Upload folder using huggingface_hub

app.py

@@ -238,18 +238,44 @@ def _watchdog():
 
 threading.Thread(target=_watchdog, daemon=True).start()
 
+@contextmanager
+def _zerogpu_patching_off():
+    """Load weights with the ZeroGPU torch hijack temporarily disabled.
+
+    `import spaces` installs a TorchFunctionMode + TorchDispatchMode over all of torch.
+    `safetensors.load_file` slices the file's storage once per tensor, and every one of
+    those slices was being routed through __torch_function__ -> dynamo ->
+    __torch_dispatch__. Across an 11B model that interception, not the I/O, is what
+    pushed a ~20s read past the 60-minute launch timeout. (The same file reads at
+    1.5 GB/s from this mount in a process that has not imported spaces.)
+
+    Reading happens entirely on CPU, so nothing here needs the hijack. It goes straight
+    back on afterwards, before any .to("cuda") — that call is what registers the fake
+    CUDA tensors and drives the pack step, and it must run patched.
+    """
+    from spaces.zero.torch import patch as zerogpu_patch, unpatch as zerogpu_unpatch
+
+    zerogpu_unpatch()
+    try:
+        yield
+    finally:
+        zerogpu_patch()
+
+
 print("Loading Open-Sora v2 weights...", flush=True)
 _t0 = time.perf_counter()
-model = _build("open-sora (11B)", cfg.model)
-model_ae = _build("hunyuan vae", cfg.ae)
-model_t5 = _build("t5-xxl", cfg.t5)
-model_clip = _build("clip-l", cfg.clip)
-optional_models = {
-    "img_flux": _build("flux (12B)", cfg.img_flux),
-    "img_flux_ae": _build("flux vae", cfg.img_flux_ae),
-}
+with _zerogpu_patching_off():
+    model = _build("open-sora (11B)", cfg.model)
+    model_ae = _build("hunyuan vae", cfg.ae)
+    model_t5 = _build("t5-xxl", cfg.t5)
+    model_clip = _build("clip-l", cfg.clip)
+    optional_models = {
+        "img_flux": _build("flux (12B)", cfg.img_flux),
+        "img_flux_ae": _build("flux vae", cfg.img_flux_ae),
+    }
 print(f"Weights read in {time.perf_counter() - _t0:.0f}s", flush=True)
 
+# Patching is back on, so these are intercepted and packed as ZeroGPU expects.
 _t0 = time.perf_counter()
 for module in (model, model_ae, model_t5, model_clip, *optional_models.values()):
     module.to("cuda")