diff --git a/javascript/gallery.js b/javascript/gallery.js
index d9945cd8d..1f3afd148 100644
--- a/javascript/gallery.js
+++ b/javascript/gallery.js
@@ -276,33 +276,34 @@ async function gallerySort(btn) {
   const arr = Array.from(el.files.children).filter((node) => node.name); // filter out separators
   const fragment = document.createDocumentFragment();
   el.files.innerHTML = '';
+  log('gallerySort', btn.charCodeAt(0));
   switch (btn.charCodeAt(0)) {
-    case 61789:
+    case 61789: // name asc
       arr
         .sort((a, b) => a.name.localeCompare(b.name))
         .forEach((node) => fragment.appendChild(node));
       break;
-    case 61790:
+    case 61790: // name dsc
       arr
         .sort((b, a) => a.name.localeCompare(b.name))
         .forEach((node) => fragment.appendChild(node));
       break;
-    case 61792:
+    case 61792: // size asc
       arr
         .sort((a, b) => a.size - b.size)
         .forEach((node) => fragment.appendChild(node));
       break;
-    case 61793:
+    case 61793: // size dsc
       arr
         .sort((b, a) => a.size - b.size)
         .forEach((node) => fragment.appendChild(node));
       break;
-    case 61794:
+    case 61794: // resolution asc
       arr
         .sort((a, b) => a.width * a.height - b.width * b.height)
         .forEach((node) => fragment.appendChild(node));
       break;
-    case 61795:
+    case 61795: // resolution dsc
       arr
         .sort((b, a) => a.width * a.height - b.width * b.height)
         .forEach((node) => fragment.appendChild(node));
diff --git a/modules/processing_args.py b/modules/processing_args.py
index 7cb380ee7..1aebca869 100644
--- a/modules/processing_args.py
+++ b/modules/processing_args.py
@@ -99,6 +99,7 @@ def set_pipeline_args(p, model, prompts: list, negative_prompts: list, prompts_2
     steps = kwargs.get("num_inference_steps", None) or len(getattr(p, 'timesteps', ['1']))
     clip_skip = kwargs.pop("clip_skip", 1)
 
+    prompt_parser_diffusers.fix_position_ids(model)
     if shared.opts.prompt_attention != 'Fixed attention' and 'Onnx' not in model.__class__.__name__ and (
         'StableDiffusion' in model.__class__.__name__ or
         'StableCascade' in model.__class__.__name__ or
diff --git a/modules/processing_vae.py b/modules/processing_vae.py
index 77f24711c..4ad77edae 100644
--- a/modules/processing_vae.py
+++ b/modules/processing_vae.py
@@ -60,8 +60,7 @@ def full_vae_decode(latents, model):
             model.vae.orig_dtype = model.vae.dtype
             model.vae = model.vae.to(dtype=torch.float32)
         latents = latents.to(torch.float32)
-    else:
-        latents = latents.to(devices.device)
+    latents = latents.to(devices.device)
     if getattr(model.vae, "post_quant_conv", None) is not None:
         latents = latents.to(next(iter(model.vae.post_quant_conv.parameters())).dtype)
 
@@ -84,6 +83,7 @@ def full_vae_decode(latents, model):
     latents_stats = f'shape={latents.shape} dtype={latents.dtype} device={latents.device}'
     stats = f'vae {vae_stats} latents {latents_stats}'
 
+    log_debug(f'VAE config: {model.vae.config}')
     try:
         decoded = model.vae.decode(latents, return_dict=False)[0]
     except Exception as e:
diff --git a/modules/prompt_parser_diffusers.py b/modules/prompt_parser_diffusers.py
index 6ebba5e0a..75e9f5b32 100644
--- a/modules/prompt_parser_diffusers.py
+++ b/modules/prompt_parser_diffusers.py
@@ -34,10 +34,12 @@ def compel_hijack(self, token_ids: torch.Tensor,
                   attention_mask: typing.Optional[torch.Tensor] = None) -> torch.Tensor:
     needs_hidden_states = self.returned_embeddings_type != 1
     try: # can crash in ATen/native/cuda/Indexing since position_ids are corrupt so index lookup fails, but its not compel specific, happens with fixed attention as well
+        sd_models.move_model(self.text_encoder, devices.device)
         text_encoder_output = self.text_encoder(token_ids, attention_mask, output_hidden_states=needs_hidden_states, return_dict=True)
     except Exception as e: # its a non-recoverable error as cuda state is corrupt
         shared.log.error(f'TE: class={self.text_encoder.__class__} device={self.text_encoder.device} dtype={self.text_encoder.dtype} {e}')
         errors.display(e, 'TE:')
+        return None
 
     if not needs_hidden_states:
         return text_encoder_output.last_hidden_state
@@ -174,7 +176,7 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
     ):
         shared.log.warning(f"Prompt parser not supported: {pipe.__class__.__name__}")
         return
-    elif shared.opts.sd_textencoder_cache and cache.get('model_type', None) == shared.sd_model_type and params_match and False:
+    elif shared.opts.sd_textencoder_cache and cache.get('model_type', None) == shared.sd_model_type and params_match:
         p.prompt_embeds = cache.get('prompt_embeds', None)
         p.positive_pooleds = cache.get('positive_pooleds', None)
         p.negative_embeds = cache.get('negative_embeds', None)
@@ -190,10 +192,14 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
             pipe.maybe_free_model_hooks()
             devices.torch_gc()
 
