feat(mm): use same pattern for vae converter as others

Add `dump_path` arg to the converter function & save the model to disk inside the conversion function. This is the same pattern as in the other conversion functions.

invokeai/backend/model_manager/convert_ckpt_to_diffusers.py

@@ -3,10 +3,10 @@
 """Conversion script for the Stable Diffusion checkpoints."""
 
 from pathlib import Path
-from typing import Dict, Optional
+from typing import Optional
 
 import torch
-from diffusers import AutoencoderKL
+from diffusers.models.autoencoders.autoencoder_kl import AutoencoderKL
 from diffusers.pipelines.stable_diffusion.convert_from_ckpt import (
     convert_ldm_vae_checkpoint,
     create_vae_diffusers_config,
@@ -19,9 +19,10 @@ from . import AnyModel
 
 
 def convert_ldm_vae_to_diffusers(
-    checkpoint: Dict[str, torch.Tensor],
+    checkpoint: torch.Tensor | dict[str, torch.Tensor],
     vae_config: DictConfig,
     image_size: int,
+    dump_path: Optional[Path] = None,
     precision: torch.dtype = torch.float16,
 ) -> AutoencoderKL:
     """Convert a checkpoint-style VAE into a Diffusers VAE"""
@@ -30,7 +31,12 @@ def convert_ldm_vae_to_diffusers(
 
     vae = AutoencoderKL(**vae_config)
     vae.load_state_dict(converted_vae_checkpoint)
-    return vae.to(precision)
+    vae.to(precision)
+
+    if dump_path:
+        vae.save_pretrained(dump_path, safe_serialization=True)
+
+    return vae
 
 
 def convert_ckpt_to_diffusers(

invokeai/backend/model_manager/load/model_loaders/vae.py

@@ -64,6 +64,6 @@ class VAELoader(GenericDiffusersLoader):
             vae_config=ckpt_config,
             image_size=512,
             precision=self._torch_dtype,
+            dump_path=output_path,
         )
-        vae_model.save_pretrained(output_path, safe_serialization=True)
         return vae_model