Turn troubleshooting advice into evidence-led cases.
You will diagnose service, download, memory, GPU, speed, and output issues by changing one factor at a time.
Before you begin
Save work and collect runtime version, model identifier, OS, hardware summary, free storage, and the complete error.
Start with exact evidence
Record the command, complete error, runtime version, model identifier, operating system, hardware, free storage, and available memory. Reproduce the issue with the smallest prompt. Change one factor at a time so you know which action mattered.
Model will not load
Confirm the download completed and the model is supported by the runtime. Check free disk space and RAM or VRAM. Reduce context length or choose a smaller quantization. Close other heavy applications. Do not download random replacement files from untrusted mirrors.
GPU is not used
Check the runtime’s current hardware-support documentation, installed driver, and logs. A GPU may be unsupported on that operating system or build. Update through the hardware vendor’s official channel and retest a known supported model. CPU fallback is not necessarily a defect.
Slow or poor output
Separate performance from quality. For speed, measure model load time and token generation while monitoring CPU, GPU, and memory. For quality, use a repeatable prompt, confirm the correct instruct model and template, reduce irrelevant context, and test a model suited to the language and task. Keep a rollback path before runtime or driver changes.
Service and connection errors
Determine whether the runtime process is running using official status or log guidance. Confirm the client addresses the expected local host and port. Restart normally and repeat the smallest request. If another process uses the port, identify it before stopping anything. Security software may affect connections, but do not disable protection broadly; use only a documented local exception when necessary.
Download and storage failures
Confirm model name and source, disk space, storage location, and connection. Interrupted downloads may leave temporary data that the runtime can resume or clean; follow documentation instead of deleting unknown folders. Check permissions and file locks. For integrity errors, download from the official source again. Keep space for extraction and updates, not only the advertised model size.
Context and output quality
Poor instruction following can result from a base model, irrelevant context, truncation, or ambiguity. Confirm the instruct variant and model identity. Reduce to one reproducible example, state the format, and compare with a known answer. Lowering context can relieve memory pressure but remove evidence. Do not confuse confident prose with system health; evaluate quality against the task.
Change control and escalation
Record each change and result. Update one layer at a time—application, runtime, driver, or model—and retain the last working version. Avoid driver or firmware changes unless official documentation requires them. When asking for help, share sanitized errors, versions, model, reproduction steps, and resource observations, never passwords, tokens, private prompts, or customer files.
Troubleshoot one software layer at a time
Decision: Identify interface, API, runtime, model, driver, memory, or network as the failing layer before changing it.
Relevant options: Ollama exposes CLI/API checks; LM Studio documents desktop/server use; llama.cpp provides lower-level runtime evidence.
Worked example: When desktop chat fails, test the configured endpoint and model list, then one known-small model.
Failure mode: Changing drivers, runtimes, and models together destroys the last known state and can add defects.
Validation: Record error, version, endpoint, model, storage, memory, backend, and one controlled change.
Useful next step: Restore the last working layer and escalate with a minimal reproduction.
Official sources checked 2026-10-24: Ollama · LM Studio · llama.cpp. Product availability, support, terms, pricing, and licenses can change. These are vendor or project sources; no independent benchmark is claimed.
Three diagnostic cases
- Observed: user Ollama unit inactive, system unit active, API responsive. Identify service scope first.
- Synthetic: “model not found” after a misspelling. Check exact identifier before reinstalling.
- Synthetic: output reaches cap. Preserve stop reason; do not call it complete.
Only case one was observed on 2026-10-27.
Authoritative references
Reference scope checked 2026-10-27. These sources support the concepts and risk controls; product-specific behavior still requires the nearby vendor documentation and a dated test.
Your next step
Capture the exact error and identify the failing layer before changing one variable.
Editorial provenance
Nyfir Studios is the publisher and editorially responsible organization. AI tools assisted with drafting and structured review for this candidate; factual claims, test labels, links, and retained evidence require human review before deployment. No named employee, credential, or unrecorded hands-on result is implied.