Add beta7 test3 int8 convrot GPTQ quant
#42
by cicalooo - opened
int8_convrot (W8A8, per-row scales) GPTQ quant of 10Eros_Max_h3_8stepTURBO_beta7_test3. Blocks 0,1,48,49 + in/out layers bf16. GPTQ damp 0.01, blocksize 128, act-order, synthetic calibration (12 samples, 8-step shift-12 schedule). Held-out mean layer output rel error 0.00919 vs RTN 0.01149.
TenStrip changed pull request status to merged