The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
bf16 manual cast works on old AMD.
comfyanonymous8c0bae50c3b1
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 05f66c9e5..535d53014 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -1111,6 +1111,8 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, maif is_amd():arch = torch.cuda.get_device_properties(device).gcnArchNameif arch in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]: # RDNA2 and older don't support bf16+ if manual_cast:+ return Truereturn Falseprops = torch.cuda.get_device_properties(device)