KV Caching Across Accelerators on Edge SoCs
What a KV cache saves, when another engine can use it, and when rebuilding the context is the better option.
Read the noteTechnical writing / From the workbench
Things learned by building, measuring, and getting it wrong.
Technical posts, research notes, and practical guides.
What a KV cache saves, when another engine can use it, and when rebuilding the context is the better option.
Read the noteNPUs like stable graphs, GPUs need enough parallel work, and shared memory is still shared.
Read the note