Check for kernels before loading, not after

ltx_packed_codec.py

@@ -511,6 +511,46 @@ def resident_bytes(model) -> int:
     return total
 
 
+def check_gpu_kernels(device: str | None = None) -> None:
+    """Refuse a wheel that has no kernels for this card, and say what to install.
+
+    Nothing in this format needs anything unusual from a GPU - no fp8 units, no
+    minimum compute capability, no bf16 tensor cores. What can be missing is
+    PyTorch's own kernels: the current default wheel on PyPI is a cu130 build,
+    and cu130 dropped Volta.
+
+    Left alone, that surfaces as
+
+        CUDA error: no kernel image is available for execution on the device
+
+    raised from the first kernel launch - which is after an 8.46 GB load, inside
+    somebody else's library, and long after `torch.cuda.is_available()` returned
+    True. Checking `get_arch_list` costs nothing and moves the failure to the
+    place where the fix makes sense.
+    """
+    if device is not None and not str(device).startswith("cuda"):
+        return
+    if not torch.cuda.is_available():
+        return
+
+    major, minor = torch.cuda.get_device_capability(0)
+    arch = f"sm_{major}{minor}"
+    compiled = torch.cuda.get_arch_list()
+    # An empty list means a build that does not report them; do not guess.
+    if not compiled or arch in compiled:
+        return
+
+    name = torch.cuda.get_device_name(0)
+    raise SystemExit(
+        f"this torch ({torch.__version__}) has no kernels for {name} ({arch}).\n"
+        f"It was built for {', '.join(compiled)}, and the first CUDA op would "
+        f"fail with 'no kernel image is available for execution on the device'.\n"
+        f"The model is fine - it needs no custom kernels. Install a torch built "
+        f"for your card, e.g. for {arch}:\n"
+        f"  pip install torch --index-url https://download.pytorch.org/whl/cu128\n"
+        f"or pass device='cpu' to load without touching the GPU.")
+
+
 def load_packed_model(model_dir: str, packed_path: Path | str,
                       gpu_budget: str | None = None, resident: bool = False,
                       fold_pre_scale: bool = False, device: str | None = None):
@@ -537,6 +577,8 @@ def load_packed_model(model_dir: str, packed_path: Path | str,
     import os
     import re
 
+    check_gpu_kernels(device)
+
     if gpu_budget is None:
         gpu_budget = os.environ.get("LTX_PACKED_GPU_BUDGET", "13GiB")