RomanNum3ral
/
ComfyUI
mirror of https://github.com/Comfy-Org/ComfyUI.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Preserve full_precision_matrix_mult in the NVFP4 metadata rewrite 

The conversion now copies the int8 layer's full_precision_matrix_mult flag into the rewritten .comfy_quant metadata instead of replacing it with a bare nvfp4 entry, so layers that require the full-precision matmul path keep it.
This commit is contained in:
F1LIP3 2026-08-15 14:39:15 -03:00
parent ea840373b8
commit a2df80be7e
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
comfy/text_encoders/minimax_music.py
View File

@ -36,7 +36,10 @@ def _to_nvfp4(sd):
sd[prefix + ".weight"] = qdata sd[prefix + ".weight"] = qdata
sd[prefix + ".weight_scale"] = params.block_scale sd[prefix + ".weight_scale"] = params.block_scale
sd[prefix + ".weight_scale_2"] = params.scale sd[prefix + ".weight_scale_2"] = params.scale
sd[key] = torch.tensor(list(b'{"format": "nvfp4"}'), dtype=torch.uint8) meta = {"format": "nvfp4"}
if "full_precision_matrix_mult" in conf:
meta["full_precision_matrix_mult"] = conf["full_precision_matrix_mult"]
sd[key] = torch.tensor(list(json.dumps(meta).encode()), dtype=torch.uint8)
MODEL_CONFIG = { MODEL_CONFIG = {