Property ScratchMemorySize
ScratchMemorySize
Gets the device bytes the context's backend holds beyond MemorySize: its temporary pool (matrix-product, attention and conversion temporaries, kept at the largest size any decode on this context requested) and its math-library workspaces, speculative-decoding sibling included. Zero on backends that keep their temporaries in the compute buffers already counted by MemorySize, and while hibernated. Returned to the device, and rebuilt by the next decode, by ReleaseIdleDeviceMemory().
public long ScratchMemorySize { get; }