Vendor CCCL and use it to reactivate tensormap

.github/dependabot.yml

@@ -48,6 +48,15 @@ updates:
 
     labels: [CI/CD, dependencies]
 
+  # CCCL submodule on main.
+  - package-ecosystem: gitsubmodule
+    directory: /
+    schedule:
+      interval: "monthly"
+      time: "09:00"
+      timezone: "America/Los_Angeles"
+    open-pull-requests-limit: 1
+
   # GitHub Actions updates targeting the 12.9.x branch
   - package-ecosystem: github-actions
     directory: /

.github/workflows/build-wheel.yml

@@ -69,6 +69,8 @@ jobs:
           # but not historical blobs.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Register problem matchers
         # Compiler and Cython errors from the wheel builds become annotations

.github/workflows/ci-pixi-source-test.yml

@@ -88,6 +88,8 @@ jobs:
           # cuda.core's "cuda.bindings 12.x or 13.x must be installed" guard.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Setup pixi
         uses: ./.github/actions/setup-pixi
@@ -145,6 +147,8 @@ jobs:
           # cuda.core's "cuda.bindings 12.x or 13.x must be installed" guard.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Setup pixi
         uses: ./.github/actions/setup-pixi
@@ -211,6 +215,8 @@ jobs:
           # cuda.core's "cuda.bindings 12.x or 13.x must be installed" guard.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Setup proxy cache
         uses: nv-gha-runners/setup-proxy-cache@main

.github/workflows/coverage.yml

@@ -71,6 +71,8 @@ jobs:
           # but not historical blobs.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Fix workspace ownership
         run: |
@@ -242,6 +244,8 @@ jobs:
           # but not historical blobs.
           fetch-depth: 0
           filter: blob:none
+          # third_party/cccl provides the CCCL headers cuda.core compiles against.
+          submodules: true
 
       - name: Set up Python
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0

.github/workflows/test-sdist-linux.yml

@@ -44,6 +44,7 @@ jobs:
           # the commit graph (`git describe`) but not historical blobs.
           fetch-depth: 0
           filter: blob:none
+          submodules: true
 
       - name: Set up Python
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0

.github/workflows/test-sdist-windows.yml

@@ -53,6 +53,7 @@ jobs:
           # the commit graph (`git describe`) but not historical blobs.
           fetch-depth: 0
           filter: blob:none
+          submodules: true
 
       - name: Set up Python
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "cuda_core/third_party/cccl"]
+	path = cuda_core/third_party/cccl
+	url = https://github.com/NVIDIA/cccl.git
+	branch = main

.pre-commit-config.yaml

@@ -55,7 +55,7 @@ repos:
         language: python
         additional_dependencies:
           - https://files.pythonhosted.org/packages/cc/20/ff623b09d963f88bfde16306a54e12ee5ea43e9b597108672ff3a408aad6/pathspec-0.12.1-py3-none-any.whl
-        exclude: '(.*pixi\.lock)|(\.git_archival\.txt)|(.*\.patch$)|(^cuda_core/cuda/core/_vendored/)'
+        exclude: '(.*pixi\.lock)|(\.git_archival\.txt)|(.*\.patch$)|(^cuda_core/cuda/core/_vendored/)|(^cuda_core/third_party/)'
         args: ["--fix"]
 
       - id: check-generated-file-seals

.spdx-ignore

@@ -5,6 +5,7 @@ LICENSE*
 *.png
 .gitattributes
 .gitignore
+.gitmodules
 requirements*.txt
 
 # Samples are synced to NVIDIA/cuda-samples on release and carry either the
@@ -15,6 +16,8 @@ samples/**/*.py
 cuda_core/cuda/core/_include/dlpack.h
 cuda_core/cuda/core/_include/aoti_shim.h
 cuda_core/cuda/core/_include/aoti_shim.def
+# CCCL submodule (Apache-2.0 WITH LLVM-exception), including the gitlink.
+cuda_core/third_party/**
 
 # Generated by stubgen-pyx; regenerated on every commit so a header would be lost
 cuda_core/cuda/**/*.pyi

CONTRIBUTING.md

@@ -66,25 +66,26 @@ version is read from the *repository root* rather than the package directory. A
 working build therefore needs all of the following:
 
 1. **A real git clone.** Source zips and GitHub "Download ZIP" archives have no
-   git metadata and the build fails outright. (Tarballs produced by
-   `git archive` do work, thanks to the `.git_archival.txt` substitutions
-   configured in `.gitattributes`.)
+   git metadata and the build fails outright. Tarballs produced by
+   `git archive` still resolve a version via the `.git_archival.txt`
+   substitutions in `.gitattributes`, but they do not include submodule
+   contents, so they cannot compile `cuda.core`.
 2. **The full repository**, not just the package subdirectory, because the
    version lookup walks up to the repository root.
 3. **Tags, reaching back at least as far as the most recent tag** matching the
    package you are building. `git describe` needs to find that tag; the history
    between it and your checkout must be present too.
+4. **The CCCL submodule** when building `cuda.core`. That package compiles
+   against the headers pinned in `cuda_core/third_party/cccl`.
 
