Add highly optimized HQ hybrid w6a8/int8 activation learned version of beta7 test1 for 16Gb vRAM
This PR adds a learned HQ quantization of 10Eros_Max_h3_8stepTURBO_beta7_test1. It uses W6A8 group-16 for 90 QKV/FC1 matrices and learned INT8 ConvRot for the remaining 110 main transformer matrices, preserving all other tensors in their original source precision.
Both formats were optimized using fresh Beta7-specific activation captures from a companion quantized from this same checkpoint. This is the fixed, quality-focused HQ baseline for testing; a smaller adaptive W4 compact version will follow, reusing compatible learned results.
Actual VRAM usage depends on the workflow and other loaded models. A full model card and standalone release will follow once repository gating is removed.
This where I'd put all my beta7 models...if I HAD ANY
I don’t know why, but I submitted my test to Affectionate, but it’s been a long time and it still hasn’t passed the review.