build: parametrize CUDA version for wheels and docker

- Add cu128/cu130 build matrix to release workflow
- Parametrize Dockerfile and docker-compose with CUDA_TAG build arg
- Allow csrc/ and setup.py in docker context via .dockerignore
- Add nvcc/torch CUDA version mismatch preflight warning in setup.py
- Add cuda_toolkit_version() helper in csrc/build.py
- Use at::IntArrayRef explicitly to fix ATen overload ambiguity
- Guard kernels with CUDART_VERSION >= 11020 check
- Remove invalid [tool.pip] section from pyproject.toml

setup.py

@@ -1,5 +1,6 @@
 import os
 import sys
+import warnings
 from pathlib import Path
 
 from setuptools import setup
@@ -32,7 +33,24 @@ if _should_build():
     import torch
     from torch.utils.cpp_extension import BuildExtension, CUDAExtension
 
-    from csrc.build import REGISTRY
+    from csrc.build import REGISTRY, cuda_toolkit_version
+
+    # Preflight: warn if nvcc major version != torch's bundled CUDA major version.
+    # A mismatch (e.g. nvcc 13.0 + cu128 torch) causes cryptic ABI/header errors.
+    nvcc_ver = cuda_toolkit_version()
+    torch_cuda = torch.version.cuda
+    if nvcc_ver is not None and torch_cuda is not None:
+        torch_major = int(torch_cuda.split(".")[0])
+        if nvcc_ver[0] != torch_major:
+            warnings.warn(
+                f"CUDA version mismatch: nvcc is {nvcc_ver[0]}.{nvcc_ver[1]} "
+                f"but torch was built with CUDA {torch_cuda}. "
+                f"This may cause compilation errors. "
+                f"Install a matching torch wheel: "
+                f"pip install torch --index-url "
+                f"https://download.pytorch.org/whl/cu{nvcc_ver[0]}{nvcc_ver[1]}",
+                stacklevel=2,
+            )
 
     _torch_lib = torch.utils.cpp_extension.library_paths()[0]