Ollama v0.31.2-rc2 unlocks multimodal projector offload on iGPUs

Ollama's latest release candidate, v0.31.2-rc2, addresses a long-standing performance bottleneck for multimodal models running on integrated GPUs (iGPUs). The update allows iGPU offload of the multimodal projector (mmproj) by properly handling fit padding—something previously disabled on non-Metal iGPUs to avoid a llama.cpp fit-pass risk. The fit pass in llama.cpp sizes text-model placement before the multimodal projector is loaded. Without this fix, Ollama completely disabled projector offload on iGPUs (excluding Apple's Metal), which forced the CLIP model onto the CPU—even on systems like NVIDIA's GB10 or AMD's Strix Halo that have ample projector memory.

With the new release, integrated GPUs can now use the same projector-memory check as discrete GPUs. When projector offload is enabled, Ollama adds the estimated projector memory plus the existing 1 GiB headroom to the LLAMA_ARG_FIT_TARGET, ensuring the fit pass reserves space for the eventual projector allocation. The update respects user-set fit targets and only adjusts Ollama-owned ones. The change, which fixes issue #16419, is a meaningful optimization for users who want to run multimodal models (e.g., LLaVA, Phi-3-vision) on devices with strong iGPUs, such as AMD's upcoming Strix Halo APUs or compact AI PCs with the NVIDIA GB10 chip.