Ollama GPU Offload Stops Working After a Driver Update

A model previously ran partly on GPU but now consumes only CPU resources. This guide narrows the failure to the responsible component rather than applying broad resets. Start with one controlled observation, follow with the smallest reversible correction, and only then use deeper inspection if the error remains.

Recognize this failure and establish a baseline

Compare ollama ps and GPU utilization with the same small model before and after driver changes. Record the observed result with its time and environment. If you have multiple sites, service accounts, installations or regions, label them clearly so a working control is not confused with the failing target.

Model capacity, inference runtime, API contract and data retrieval infrastructure have independent failure modes. Reproduce with a small controlled prompt or request before changing production parameters. A precise comparison is usually more informative than clearing caches, restarting every service or changing unrelated permissions. Confirm the reproduction using a harmless sample and preserve the current configuration before editing it.

Three ways to isolate the cause

First safe fix: change one variable

Recommended correction: Restart the Ollama service after verifying supported drivers and available VRAM. Apply it first in a staging environment or on one limited test object when production traffic, orders or private information could be affected. Write down the previous setting and expected result so a rollback is straightforward.

Why not reset everything? Driver/runtime detection, model size or environment variables may alter acceleration. A broad reset can hide the original cause and create additional failures in permissions, orders, certificates or background jobs. If the proposed correction does not change the exact reproduction test, revert it and move to the advanced inspection instead of stacking speculative fixes.

Advanced diagnosis: inspect the first failing boundary

Inspect service logs, GPU device visibility and container GPU passthrough where applicable. When reading logs, match the same request or transaction across components. A message produced downstream can be secondary to an earlier denial, timeout, incompatible contract or unavailable resource. Use a small known-good request to remove unrelated data from the comparison.

Evidence to keep: Capture the software or protocol version, a masked timestamp, the operation being attempted, the relevant configuration key, and the outcome of the control test. Note both positive and negative findings. If a security rule is involved, preserve the original protection and grant only a narrowly scoped supported exception.

How to verify the repair

Passing result: The same model consistently reports GPU residency or expected partial offload. Verify the exact operation rather than assuming that an adjacent screen, health endpoint or cached response proves success. Repeat once under normal conditions and check one previously healthy workflow for regressions.

If the result changes only in staging, compare release version, permissions, environment variables and network path with production. If success is intermittent, compare two successive runs and retain the corresponding log evidence before declaring the incident closed.

Safety and rollback

Avoid: Do not install random CUDA DLLs or unofficial patched drivers. Store backups, before-and-after settings and sanitized evidence securely. If the action involves orders, identities, databases or customer information, complete an authorized rollback or request maintenance approval before making another production change.

If it still fails

Return to the first comparison: Compare ollama ps and GPU utilization with the same small model before and after driver changes. Then investigate the remaining owner: Inspect service logs, GPU device visibility and container GPU passthrough where applicable. Escalate with a concise timeline and the exact point at which expected behavior differs. Do not send passwords, API keys, recovery codes, complete customer records or private signing material to outside support.

Prevent the next occurrence

Pin model/runtime versions and maintain representative test prompts, resource measurements and dataset provenance for upgrades. Record the minimum regression test for this exact workflow: The same model consistently reports GPU residency or expected partial offload. This is more reliable than applying generic maintenance advice each time a dependency changes.