← PR #1917
PR #1919→

PR #1918

open

RECORD: 1855 base + AWQ-lite mixed-precision GPTQ — val_bpb 1.06086 (3-seed mean)

by aquariouseworkmanView on GitHub
val_bpb
1.0609
Architecture
Transformer
Optimizer
—
Artifact Size
15,978,503 bytes

Training Techniques

Quantization
GPTQ
bits: 6
scope: all weights with top-1 salient 64-column group per matrix at int8
mixed int6/int8
bits: null
scope: top-1 salient 64-column group per matrix
Test-Time Training
full TTT
parameters: null

Novel Contributions

  • Applies activation-aware mixed-precision GPTQ on top of the PR #1855 base stack
  • Uses AWQ-lite saliency scoring from activation RMS and mean absolute weight magnitude
  • Quantizes the most salient 64-column group per matrix at int8 while keeping the rest at int6
  • Reports a 3-seed mean validation bpb of 1.06086