanArt+COCO 640x640 and RTMPose-m Body17 256x192, ONNX Runtime GPU with CUDA EP first.
+- Alternatives: `openmmlab`, `rtmlib_wholebody` (DWPose-m 133), and `rtmlib_rtmw_l`. The original RTMDet-tiny + RTMW-m OpenMMLab path remains available as the baseline.
 - NDI binding: cyndilib 0.1.1 and its distributed NDI runtime. NDI's redistributable license applies; do not redistribute the image outside this private deployment without reviewing that license.
 
 Rotate the integration token after onboarding clients:
@@ -19,7 +20,9 @@ docker compose restart tail2-pose-server
 
 ## Current acceptance status
 
-`blocked` for the requested performance SLO only. Functional/API/privacy/stability acceptance passed, including a 600.072-second NDI soak with 6,355 strictly increasing messages and no WS failures. Live p50/p95/p99 source-to-send latency was 91.629/122.584/139.033 ms (~10.59 output FPS), so p95 exceeds 100 ms and output is below 15 FPS. Offline 1/2/4-person p50 was 144.228/154.546/153.313 ms. No OOM; maximum selected-GPU memory was 14,145 MiB and temperature 43 C.
+The RTMLib optimization passes the offline inference target. Body17 1/2/4-person p50 is 19.405/24.543/34.781 ms and p95 is 19.734/25.727/36.191 ms. DWPose-m WholeBody 1/2/4-person p50 is 19.905/26.193/37.770 ms and p95 is 20.237/27.489/39.592 ms. All cases used 100 warmups and 300 measured iterations with real 1/2/4-person detector results.
+
+Live acceptance is currently blocked externally: the camera IP responds, but the NDI finder repeatedly reports `last_sources=[]`. Re-run the live WebSocket and 30-minute soak acceptance after `TAIL 2_1621D2 (OBSBOT)` resumes broadcasting.
 
 The one-shot validation services are profile-gated and do not run during normal Arcane lifecycle:
 

servers/unraid/tail2-pose-server/compose.yaml

@@ -166,6 +166,20 @@ services:
         } > "$$out/process-map.csv"
         chmod -R a+rX,u+w "$$out"
 
+  gpu-final-snapshot:
+    image: tail2-pose-server:0.1.0
+    container_name: tail2-pose-gpu-final-snapshot
+    gpus: all
+    restart: "no"
+    volumes:
+      - ./validation:/validation
+    entrypoint: ["/bin/bash", "-lc"]
+    command:
+      - |
+        nvidia-smi --query-gpu=index,uuid,pstate,utilization.gpu,utilization.memory,memory.used,memory.total,power.draw,clocks.current.sm,temperature.gpu --format=csv,noheader,nounits > /validation/gpu-final-snapshot.csv
+        nvidia-smi --query-compute-apps=gpu_uuid,pid,process_name,used_gpu_memory --format=csv,noheader,nounits > /validation/gpu-final-processes.csv
+        chmod 0666 /validation/gpu-final-snapshot.csv /validation/gpu-final-processes.csv
+
   benchmark-rtmlib-body:
     profiles: ["validation"]
     image: tail2-pose-server:0.1.0
@@ -265,6 +279,7 @@ services:
         done
 
   benchmark-openmmlab:
+    profiles: ["validation"]
     image: tail2-pose-server:0.1.0
     container_name: tail2-pose-benchmark-openmmlab
     gpus: all