Weights-only VRAM
Weights-only VRAM
Current VRAM labeling:
| Component | Status |
|---|---|
| Artifact weights ≈ file size | SOURCE |
| KV cache | NOT KNOWN YET |
| Runtime buffers | NOT KNOWN YET |
| Safety reserve | NOT KNOWN YET |
| Weights-only memory class | SOURCE + POLICY |
Policy thresholds live in data/source/rules/weights-only-vram.yaml. The band is the smallest listed class (8 / 12 / 16 / 24 / 32 / 48 / 64 GB) whose capacity is at least the artifact size.
This is not:
- a full fit recommendation
- a max-context result
- a guarantee the model runs at a useful context on that GPU