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 slow performance on 10 series Nvidia GPUs.
comfyanonymousa60620dcea13
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex bcc937792..edbe6a8a4 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -668,6 +668,7 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.floif bf16_supported and weight_dtype == torch.bfloat16:return None+ fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)for dt in supported_dtypes:if dt == torch.float16 and fp16_supported:return torch.float16