fix lora on model change

Signed-off-by: Vladimir Mandic <[email protected]>

CHANGELOG.md

@@ -36,6 +36,7 @@ NYE refresh release with quite a few optimizatios and bug fixes...
 - **Fixes**:  
   - flux pipeline switches: txt/img/inpaint  
   - flux custom unet loader for bnb  
+  - flux do not requantize already quantized model
   - interrogate caption with T5  
   - on-the-fly quantization using TorchAO  
   - remove concurrent preview requests  

installer.py

@@ -255,7 +255,7 @@ def uninstall(package, quiet = False):
 
 
 @lru_cache()
-def pip(arg: str, ignore: bool = False, quiet: bool = False, uv = True):
+def pip(arg: str, ignore: bool = False, quiet: bool = True, uv = True):
     originalArg = arg
     arg = arg.replace('>=', '==')
     package = arg.replace("install", "").replace("--upgrade", "").replace("--no-deps", "").replace("--force", "").replace(" ", " ").strip()

modules/model_flux.py

@@ -143,7 +143,9 @@ def quant_flux_bnb(checkpoint_info, transformer, text_encoder_2):
 """
 
 
-def load_quants(kwargs, repo_id, cache_dir):
+def load_quants(kwargs, repo_id, cache_dir, allow_quant):
+    if not allow_quant:
+        return kwargs
     quant_args = {}
     quant_args = model_quant.create_bnb_config(quant_args)
     if quant_args:
@@ -359,10 +361,10 @@ def load_flux(checkpoint_info, diffusers_load_config): # triggered by opts.sd_ch
             except Exception:
                 pass
 
-    allow_quant = 'gguf' not in (sd_unet.loaded_unet or '')
+    allow_quant = 'gguf' not in (sd_unet.loaded_unet or '') and (quant is None or quant == 'none')
     fn = checkpoint_info.path
     if (fn is None) or (not os.path.exists(fn) or os.path.isdir(fn)):
-        kwargs = load_quants(kwargs, repo_id, cache_dir=shared.opts.diffusers_dir)
+        kwargs = load_quants(kwargs, repo_id, cache_dir=shared.opts.diffusers_dir, allow_quant=allow_quant)
     kwargs = model_quant.create_bnb_config(kwargs, allow_quant)
     kwargs = model_quant.create_ao_config(kwargs, allow_quant)
     if fn.endswith('.safetensors') and os.path.isfile(fn):

modules/sd_models.py

@@ -1622,6 +1622,11 @@ def unload_model_weights(op='model'):
             model_data.sd_model = None
             devices.torch_gc(force=True)
             shared.log.debug(f'Unload weights {op}: {memory_stats()}')
+            if not shared.opts.lora_legacy:
+                from modules.lora import networks
+                networks.loaded_networks.clear()
+                networks.previously_loaded_networks.clear()
+                networks.lora_cache.clear()
     elif op == 'refiner':
         if model_data.sd_refiner:
             if not shared.native:

modules/sd_samplers.py

@@ -61,7 +61,7 @@ def create_sampler(name, model):
                 model.prior_pipe.scheduler = copy.deepcopy(model.default_scheduler)
                 model.prior_pipe.scheduler.config.clip_sample = False
         config = {k: v for k, v in model.scheduler.config.items() if not k.startswith('_')}
-        shared.log.debug(f'Sampler: default class={current}: {config}')
+        shared.log.debug(f'Sampler: "default" class={current}: {config}')
         if "flow" in model.scheduler.__class__.__name__.lower():
             shared.state.prediction_type = "flow_prediction"
         elif hasattr(model.scheduler, "config") and hasattr(model.scheduler.config, "prediction_type"):