Use GPU for loading safetensors, disable export

modules/sd_models.py

@@ -147,8 +147,9 @@ def torch_load(model_filename, model_info, map_override=None):
     map_override=shared.weight_load_location if not map_override else map_override
     if(checkpoint_types[model_info.exttype] == 'safetensors'):
         # safely load weights
-        # TODO: safetensors supports zero copy fast load to gpu, see issue #684
-        return load_file(model_filename, device=map_override)
+        # TODO: safetensors supports zero copy fast load to gpu, see issue #684.  
+        # GPU only for now, see https://github.com/huggingface/safetensors/issues/95
+        return load_file(model_filename, device='cuda')
     else:
         return torch.load(model_filename, map_location=map_override)
 

modules/ui.py

@@ -1187,7 +1187,8 @@ def create_ui(wrap_gradio_gpu_call):
                 interp_amount = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Multiplier (M) - set to 0 to get model A', value=0.3)
                 interp_method = gr.Radio(choices=["Weighted sum", "Add difference"], value="Weighted sum", label="Interpolation Method")
                 save_as_half = gr.Checkbox(value=False, label="Save as float16")
-                save_as_safetensors = gr.Checkbox(value=False, label="Save as safetensors format")
+                # invisible until feature can be verified
+                save_as_safetensors = gr.Checkbox(value=False, label="Save as safetensors format", visible=False)
                 modelmerger_merge = gr.Button(elem_id="modelmerger_merge", label="Merge", variant='primary')
 
             with gr.Column(variant='panel'):