save current model for SDXL (with --medvram-sdxl option) - Githubissues

wkpark / sd-webui-model-mixer

Checkpoint model mixer/merger extension

GNU Affero General Public License v3.0

103 stars 4 forks source link

save current model for SDXL (with --medvram-sdxl option) #11

Closed wkpark closed 1 year ago

wkpark commented 1 year ago

With this fix, I've verified that the SDXL was saved correctly.

[x] sdxl with --medvram-sdxl works
[x] SD1.5 also works fine.

[x] support load already loaded or merged model from shared.sd_model


@@ -823,8 +823,16 @@ class ModelMixerScript(scripts.Script):
         metadata["sd_merge_models"] = json.dumps(metadata["sd_merge_models"])

         if shared.sd_model is not None:

print("load from shared.sd_model..")
print("Load state_dict from shared.sd_model..")
save to cpu
sd_models.send_model_to_cpu(shared.sd_model)
sd_hijack.model_hijack.undo_hijack(shared.sd_model)

         state_dict = shared.sd_model.state_dict()

restore to gpu
sd_models.send_model_to_device(shared.sd_model)
sd_hijack.model_hijack.hijack(shared.sd_model) else: print("No loaded model found") return gr.update(value="No loaded model found")