lora: restore pipeline type if reload/recompile needed

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -40,6 +40,7 @@ TBD
   - ui: fix collapsible panels  
   - svd: fix stable-video-diffusion dtype mismatch  
   - animatediff: disable sdnq if used  
+  - lora: restore pipeline type if reload/recompile needed  
   - process: improve send-to functionality  
   - control: safe load non-sparse controlnet  
   - control: fix marigold preprocessor with bfloat16  

modules/intel/openvino/__init__.py

@@ -73,6 +73,13 @@ if hasattr(torch, "float8_e8m0fnu"):
     dtype_mapping[torch.float8_e8m0fnu] = Type.f8e8m0
 
 
+warned = False
+def warn_once(msg):
+    global warned
+    if not warned:
+        shared.log.warning(msg)
+        warned = True
+
 class OpenVINOGraphModule(torch.nn.Module):
     def __init__(self, gm, partition_id, use_python_fusion_cache, model_hash_str: str = None, file_name="", int_inputs=[]):
         super().__init__()
@@ -128,7 +135,7 @@ def get_device():
         device = "GPU.0"
     else:
         device = core.available_devices[-1]
-        shared.log.warning(f"OpenVINO: No compatible GPU detected! Using {device}")
+        warn_once(f"OpenVINO: device={device} no compatible GPU detected")
     return device
 
 

modules/lora/lora_load.py

@@ -157,11 +157,14 @@ def maybe_recompile_model(names, te_multipliers):
             recompile_model = True
             shared.compiled_model_state.lora_model = []
     if recompile_model:
+        current_task = sd_models.get_diffusers_task(shared.sd_model)
+        shared.log.debug(f'Compile: task={current_task} force model reload')
         backup_cuda_compile = shared.opts.cuda_compile
         backup_scheduler = getattr(sd_model, "scheduler", None)
         sd_models.unload_model_weights(op='model')
         shared.opts.cuda_compile = []
         sd_models.reload_model_weights(op='model')
+        shared.sd_model = sd_models.set_diffuser_pipe(shared.sd_model, current_task)
         shared.opts.cuda_compile = backup_cuda_compile
         if backup_scheduler is not None:
             sd_model.scheduler = backup_scheduler
@@ -247,7 +250,7 @@ def network_load(names, te_multipliers=None, unet_multipliers=None, dyn_dims=Non
             try:
                 lora_scale = te_multipliers[i] if te_multipliers else shared.opts.extra_networks_default_multiplier
                 lora_module = lora_modules[i] if lora_modules and len(lora_modules) > i else None
-                if recompile_model:
+                if recompile_model and shared.compiled_model_state is not None:
                     shared.compiled_model_state.lora_model.append(f"{name}:{lora_scale}")
                 lora_method = lora_overrides.get_method(shorthash)
                 if lora_method == 'diffusers':
@@ -307,9 +310,13 @@ def network_load(names, te_multipliers=None, unet_multipliers=None, dyn_dims=Non
 
     if recompile_model:
         shared.log.info("Network load: type=LoRA recompiling model")
-        backup_lora_model = shared.compiled_model_state.lora_model
+        if shared.compiled_model_state is not None:
+            backup_lora_model = shared.compiled_model_state.lora_model
+        else:
+            backup_lora_model = []
         if 'Model' in shared.opts.cuda_compile:
             sd_model = sd_models_compile.compile_diffusers(sd_model)
-        shared.compiled_model_state.lora_model = backup_lora_model
+        if shared.compiled_model_state is not None:
+            shared.compiled_model_state.lora_model = backup_lora_model
 
     l.timer.load = time.time() - t0

modules/sd_models.py

@@ -1263,6 +1263,7 @@ def clear_caches(full:bool=False):
 
 
 def unload_model_weights(op='model'):
+    fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access
     clear_caches(full=True)
     if shared.compiled_model_state is not None:
         shared.compiled_model_state.compiled_cache.clear()
@@ -1275,14 +1276,14 @@ def unload_model_weights(op='model'):
             move_model(model_data.sd_model, 'meta')
         model_data.sd_model = None
         devices.torch_gc(force=True, reason='unload')
-        shared.log.debug(f'Unload {op}: {memory_stats()} after')
+        shared.log.debug(f'Unload {op}: {memory_stats()} fn={fn}')
     elif (op == 'refiner') and model_data.sd_refiner:
         shared.log.debug(f'Current {op}: {memory_stats()}')
         disable_offload(model_data.sd_refiner)
         move_model(model_data.sd_refiner, 'meta')
         model_data.sd_refiner = None
         devices.torch_gc(force=True, reason='unload')
-        shared.log.debug(f'Unload {op}: {memory_stats()}')
+        shared.log.debug(f'Unload {op}: {memory_stats()}  fn={fn}')
 
 
 def hf_auth_check(checkpoint_info, force:bool=False):