fix dtype logic

TODO.md

@@ -6,12 +6,16 @@ Stuff to be fixed...
 
 - Fix extensions imports
 - ClipSkip not updated on read gen info
+- Usage of `sd_vae` in quick settings
+- Run VAE with hires at 1280
+
 
 ## Features
 
 Stuff to be added...
 
 - Add Gradio base themes: <https://gradio.app/theming-guide/#using-the-theme-builder>
+- Add gradio theme maker
 - Create new GitHub hooks/actions for CI/CD  
 - Move Restart Server from WebUI to Launch and reload modules
 - Redo Extensions tab: see <https://vladmandic.github.io/sd-extension-manager/pages/extensions.html>

modules/devices.py

@@ -54,17 +54,17 @@ def torch_gc():
 
 
 def set_cuda_params():
-    if not torch.cuda.is_available():
-        return
     from modules import shared
-    if torch.backends.cudnn.is_available():
-        torch.backends.cudnn.benchmark = shared.opts.cudnn_benchmark
-        torch.backends.cudnn.benchmark_limit = 0
-        torch.backends.cudnn.allow_tf32 = shared.opts.cuda_allow_tf32
-    torch.backends.cuda.matmul.allow_tf32 = shared.opts.cuda_allow_tf32
-    torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = shared.opts.cuda_allow_tf16_reduced
-    torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = shared.opts.cuda_allow_tf16_reduced
+    if torch.cuda.is_available():
+        torch.backends.cuda.matmul.allow_tf32 = shared.opts.cuda_allow_tf32
+        torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = shared.opts.cuda_allow_tf16_reduced
+        torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = shared.opts.cuda_allow_tf16_reduced
+        if torch.backends.cudnn.is_available():
+            torch.backends.cudnn.benchmark = shared.opts.cudnn_benchmark
+            torch.backends.cudnn.benchmark_limit = 0
+            torch.backends.cudnn.allow_tf32 = shared.opts.cuda_allow_tf32
     global dtype, dtype_vae, dtype_unet, unet_needs_upcast # pylint: disable=global-statement
+    # set dtype
     if shared.opts.cuda_dtype == 'FP16':
         dtype = torch.float16
         dtype_vae = torch.float16
@@ -73,10 +73,12 @@ def set_cuda_params():
         dtype = torch.bfloat16
         dtype_vae = torch.bfloat16
         dtype_unet = torch.bfloat16
-    if shared.opts.cuda_dtype == 'FP32':
+    if shared.opts.cuda_dtype == 'FP32' or shared.opts.no_half:
         dtype = torch.float32
         dtype_vae = torch.float32
         dtype_unet = torch.float32
+    if shared.opts.no_half_vae: # set dtype again as no-half-vae options take priority
+        dtype_vae = torch.float32
     unet_needs_upcast = shared.opts.upcast_sampling
 
 

modules/sd_models.py

@@ -403,7 +403,6 @@ def load_model(checkpoint_info=None, already_loaded_state_dict=None):
         current_checkpoint_info = shared.sd_model.sd_checkpoint_info
         sd_hijack.model_hijack.undo_hijack(shared.sd_model)
         shared.sd_model = None
-
     gc.collect()
     devices.torch_gc()
 
@@ -526,9 +525,7 @@ def reload_model_weights(sd_model=None, info=None):
 def unload_model_weights(sd_model=None, _info=None):
     from modules import sd_hijack
     timer = Timer()
-
     if shared.sd_model:
-
         # shared.sd_model.cond_stage_model.to(devices.cpu)
         # shared.sd_model.first_stage_model.to(devices.cpu)
         shared.sd_model.to(devices.cpu)
@@ -538,9 +535,7 @@ def unload_model_weights(sd_model=None, _info=None):
         gc.collect()
         devices.torch_gc()
         torch.cuda.empty_cache()
-
     print(f"Unloaded weights {timer.summary()}")
-
     return sd_model
 
 

modules/textual_inversion/textual_inversion.py

@@ -130,6 +130,9 @@ class EmbeddingDatabase:
         return embedding
 
     def get_expected_shape(self):
+        if shared.sd_model is None:
+            print('Model not loaded')
+            return 0
         vec = shared.sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
         return vec.shape[1]