Configure video pipeline and text encoder GPUs with memory guidance

video_worker.py

@@ -3,10 +3,16 @@ import json
 import os
 import sys
 
-# Keep stdout exclusively for bounded protocol replies, discard library output.
-protocol=os.fdopen(os.dup(sys.stdout.fileno()),'w',buffering=1)
-os.dup2(os.open(os.devnull,os.O_WRONLY),sys.stdout.fileno())
 def reply(value):protocol.write(json.dumps(value)+'\n')
+
+class EncoderOnDevice:
+ """Encode on a separate GPU; move all returned conditioning to the diffusion GPU."""
+ def __init__(self,encoder,source,target):self.encoder=encoder;self.source=source;self.target=target
+ def __call__(self,*args,**kwargs):
+  import torch
+  with torch.cuda.device(self.source):outputs=self.encoder(*args,**kwargs)
+  return [type(output)(*(value.to(self.target) if value is not None else None for value in output)) for output in outputs]
+
 def main():
  import torch
  from ltx_pipelines.distilled import DistilledPipeline
@@ -17,6 +23,11 @@ def main():
  config=json.loads(sys.stdin.readline())
  paths=ModelPaths.from_split(**config['paths'])
  pipeline=DistilledPipeline(model_paths=paths,spatial_upsampler_path=config['spatial_upsampler'],loras=(),device=torch.device('cuda:0'),offload_mode=OffloadMode.DISK)
+ encoder_device=torch.device(config.get('text_encoder_device','cuda:0'))
+ if encoder_device!=pipeline.device:
+  from ltx_pipelines.utils.blocks import PromptEncoder
+  encoder=PromptEncoder(paths,torch.bfloat16,encoder_device,offload_mode=OffloadMode.DISK)
+  pipeline.prompt_encoder=EncoderOnDevice(encoder,encoder_device,pipeline.device)
  # Validate all split pack metadata without materializing BF16 weights in RAM.
  from ltx_core.text_encoders.gemma.gemma_assets import GemmaAssets
  GemmaAssets.load(config['paths']['text_encoder_path'])
@@ -24,7 +35,7 @@ def main():
   def __init__(self,target,label):self.target=target;self.label=label
   def __getattr__(self,name):return getattr(self.target,name)
   def __call__(self,*args,**kwargs):
-   reply({'state':'progress','phase':self.label})
+   reply({'state':'progress','phase':self.label+(' · '+config.get('device_names',{}).get('text_encoder' if self.label.startswith('Textencoder') else 'video',''))})
    return self.target(*args,**kwargs)
  for attr,label in [('prompt_encoder','Textencoder · Prompt verarbeiten'),('stage','Videogewichte laden und Diffusion berechnen'),('upsampler','Video hochskalieren'),('video_decoder','Videobilder dekodieren'),('audio_decoder','Audiospur dekodieren')]:setattr(pipeline,attr,Phase(getattr(pipeline,attr),label))
  reply({'state':'ready'})
@@ -37,5 +48,9 @@ def main():
    reply({'state':'complete'})
   except torch.OutOfMemoryError:reply({'state':'failed','error':'GPU-Speicher erschöpft (CUDA OOM). Auflösung oder Bildanzahl reduzieren.'});return
   except Exception:reply({'state':'failed','error':'LTX-Generierung fehlgeschlagen. Modellformat, Laufzeit und Speicher prüfen.'});return
-try:main()
-except Exception:reply({'state':'failed','error':'LTX-Profil konnte nicht vorbereitet werden. Komponenten oder Laufzeit inkompatibel.'})
+if __name__=='__main__':
+ # Keep stdout exclusively for bounded protocol replies, discard library output.
+ protocol=os.fdopen(os.dup(sys.stdout.fileno()),'w',buffering=1)
+ os.dup2(os.open(os.devnull,os.O_WRONLY),sys.stdout.fileno())
+ try:main()
+ except Exception:reply({'state':'failed','error':'LTX-Profil konnte nicht vorbereitet werden. Komponenten oder Laufzeit inkompatibel.'})