Troubleshooting
Begin with the actual state
Section titled “Begin with the actual state”Record the app version, selected artifact, backend, exact context, GPU layers and original error. Open System for hardware capture time and runtime logs. Reproduce once with one controlled change. Repeatedly patching several settings can hide the real cause.
| Symptom | First check |
|---|---|
| Library is empty | Add the actual model directory or import a GGUF; refresh after files finish downloading |
| Projector cannot load | Select the language-model artifact; projector-only chat is unsupported |
| Split GGUF blocked | Current core does not qualify incomplete shard sets |
| Load button disabled | Read context, artifact, backend-readiness and memory-gate messages |
| Context rejected | Use an exact whole integer within host and model bounds |
| CUDA unavailable | Check detected NVIDIA GPU, selected backend and required runtime components |
| “Another operation” / busy | Wait for loading or refresh to settle; Stop an active generation first |
| Out of memory | Refresh availability, inspect host/VRAM budgets and change placement deliberately |
| Slow token generation | Separate layer offload, CPU work, prefill, reasoning and decode; do not assume Strata is active |
A successful load with slow tokens
Section titled “A successful load with slow tokens”The 0.0.1 baseline uses ordinary layer offload. Host-resident work and memory bandwidth can dominate a large MoE model even when the app fits it in RAM. The current functionality proof is not a performance win. Follow the comparison protocol before claiming a regression or improvement, and keep the artifact and prompt fixed.
Runtime startup or backend errors
Section titled “Runtime startup or backend errors”Use the host’s actual error and log stage. The application manages a private instance with its own data directory, credentials and port. Do not kill every process on a familiar port, overwrite another program’s backend, or reuse an unrelated server’s administrative key to make startup appear healthy.
If a generation was canceled, wait for the request to settle and try a small fresh request. If the engine remains unhealthy, unload or close the app deliberately. Load cancellation and shared-service recovery require different ownership rules.
Share a useful report
Section titled “Share a useful report”Export diagnostics from System, review the file for private paths and log contents, then provide the exact error and minimal reproduction. A screenshot of Generating is not enough to prove a hung engine: changing views can hide the answer while the stream completes. Preserve request terminal events and runtime state as evidence.
Do not share API keys, entire private conversation databases or unrelated account files. See security and validation for the relevant boundaries.