Training
Lossless weight coding (rANS / exponent codebook)
Reduce model sizeLossless codingReferenceEntropy-code the weight bytes for a bit-exact ~1.5× shrink — no precision lost, unlike quantization.
On the Spark
The one shrink method with zero quality cost. BF16 sign+exponent fields take only ~31–56 distinct values per tensor, so rANS / a small exponent codebook recovers ~34% with an exact round-trip. On a bandwidth-bound Spark (273 GB/s) a fusible in-register decoder could cut bytes-per-token, not just disk. Research-stage (Brian Bell's weight-compression ledger); no turnkey Spark runtime yet.
- Objective
- Reduce model size · Lossless
- Relieves
- CapacityBandwidth
- Targets
- Weights
- Format
- Entropy-coded BF16 (rANS + exponent codebook)
- Granularity
- Per-tensor codebook, block-granular tiles
- Lifecycle
- PTQ · no gradients
- Calibration
- No calibration
- Compression
- ~1.5× (BF16 → ~10.5 b/w)
- Quality
- 100% — bit-exact round-trip
- Hardware
- In-register (fusible) decode kernels
- Runtimes
In the wild
Lossless weight compression (Brian Bell)
Write-up on entropy-coding model weights for a bit-exact ~30% size reduction — no quality axis at all.
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.