refactor: split legacy loaders

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

CHANGELOG.md

@@ -22,7 +22,8 @@
   - refactored progress monitoring, job updates and live preview  
   - improved metadata save and restore  
   - startup tracing and optimizations  
-  - threading load locks on model loads
+  - threading load locks on model loads  
+  - refactor native vs legacy model loader  
 - **Schedulers**:
   - [TDD](https://github.com/RedAIGC/Target-Driven-Distillation) new super-fast scheduler that can generate images in 4-8 steps  
     recommended to use with [TDD LoRA](https://huggingface.co/RED-AIGC/TDD/tree/main)  
@@ -40,7 +41,7 @@
   - **XYZ Grid**: add prompt search&replace options: *primary, refine, detailer, all*
   - **SysInfo**: update to collected data and benchmarks  
 - [Wiki/Docs](https://vladmandic.github.io/sdnext-docs/):
-  - updated: Detailer, Install, Debug, Control-HowTo, ZLUDA  
+  - updated: Detailer, Install, Update, Debug, Control-HowTo, ZLUDA  
 - **Fixes**:
   - explict clear caches on model load  
   - lock adetailer commit: `#a89c01d`  
@@ -61,6 +62,8 @@
   - restore args after batch run  
   - flux controlnet  
   - zluda installer  
+  - control inherit parent pipe settings  
+  - control logging  
 
 ## Update for 2024-12-31
 

installer.py

@@ -150,6 +150,11 @@ def get(self):
     log.addHandler(rb)
     log.buffer = rb.buffer
 
+    def quiet_log(quiet: bool=False, *args, **kwargs): # pylint: disable=redefined-outer-name,keyword-arg-before-vararg
+        if not quiet:
+            log.debug(*args, **kwargs)
+    log.quiet = quiet_log
+
     # overrides
     logging.getLogger("urllib3").setLevel(logging.ERROR)
     logging.getLogger("httpx").setLevel(logging.ERROR)

modules/cmd_args.py

@@ -78,7 +78,6 @@ def compatibility_args():
     group_compat.add_argument("--disable-queue", default=os.environ.get("SD_DISABLEQUEUE", False), action='store_true', help=argparse.SUPPRESS)
 
 
-
 def settings_args(opts, args):
     # removed args are added here as hidden in fixed format for compatbility reasons
     group_compat = parser.add_argument_group('Compatibility options')

modules/control/units/controlnet.py

@@ -411,13 +411,14 @@ def __init__(self,
         if dtype is not None:
             self.pipeline = self.pipeline.to(dtype)
 
+        sd_models.copy_diffuser_options(self.pipeline, pipeline)
         if opts.diffusers_offload_mode == 'none':
             sd_models.move_model(self.pipeline, devices.device)
         from modules.sd_models import set_diffuser_offload
         set_diffuser_offload(self.pipeline, 'model')
 
         t1 = time.time()
-        log.debug(f'Control {what} pipeline: class={self.pipeline.__class__.__name__} time={t1-t0:.2f}')
+        debug_log(f'Control {what} pipeline: class={self.pipeline.__class__.__name__} time={t1-t0:.2f}')
 
     def restore(self):
         self.pipeline.unload_lora_weights()

modules/processing_original.py

@@ -27,6 +27,18 @@ def get_conds_with_caching(function, required_prompts, steps, cache):
     cache[0] = (required_prompts, steps)
     return cache[1]
 
+def check_rollback_vae():
+    if shared.cmd_opts.rollback_vae:
+        if not torch.cuda.is_available():
+            shared.log.error("Rollback VAE functionality requires compatible GPU")
+            shared.cmd_opts.rollback_vae = False
+        elif torch.__version__.startswith('1.') or torch.__version__.startswith('2.0'):
+            shared.log.error("Rollback VAE functionality requires Torch 2.1 or higher")
+            shared.cmd_opts.rollback_vae = False
+        elif 0 < torch.cuda.get_device_capability()[0] < 8:
+            shared.log.error('Rollback VAE functionality device capabilities not met')
+            shared.cmd_opts.rollback_vae = False
+
 
 def process_original(p: processing.StableDiffusionProcessing):
     cached_uc = [None, None]
@@ -42,6 +54,7 @@ def process_original(p: processing.StableDiffusionProcessing):
         for x in x_samples_ddim:
             devices.test_for_nans(x, "vae")
     except devices.NansException as e:
+        check_rollback_vae()
         if not shared.opts.no_half and not shared.opts.no_half_vae and shared.cmd_opts.rollback_vae:
             shared.log.warning('Tensor with all NaNs was produced in VAE')
             devices.dtype_vae = torch.bfloat16

modules/sd_detect.py

@@ -49,12 +49,6 @@ def detect_pipeline(f: str, op: str = 'model', warning=True, quiet=False):
                 elif (size > 20000 and size < 40000):
                     guess = 'FLUX'
             # guess by name
-            """
-            if 'LCM_' in f.upper() or 'LCM-' in f.upper() or '_LCM' in f.upper() or '-LCM' in f.upper():
-                if shared.backend == shared.Backend.ORIGINAL:
-                    warn(f'Model detected as LCM model, but attempting to load using backend=original: {op}={f} size={size} MB')
-                guess = 'Latent Consistency Model'
-            """
             if 'instaflow' in f.lower():
                 guess = 'InstaFlow'
             if 'segmoe' in f.lower():