TurboQuant v1.5.2 — PPL 21%→6%, Precision Fix, Deterministic Kernel
v1.5.2 — PPL + Math Bench Dual Improvement
3-bit KV cache now achieves f16-equivalent quality on both PPL and math benchmarks.
PPL (wikitext-2, ctx=2048, Gemma 4 26B MoE)
| Config | PPL | vs f16 |
|---|---|---|
| f16/f16 | 419.8 | 1.00x |
| tbqp3/tbq3 | 454.7 | 1.08x |
v1.5.1 was 1.21x → now 1.08x (21% gap → 8% gap).
Math Accuracy (35 problems × 10 runs, 262K ctx, temp=0)
| Config | Average | Peak |
|---|---|---|
| tbqp3/tbq3 | 19.1/35 | 23/35 |
| f16/f16 | 20.1/35 | 21/35 |
Key Changes
- Attention Sharpening (α = 1 + 1/(2×SQNR)): Compensates softmax flattening from quantization noise. TBQP3 α=1.036, TBQ3 α=1.016. Derived from MMSE theory.
- V Rotation Bugfix:
attn_rot_vwas enabled but IWHT decode has no inverse rotation — V output was corrupted. - Per-block Norm for TBQ3 D=512 encode.
- 1.15x V compensation removed (replaced by principled sharpening).
- tbq4_0 D=512 OOB read fix.