App: overcome 50GB limit by using mounted storage

app.py

@@ -6,11 +6,6 @@ import sys
 os.environ["TORCH_COMPILE_DISABLE"] = "1"
 os.environ["TORCHDYNAMO_DISABLE"] = "1"
 
-# Place caches in persistent storage
-os.environ["HF_HOME"] = "/data/.huggingface"
-os.environ["HF_HUB_CACHE"] = "/data/.cache/huggingface/hub"
-os.environ["HF_DATASETS_CACHE"] = "/data/.cache/huggingface/datasets"
-
 # Install xformers for memory-efficient attention
 subprocess.run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)
 
@@ -86,22 +81,15 @@ RESOLUTIONS = {
     "low": {"16:9": (768, 512), "9:16": (512, 768), "1:1": (768, 768)},
 }
 
-# Model repos
-LTX_MODEL_REPO = "Lightricks/LTX-2.3"
-GEMMA_REPO = "google/gemma-3-12b-it-qat-q4_0-unquantized"
-
-# Download model checkpoints
-print("=" * 80)
-print("Downloading LTX-2.3 distilled model + Gemma...")
-print("=" * 80)
+LTX_MOUNT = "/models/ltx"
+GEMMA_MOUNT = "/models/gemma"
 
-checkpoint_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-22b-distilled-1.1.safetensors")
-spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors")
-gemma_root = snapshot_download(repo_id=GEMMA_REPO)
+DISTILLED_FILENAME = "ltx-2.3-22b-distilled.safetensors"
+UPSCALER_FILENAME = "ltx-2.3-spatial-upscaler-x2-1.0.safetensors"
 
-print(f"Checkpoint: {checkpoint_path}")
-print(f"Spatial upsampler: {spatial_upsampler_path}")
-print(f"Gemma root: {gemma_root}")
+ltx_checkpoint = os.path.join(LTX_MOUNT, DISTILLED_FILENAME)
+ltx_upscaler = os.path.join(LTX_MOUNT, UPSCALER_FILENAME)
+gemma_root = GEMMA_MOUNT
 
 # Initialize pipeline WITH text encoder
 pipeline = DistilledPipeline(