decode-tokens

A discussion questioning whether extremely high decode speeds (1k-10k tok/s) for large models like Qwen 3.5 397B or GLM-5.2 would unlock new use cases or be better spent on loading larger models.