ollama - 💡(How to fix) Fix [Regression 0.20.x] memory layout cannot be allocated for all models >10 GB on Windows — GPU abandoned after vision encoder CPU buffer allocation failure [10 comments, 2 participants]

Environment

• OS: Windows 10 Home 22H2 (build 19045)
• GPU: NVIDIA RTX 3090 24 GB (CUDA 12.x)
• Ollama version: 0.20.2 (auto-updated ~April 5 2026)
• Previously working version: 0.19.x

Affected models (all fail)

Working model

• gemma4:e2b (7.2 GB) — loads and runs correctly

Observed behaviour

When loading any model larger than ~10 GB, Ollama logs show:

1. Vision encoder tensors load to GPU successfully
2. A subsequent CPU buffer allocation (~20 MB) fails with
3. Ollama then falls back to CPU-only mode ()
4. CPU RAM is insufficient for the full model → generation fails or produces garbage

The model never recovers GPU access once the CPU fallback is triggered. Restarting Ollama does not help.

Workarounds attempted

• Restarting the Ollama service
• Running Ollama in an interactive desktop session (not Session 0) to ensure GPU access
• Reducing to 8192
• Pulling the model fresh

None resolved the issue. The 7.2 GB model works, confirming the GPU is functional.

Regression

This worked correctly on Ollama 0.19.x. The issue appeared immediately after an automatic update to 0.20.2.

Additional info

The failure appears to be in memory layout/planning for large models — specifically a CPU staging buffer allocation that regressed in 0.20.x. Once that allocation fails, the GPU is abandoned entirely for the session.