-        fix_position_ids(pipe)
-        prompt_embeds, positive_pooleds, negative_embeds, negative_pooleds = [], [], [], []
+        p.prompt_embeds = []
+        p.positive_pooleds = []
+        p.negative_embeds = []
+        p.negative_pooleds = []
+        p.scheduled_prompt = False
         last_prompt, last_negative = None, None
         for prompt, negative in zip(prompts, negative_prompts):
+            prompt_embeds, positive_pooleds, negative_embeds, negative_pooleds = [], [], [], []
             prompt_embed, positive_pooled, negative_embed, negative_pooled = None, None, None, None
             if last_prompt == prompt and last_negative == negative:
                 prompt_embeds.append(prompt_embeds[-1])
@@ -205,11 +211,7 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
                 continue
             positive_schedule, scheduled = get_prompt_schedule(prompt, steps)
             negative_schedule, neg_scheduled = get_prompt_schedule(negative, steps)
-            p.scheduled_prompt = scheduled or neg_scheduled
-            p.prompt_embeds = []
-            p.positive_pooleds = []
-            p.negative_embeds = []
-            p.negative_pooleds = []
+            p.scheduled_prompt = p.scheduled_prompt or scheduled or neg_scheduled
 
             for i in range(max(len(positive_schedule), len(negative_schedule))):
                 positive_prompt = positive_schedule[i % len(positive_schedule)]
@@ -228,25 +230,25 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
                     negative_pooleds.append(negative_pooled)
             last_prompt, last_negative = prompt, negative
 
-        def fix_length(embeds):
-            max_len = max([e.shape[1] for e in embeds if e is not None])
-            for i, e in enumerate(embeds):
-                if e is not None and e.shape[1] < max_len:
-                    expanded = torch.zeros((e.shape[0], max_len, e.shape[2]), device=e.device, dtype=e.dtype)
-                    expanded[:, :e.shape[1], :] = e
-                    embeds[i] = expanded
-            return torch.cat(embeds, dim=0).to(devices.device, dtype=devices.dtype)
+            def fix_length(embeds):
+                max_len = max([e.shape[1] for e in embeds if e is not None])
+                for i, e in enumerate(embeds):
+                    if e is not None and e.shape[1] < max_len:
+                        expanded = torch.zeros((e.shape[0], max_len, e.shape[2]), device=e.device, dtype=e.dtype)
+                        expanded[:, :e.shape[1], :] = e
+                        embeds[i] = expanded
+                return torch.cat(embeds, dim=0).to(devices.device, dtype=devices.dtype)
 
-        if len(prompt_embeds) > 0:
-            p.prompt_embeds.append(fix_length(prompt_embeds))
-        if len(negative_embeds) > 0:
-            p.negative_embeds.append(fix_length(negative_embeds))
-        if len(positive_pooleds) > 0:
-            p.positive_pooleds.append(fix_length(positive_pooleds))
-        if len(negative_pooleds) > 0:
-            p.negative_pooleds.append(fix_length(negative_pooleds))
+            if len(prompt_embeds) > 0:
+                p.prompt_embeds.append(fix_length(prompt_embeds))
+            if len(negative_embeds) > 0:
+                p.negative_embeds.append(fix_length(negative_embeds))
+            if len(positive_pooleds) > 0:
+                p.positive_pooleds.append(fix_length(positive_pooleds))
+            if len(negative_pooleds) > 0:
+                p.negative_pooleds.append(fix_length(negative_pooleds))
 
-        if shared.opts.sd_textencoder_cache and p.batch_size == 1:
+        if p.batch_size == 1:
             cache.update({
                 'prompt_embeds': p.prompt_embeds,
                 'negative_embeds': p.negative_embeds,
diff --git a/modules/sd_models.py b/modules/sd_models.py
index 613a33bfa..577a74ff5 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -1035,8 +1035,6 @@ def load_diffuser(checkpoint_info=None, already_loaded_state_dict=None, timer=No
     if shared.opts.diffusers_pipeline == 'Custom Diffusers Pipeline' and len(shared.opts.custom_diffusers_pipeline) > 0:
         shared.log.debug(f'Model pipeline: pipeline="{shared.opts.custom_diffusers_pipeline}"')
         diffusers_load_config['custom_pipeline'] = shared.opts.custom_diffusers_pipeline
-    # if 'LCM' in checkpoint_info.path:
-        #    diffusers_load_config['custom_pipeline'] = 'latent_consistency_txt2img'
     if shared.opts.data.get('sd_model_checkpoint', '') == 'model.ckpt' or shared.opts.data.get('sd_model_checkpoint', '') == '':
         shared.opts.data['sd_model_checkpoint'] = "stabilityai/stable-diffusion-xl-base-1.0"
 
