Update app.py

app.py

@@ -266,17 +266,6 @@ print("=" * 80)
 print("Downloading LTX-2.3 distilled model + Gemma...")
 print("=" * 80)
 
-# ----------------------------
-# Pipeline cache for LoRA strengths (keeps at most 2 pipelines to limit VRAM)
-# ----------------------------
-# Use rounded strengths as keys (2 decimal places)
-pipeline_cache: OrderedDict[float, LTX23DistilledA2VPipeline] = OrderedDict()
-# Record the current pipeline's LoRA strength (we built the module above with lora_descriptor default 1.0)
-current_lora_strength: float = round(1.0, 2)
-pipeline_cache[current_lora_strength] = pipeline
-CACHE_MAX_SIZE = 2  # keep at most two pipeline instances in memory
-print(f"[CACHE] initialized pipeline cache with strength={current_lora_strength}")
-
 checkpoint_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-22b-distilled.safetensors")
 spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
 gemma_root = snapshot_download(repo_id=GEMMA_REPO)
@@ -338,6 +327,14 @@ print("=" * 80)
 print("Pipeline ready!")
 print("=" * 80)
 
+# ----------------------------
+# Pipeline cache for LoRA strengths (keeps at most 2 pipelines to limit VRAM)
+# ----------------------------
+pipeline_cache: OrderedDict[float, LTX23DistilledA2VPipeline] = OrderedDict()
+current_lora_strength: float = round(1.0, 2)
+pipeline_cache[current_lora_strength] = pipeline
+CACHE_MAX_SIZE = 2
+print(f"[CACHE] initialized pipeline cache with strength={current_lora_strength}")
 
 def log_memory(tag: str):
     if torch.cuda.is_available():