Ollama fork rebalances mixed GPUs

This is the kind of infra work that actually moves tokens/sec on real hardware: mixed-GPU setups fail when the scheduler optimizes for memory only and ignores where the compute is. The post reads like a practical fork-level fix, not a theoretical rewrite.

• –Compute-weighted capacity is the key idea here: it shifts more layers onto the faster GPU without pretending VRAM is the only constraint
• –Reversing greedy placement matters because it stops the slow card from monopolizing the heaviest layers
• –Forcing the output layer onto the strongest GPU is a sensible latency win, since that layer is disproportionately expensive
• –The redistribution pass for FFN-heavy layers is a targeted heuristic, which is exactly what heterogeneous rigs usually need
• –This looks useful for power users, but upstreaming will hinge on correctness across many GPU combinations and keeping the heuristics understandable