Ollama 0.32.10 regression causes model load timeouts on Linux iGPUs

A regression in Ollama version 0.32.10 and later is causing model loading to time out, particularly on Linux systems using containers (Podman or Docker) with integrated GPUs on the Vulkan backend. This issue stems from a change in the default load mode from 'mmap' to 'none', which requires the entire model to be staged in CPU memory before being handed to the GPU. This can lead to excessive memory usage and disk thrashing on systems with limited RAM, causing loads to exceed the default 5-minute timeout. Potential workarounds include pinning Ollama to version 0.32.9, capping the context window, increasing the load timeout, or monitoring a community pull request for a more permanent fix. AI

IMPACT This regression impacts users running local LLMs on specific hardware configurations, potentially hindering adoption until a fix is widely available.

RANK_REASON The item details a regression and workaround for a specific software version, not a new release or major industry event.

Read on dev.to — LLM tag →

• 0.32.10
• 0.32.9
• 0.33.x
• Docker
• Linux
• llama.cpp
• Ollama
• Podman
• Radeon 780M
• Vulkan

• infra
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →