Add highly optimized hybrid w4a8/int8 learned version for 16Gb vRAM
#10
by LokkenJP - opened
Adds a quality-focused mixed-precision quantization of ErosMax H3 beta5, targeting 16 GB VRAM workflows with a checkpoint just under 14 GB.
- Native W4A8 with per-tensor group sizes selected from fresh calibration data.
- Learned INT8 ConvRot for the 20 early/late-block matrices, with all non-target tensors preserved in BF16.
- Validated checkpoint integrity and successful ComfyUI generation.
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.
TenStrip changed pull request status to merged