Archive notes
The number beside VRAM needs a footnote.
Weight size, context length and runtime overhead belong in the same deployment record.
OpenWeightArchive ·
The archive’s memory matrix separates the size of the weights from the memory needed to run them. Each row names a quantization; each context column records a distinct condition.
State the basis
Every memory value is marked as estimated or measured. A measured value should eventually link to the runtime version, cache format, batch size, context length, offload configuration and peak-memory observation that produced it.
An estimate helps plan a trial. It is not a promise that a particular device can run the model at the stated context length.
Leave gaps visible
The reference records contain deliberate gaps. An unavailable value stays visible as unavailable instead of being inferred from another quantization or replaced with a zero.
The current Spark record reports a 2.68 GB KV cache per rank. It does not establish total VRAM requirements at other context lengths. Missing memory measurements remain unavailable; no extrapolated memory matrix is published.