Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix int8 mm being skipped on offloaded lora weights. (#14669)

comfyanonymous79c555ce6bfe
diff --git a/comfy/ops.py b/comfy/ops.py
index 6a5090548..69d32e254 100644
--- a/comfy/ops.py
+++ b/comfy/ops.py
@@ -1216,7 +1216,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
bias_dtype=input.dtype,
offloadable=True,
compute_dtype=compute_dtype,
- want_requant=want_requant,
+ want_requant=True,
)
weight = weight.to(dtype=input.dtype)
else: