Update app.py

app.py

@@ -300,15 +300,44 @@ print(f"Checkpoint: {checkpoint_path}")
 print(f"Spatial upsampler: {spatial_upsampler_path}")
 print(f"[Gemma] Root ready: {gemma_root}")
 
-# Initialize pipeline WITH text encoder and optional audio support
-# ---- Replace block (pipeline init) lines 275-281 ----
-pipeline = LTX23DistilledA2VPipeline(
-    distilled_checkpoint_path=checkpoint_path,
-    spatial_upsampler_path=spatial_upsampler_path,
-    gemma_root=gemma_root,
-    loras=[],
-    quantization=None,  # keep FP8 quantization unchanged
-)
+pipeline = None
+_transformer = None  
+
+def get_pipeline():
+    global pipeline, _transformer
+    if pipeline is None:
+        print("Loading LTX-2.3 Pipeline safely inside ZeroGPU...")
+        pipeline = LTX23DistilledA2VPipeline(
+            distilled_checkpoint_path=checkpoint_path,
+            spatial_upsampler_path=spatial_upsampler_path,
+            gemma_root=gemma_root,
+            loras=[],
+            quantization=None,
+        )
+        
+        ledger = pipeline.model_ledger
+        _transformer = ledger.transformer()
+        _video_encoder = ledger.video_encoder()
+        _video_decoder = ledger.video_decoder()
+        _audio_encoder = ledger.audio_encoder()
+        _audio_decoder = ledger.audio_decoder()
+        _vocoder = ledger.vocoder()
+        _spatial_upsampler = ledger.spatial_upsampler()
+        _text_encoder = ledger.text_encoder()
+        _embeddings_processor = ledger.gemma_embeddings_processor()
+        
+        ledger.transformer = lambda: _transformer
+        ledger.video_encoder = lambda: _video_encoder
+        ledger.video_decoder = lambda: _video_decoder
+        ledger.audio_encoder = lambda: _audio_encoder
+        ledger.audio_decoder = lambda: _audio_decoder
+        ledger.vocoder = lambda: _vocoder
+        ledger.spatial_upsampler = lambda: _spatial_upsampler
+        ledger.text_encoder = lambda: _text_encoder
+        ledger.gemma_embeddings_processor = lambda: _embeddings_processor
+    return pipeline
+
+
 # ----------------------------------------------------------------
 
 def _make_lora_key(singularity_strength: float, teneros_strength: float, pose_strength: float, general_strength: float, motion_strength: float, dreamlay_strength: float, mself_strength: float, dramatic_strength: float, fluid_strength: float, liquid_strength: float, demopose_strength: float, voice_strength: float, realism_strength: float, transition_strength: float, physics_strength: float, reasoning_strength: float, twostep_strength: float, mcfurry_strength: float, dm_strength: float, praxis_strength: float, threed_strength: float, concept_strength: float, bulge_strength: float) -> tuple[str, str]:
@@ -375,7 +404,8 @@ def prepare_lora_cache(
     """
     global PENDING_LORA_KEY, PENDING_LORA_STATE, PENDING_LORA_STATUS
 
-    ledger = pipeline.model_ledger
+        active_pipeline = get_pipeline()
+    ledger = active_pipeline.model_ledger
     key, _ = _make_lora_key(singularity_strength, teneros_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength)
     cache_path = LORA_CACHE_DIR / f"{key}.safetensors"
 
@@ -501,49 +531,6 @@ def apply_prepared_lora_state_to_pipeline():
     print("[LoRA] Prepared LoRA state applied to the pipeline.")
     return True
 
-# ---- REPLACE PRELOAD BLOCK START ----
-# Preload all models for ZeroGPU tensor packing.
-print("Preloading all models (including Gemma and audio components)...")
-ledger = pipeline.model_ledger
-
-# Save the original factory methods so we can rebuild individual components later.
-# These are bound callables on ledger that will call the builder when invoked.
-_orig_transformer_factory = ledger.transformer
-_orig_video_encoder_factory = ledger.video_encoder
-_orig_video_decoder_factory = ledger.video_decoder
-_orig_audio_encoder_factory = ledger.audio_encoder
-_orig_audio_decoder_factory = ledger.audio_decoder
-_orig_vocoder_factory = ledger.vocoder
-_orig_spatial_upsampler_factory = ledger.spatial_upsampler
-_orig_text_encoder_factory = ledger.text_encoder
-_orig_gemma_embeddings_factory = ledger.gemma_embeddings_processor
-
-# Call the original factories once to create the cached instances we will serve by default.
-_transformer = _orig_transformer_factory()
-_video_encoder = _orig_video_encoder_factory()
-_video_decoder = _orig_video_decoder_factory()
-_audio_encoder = _orig_audio_encoder_factory()
-_audio_decoder = _orig_audio_decoder_factory()
-_vocoder = _orig_vocoder_factory()
-_spatial_upsampler = _orig_spatial_upsampler_factory()
-_text_encoder = _orig_text_encoder_factory()
-_embeddings_processor = _orig_gemma_embeddings_factory()
-
-# Replace ledger methods with lightweight lambdas that return the cached instances.
-# We keep the original factories above so we can call them later to rebuild components.
-ledger.transformer = lambda: _transformer
-ledger.video_encoder = lambda: _video_encoder
-ledger.video_decoder = lambda: _video_decoder
-ledger.audio_encoder = lambda: _audio_encoder
-ledger.audio_decoder = lambda: _audio_decoder
-ledger.vocoder = lambda: _vocoder
-ledger.spatial_upsampler = lambda: _spatial_upsampler
-ledger.text_encoder = lambda: _text_encoder
-ledger.gemma_embeddings_processor = lambda: _embeddings_processor
-
-print("All models preloaded (including Gemma text encoder and audio encoder)!")
-# ---- REPLACE PRELOAD BLOCK END ----
-
 print("=" * 80)
 print("Pipeline ready!")
 print("=" * 80)
@@ -700,10 +687,14 @@ def generate_video(
 
         log_memory("before pipeline call")
 
-        apply_prepared_lora_state_to_pipeline()
+       # Safely get the pipeline inside ZeroGPU right before execution
+active_pipeline = get_pipeline()
+
+apply_prepared_lora_state_to_pipeline()
+
+video, audio = active_pipeline(
+    prompt=prompt,
 
-        video, audio = pipeline(
-            prompt=prompt,
             seed=current_seed,
             height=int(height),
             width=int(width),