NOC Thoughts / Experiments Simulated data
Local inference replay

Local LLM Observatory

Follow a chat request as it hands a render job to Rapid. Replay the moment the chat model releases the GPU and the renderer takes over.

qwen3.5-9bRapid
AMD Radeon RX 6900 XT

GPU timeline

GPU busy68%
VRAM used8.7 GiB
Faded traces preview the rest of the replay. VRAM is device-wide.

Handoff state

Chat inference
Phase timer
0s in phase
Rapid worker
Idle
Routed now
qwen3.5-9b1 active request
Last completed
qwen3.5-9b63.8 tok/s · prior request
9B model
Ready
GPU release
Pending

A chat request is routed to the ready 9B model. Rapid is idle.

Evidence behind this phase

Runtime readiness and routed request counts
The timeline aligns a loaded model with current requests. Token rate comes from the most recently completed request.

Jump to a milestone