Property BackendScratchMemorySize
BackendScratchMemorySize
Gets the device bytes this process's backend instances hold on this device beyond the buffers of weights, KV caches and compute graphs: each backend's temporary pool (matrix-product, attention and conversion temporaries, kept at the largest size ever requested) and its math-library workspaces, summed over every inference context, vision encoder and other graph runner alive on the device. Zero on backends that keep their temporaries in the compute buffers.
public ulong BackendScratchMemorySize { get; }
Property Value
Remarks
Idle scratch returns to the device through ReleaseIdleDeviceMemory(), and the next decode rebuilds what it needs.