Add beta7 test2 W4A8 conservative GPTQ (activation-calibrated)
#39
by sandpies - opened
cicalooo's conservative W4A8 layout (W4 g16 on fc1 blocks 10-37, int8 ConvRot elsewhere), GPTQ calibrated on real activations captured from test2 TURBO. No learned rounding. W4 scales match h3quant's GPU output bit-for-bit. Held-out output error vs RTN: W4 layers 0.0618 -> 0.0260 (-58). 19.08 GB.
TenStrip changed pull request status to merged