Add native int8 HQ quantization of Beta7 Fix1(INT8 ConvRot + protected edges)
#86
by LokkenJP - opened
This PR adds a native quantized HQ checkpoint of 10Eros_Max_h3_8stepTURBO_beta7_fix1.
It uses native INT8 ConvRot for the inner 180 main transformer matrices and native precision on protected edge blocks (0,1,47-49), preserving all other auxiliary tensors in their original source precision.
This is the fixed native int8 recipe, a quality-focused HQ activation learned w6a8+int8 version will follow for lower memory requirements and lower-end GPUs in my own repo.
Actual VRAM usage depends on the workflow and other loaded models.
A full model card and standalone release will follow.
TenStrip changed pull request status to merged