@@ -1737,7 +1735,6 @@ def reload_text_encoder(initial=False):
 
 
 def reload_model_weights(sd_model=None, info=None, reuse_dict=False, op='model', force=False):
-    devices.set_cuda_params()
     load_dict = shared.opts.sd_model_dict != model_data.sd_dict
     from modules import lowvram, sd_hijack
     checkpoint_info = info or select_checkpoint(op=op) # are we selecting model or dictionary
@@ -1749,10 +1746,10 @@ def reload_model_weights(sd_model=None, info=None, reuse_dict=False, op='model',
     shared.state = shared_state.State()
     shared.state.begin('Load')
     if load_dict:
-        shared.log.debug(f'Model dict: existing={sd_model is not None} target={checkpoint_info.filename} info={info}')
+        shared.log.debug(f'Load {op} dict: target="{checkpoint_info.filename}" existing={sd_model is not None} info={info}')
     else:
         model_data.sd_dict = 'None'
-        shared.log.debug(f'Load model: existing={sd_model is not None} target={checkpoint_info.filename} info={info}')
+        shared.log.debug(f'Load {op}: target="{checkpoint_info.filename}" existing={sd_model is not None} info={info}')
     if sd_model is None:
         sd_model = model_data.sd_model if op == 'model' or op == 'dict' else model_data.sd_refiner
     if sd_model is None:  # previous model load failed
@@ -1766,7 +1763,7 @@ def reload_model_weights(sd_model=None, info=None, reuse_dict=False, op='model',
         else:
             move_model(sd_model, devices.cpu)
         if (reuse_dict or shared.opts.model_reuse_dict) and not getattr(sd_model, 'has_accelerate', False):
-            shared.log.info('Reusing previous model dictionary')
+            shared.log.info(f'Load {op}: reusing dictionary')
             sd_hijack.model_hijack.undo_hijack(sd_model)
         else:
             unload_model_weights(op=op)
diff --git a/modules/shared.py b/modules/shared.py
index 597b72079..6e377be57 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -1081,9 +1081,11 @@ cmd_opts.disable_extension_access = (cmd_opts.share or cmd_opts.listen or (cmd_o
 devices.backend = devices.get_backend(cmd_opts, opts)
 devices.device = devices.get_optimal_device()
 devices.onnx = [opts.onnx_execution_provider]
+devices.set_cuda_params()
 if opts.onnx_cpu_fallback and 'CPUExecutionProvider' not in devices.onnx:
     devices.onnx.append('CPUExecutionProvider')
 device = devices.device
+
 batch_cond_uncond = opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram)
 parallel_processing_allowed = not cmd_opts.lowvram
 mem_mon = modules.memmon.MemUsageMonitor("MemMon", devices.device)
diff --git a/scripts/xyz_grid_on.py b/scripts/xyz_grid_on.py
index cb13acf10..ac1bc4c2f 100644
--- a/scripts/xyz_grid_on.py
+++ b/scripts/xyz_grid_on.py
@@ -143,9 +143,9 @@ class Script(scripts.Script):
 
     def process(self, p, enabled, x_type, x_values, x_values_dropdown, y_type, y_values, y_values_dropdown, z_type, z_values, z_values_dropdown, csv_mode, draw_legend, no_fixed_seeds, include_grid, include_subgrids, include_images, margin_size): # pylint: disable=W0221
         global active, cache # pylint: disable=W0603
+        cache = None
         if not enabled or active:
             return
-        cache = None
         active = True
         if not no_fixed_seeds:
             processing.fix_seed(p)
diff --git a/scripts/xyz_grid_shared.py b/scripts/xyz_grid_shared.py
index 972e1d2a4..b182a4e80 100644
--- a/scripts/xyz_grid_shared.py
+++ b/scripts/xyz_grid_shared.py
@@ -196,6 +196,10 @@ def apply_lora(p, x, xs):
         return
     x = os.path.basename(x)
     p.prompt = p.prompt + f" <lora:{x}:{shared.opts.extra_networks_default_multiplier}>"
+    if p.all_prompts is not None:
+        p.all_prompts = len(p.all_prompts) * [p.prompt]
+    if p.all_negative_prompts is not None:
+        p.all_negative_prompts = len(p.all_negative_prompts) * [p.prompt]
     shared.log.debug(f'XYZ grid apply LoRA: "{x}"')