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.

Don't run TE on cpu when dynamic vram enabled. (#12815)

comfyanonymousd69d30819b91
diff --git a/comfy/model_management.py b/comfy/model_management.py
index ee28ea107..39b4aa483 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -939,7 +939,7 @@ def text_encoder_offload_device():
def text_encoder_device():
if args.gpu_only:
return get_torch_device()
- elif vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.NORMAL_VRAM:
+ elif vram_state in (VRAMState.HIGH_VRAM, VRAMState.NORMAL_VRAM) or comfy.memory_management.aimdo_enabled:
if should_use_fp16(prioritize_performance=False):
return get_torch_device()
else: