Install sub directories with folders correctly, ensure consistent dtype of tensors in flux pipeline and vae

invokeai/app/invocations/flux_text_to_image.py

@@ -72,7 +72,7 @@ class FluxTextToImageInvocation(BaseInvocation, WithMetadata, WithBoard):
         t5_embeddings: torch.Tensor,
     ):
         transformer_info = context.models.load(self.transformer.transformer)
-        inference_dtype = TorchDevice.choose_torch_dtype()
+        inference_dtype = torch.bfloat16
 
         # Prepare input noise.
         # TODO(ryand): Does the seed behave the same on different devices? Should we re-implement this to always use a

invokeai/app/services/model_install/model_install_default.py

@@ -783,8 +783,9 @@ class ModelInstallService(ModelInstallServiceBase):
         # So what we do is to synthesize a folder named "sdxl-turbo_vae" here.
         if subfolder:
             top = Path(remote_files[0].path.parts[0])  # e.g. "sdxl-turbo/"
-            path_to_remove = top / subfolder.parts[-1]  # sdxl-turbo/vae/
-            path_to_add = Path(f"{top}_{subfolder}")
+            path_to_remove = top / subfolder  # sdxl-turbo/vae/
+            subfolder_rename = subfolder.name.replace('/', '_').replace('\\', '_')
+            path_to_add = Path(f"{top}_{subfolder_rename}")
         else:
             path_to_remove = Path(".")
             path_to_add = Path(".")

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

@@ -64,8 +64,7 @@ class FluxVAELoader(GenericDiffusersLoader):
             params = AutoEncoderParams(**filtered_data)
 
             with SilenceWarnings():
-                model = load_class(params).to(self._torch_dtype)
-                # load_sft doesn't support torch.device
+                model = load_class(params)
                 sd = load_file(model_path)
                 model.load_state_dict(sd, strict=False, assign=True)
 
@@ -203,8 +202,6 @@ class FluxBnbQuantizednf4bCheckpointModel(GenericDiffusersLoader):
             with accelerate.init_empty_weights():
                 model = load_class(params)
                 model = quantize_model_nf4(model, modules_to_not_convert=set(), compute_dtype=torch.bfloat16)
-            # TODO(ryand): Right now, some of the weights are loaded in bfloat16. Think about how best to handle
-            # this on GPUs without bfloat16 support.
             sd = load_file(model_path)
             model.load_state_dict(sd, strict=False, assign=True)
         return model

invokeai/backend/model_manager/util/select_hf_files.py

@@ -69,7 +69,7 @@ def filter_files(
     # limit search to subfolder if requested
     if subfolder:
         subfolder = root / subfolder
-        paths = [x for x in paths if x.parent == Path(subfolder)]
+        paths = [x for x in paths if Path(subfolder) in x.parents]
 
     # _filter_by_variant uniquifies the paths and returns a set
     return sorted(_filter_by_variant(paths, variant))

invokeai/backend/quantization/bnb_nf4.py

@@ -116,7 +116,7 @@ def _convert_linear_layers_to_nf4(
                 child.in_features,
                 child.out_features,
                 bias=has_bias,
-                compute_dtype=torch.float16,
+                compute_dtype=compute_dtype,
                 compress_statistics=compress_statistics,
             )
             if has_bias: