llama.cpp VRAM Spikes with Long Contexts on Intel Arc B580

WizardlyBump17 · reddit · 2026-08-14

A developer running llama.cpp locally via SYCL on an Intel Arc B580 GPU encountered a VRAM bottleneck. While the model consumes 10.7GB of VRAM upon loading, sending a prompt immediately spikes usage to 11.6GB. As the conversation context grows, the VRAM usage eventually fills up the entire 12GB and can even cause a crash.

The author noted that switching to the Vulkan backend keeps VRAM usage flat, but at the cost of significantly slower inference speeds. They shared their Docker configuration running a Qwen model and asked the community if this VRAM scaling issue is a common trait across Nvidia and AMD cards, seeking advice on optimizing long-context handling.

Original post →

More from Infra

Infra channel →