Update app.py

app.py

@@ -14,16 +14,13 @@ if not os.path.exists(LTX_REPO_DIR):
     print(f"Cloning {LTX_REPO_URL}...")
     subprocess.run(["git", "clone", "--depth", "1", LTX_REPO_URL, LTX_REPO_DIR], check=True)
 
-try:
-    import ltx_pipelines  # noqa: F401
-except ImportError:
-    print("Installing ltx-core and ltx-pipelines...")
-    subprocess.run(
-        [sys.executable, "-m", "pip", "install", "-e",
-         os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
-         "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines")],
-        check=True,
-    )
+print("Installing ltx-core and ltx-pipelines from cloned repo...")
+subprocess.run(
+    [sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", "-e",
+     os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
+     "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines")],
+    check=True,
+)
 
 sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
 sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))
@@ -93,6 +90,26 @@ pipeline = DistilledPipeline(
     quantization=QuantizationPolicy.fp8_cast(),
 )
 
+# Preload all models so first request is fast.
+# On ZeroGPU, .to('cuda') is intercepted and actual GPU allocation
+# happens inside the @spaces.GPU decorated function.
+print("Preloading models...")
+ledger = pipeline.model_ledger
+_transformer = ledger.transformer()
+_video_encoder = ledger.video_encoder()
+_video_decoder = ledger.video_decoder()
+_audio_decoder = ledger.audio_decoder()
+_vocoder = ledger.vocoder()
+_spatial_upsampler = ledger.spatial_upsampler()
+
+ledger.transformer = lambda: _transformer
+ledger.video_encoder = lambda: _video_encoder
+ledger.video_decoder = lambda: _video_decoder
+ledger.audio_decoder = lambda: _audio_decoder
+ledger.vocoder = lambda: _vocoder
+ledger.spatial_upsampler = lambda: _spatial_upsampler
+print("All models preloaded!")
+
 # Connect to text encoder space
 print(f"Connecting to text encoder space: {TEXT_ENCODER_SPACE}")
 try: