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.

Always use fp16 for the text encoders.

comfyanonymous4b0239066daa
diff --git a/comfy/model_management.py b/comfy/model_management.py
index e12146d11..cbaa80874 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -546,10 +546,8 @@ def text_encoder_dtype(device=None):
if is_device_cpu(device):
return torch.float16
- if should_use_fp16(device, prioritize_performance=False):
- return torch.float16
- else:
- return torch.float32
+ return torch.float16
+
def intermediate_device():
if args.gpu_only: