diff --git a/packages_3rdparty/gguf/quants.py b/packages_3rdparty/gguf/quants.py index c0d144d5..4e013141 100644 --- a/packages_3rdparty/gguf/quants.py +++ b/packages_3rdparty/gguf/quants.py @@ -620,6 +620,8 @@ class Q8_0(__Quant, qtype=GGMLQuantizationType.Q8_0): if d.device != x.device: d = d.to(device=x.device) + x = x.to(cls.computation_dtype) + return x * d @classmethod