 ### Recommended clone
 
-The default `git clone` gives you everything you need:
+The default `git clone` plus the submodules gives you everything you need:
 
 ```console
-$ git clone https://github.com/NVIDIA/cuda-python.git
+$ git clone --recurse-submodules https://github.com/NVIDIA/cuda-python.git
 ```
 
-
-
 ### Fixing an existing clone
 
 If you already have a shallow clone:
@@ -93,6 +94,12 @@ If you already have a shallow clone:
 $ git fetch --unshallow --tags
 ```
 
+If you cloned without submodules:
+
+```console
+$ git submodule update --init cuda_core/third_party/cccl
+```
+
 If you are working from a personal fork, your fork's tags stop tracking upstream
 the moment new releases are cut, which silently yields a stale version. Fetch
 tags from upstream directly:

cuda_core/AGENTS.md

@@ -35,6 +35,10 @@ This file describes `cuda_core`, the high-level Pythonic CUDA subpackage in the
   or CUDA headers (`CUDA_HOME`/`CUDA_PATH`) and uses it for build decisions.
 - Source builds require CUDA headers available through `CUDA_HOME` or
   `CUDA_PATH`.
+- CCCL headers come from the pinned `cuda_core/third_party/cccl` submodule, not
+  from the CTK, so the CCCL feature set does not vary with the build
+  environment. There is no fallback: a source build without the submodule
+  fails.
 - `cuda_core` requires `cuda-bindings` at or above a per-major *floor* at build
   time and at run time. At build time it also requires a `cuda.h` of the same
   major.minor as that `cuda-bindings` (NVIDIA/cuda-python#2783). The floors are

cuda_core/MANIFEST.in

@@ -9,3 +9,8 @@ include cuda/core/py.typed
 include NOTICE
 # package-local shared PEP 517 helper
 include _build_shared.py
+# setuptools-scm only records the submodule gitlink, so graft the libcu++
+# headers needed for wheel-from-sdist builds and include their licenses.
+graft third_party/cccl/libcudacxx/include
+include third_party/cccl/LICENSE
+include third_party/cccl/libcudacxx/LICENSE.TXT

cuda_core/NOTICE

@@ -28,3 +28,10 @@ Source: https://github.com/pytorch/pytorch
 Vendored at: cuda/core/_include/aoti_shim.h, and the accompanying
 cuda/core/_include/aoti_shim.def, which declares the same AOT Inductor
 stable C ABI symbol names for the MSVC linker on Windows.
+
+CCCL (libcu++)
+Copyright (c) NVIDIA CORPORATION & AFFILIATES.
+Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+Source: https://github.com/NVIDIA/cccl
+Pinned as the ``cuda_core/third_party/cccl`` git submodule. The libcu++
+headers are used at build time and are not installed into the wheel.

cuda_core/build_hooks.py

@@ -178,6 +178,13 @@ def _read_cuda_h_version(cuda_path: str) -> int:
     )
 
 
+def _cccl_toolchain_flags(name, compile_args, link_args):
+    if name == "msvc":
+        # CCCL headers require the conforming MSVC preprocessor.
+        compile_args.append("/Zc:preprocessor")
+    return compile_args, link_args
+
+
 def _resolve_toolchain(debug=False, compile_for_coverage=False):
     """Resolve the C/C++ toolchain from CUDA_PYTHON_TOOLCHAIN (cuda.core flags).
 
@@ -190,6 +197,7 @@ def _resolve_toolchain(debug=False, compile_for_coverage=False):
         debug=debug,
         compile_for_coverage=compile_for_coverage,
         warnings_as_errors=WARNINGS_AS_ERRORS,
+        tweak=_cccl_toolchain_flags,
     )
 
 
@@ -319,6 +327,10 @@ def _build_define_macros(cuda_major: str, bindings_cuda_version: int | None = No
 # used later by setup()
 _extensions = None
 
+# Keep the pinned CCCL checkout inside cuda_core so it also travels into
+# cibuildwheel's Linux container, which copies only this package directory.
+_CCCL_SUBMODULE_DIR = Path(__file__).resolve().parent / "third_party" / "cccl"
+
 # Records the build configuration (CUDA major, toolchain, debug/coverage) of
 # the last completed build for this extension ABI, so setup.py can force
 # build_ext when it changes. Written by record_build_config() after the
@@ -372,7 +384,7 @@ def _relativize_extension_sources(extensions) -> None:
 def _extension_sources(mod_name):
     """The module's .pyx plus its C++, if any: every .cpp under
     cuda/core/_cpp/<stem>/, or the single legacy file cuda/core/_cpp/<stem>.cpp.
-    Example: _tensor_map.pyx compiles _cpp/tensor_map.cpp."""
+    Example: _rt.pyx compiles every .cpp under _cpp/rt/."""
     sources = [f"cuda/core/{mod_name}.pyx"]
     cpp_stem = Path("cuda", "core", "_cpp", mod_name.lstrip("_"))
     if cpp_stem.is_dir():
@@ -408,6 +420,31 @@ def _extension_depends():
     )
 
 
+@functools.cache
+def _get_cccl_include_dirs() -> list[str]:
+    """Return the include dirs providing the CCCL headers to compile against.
+
+    Always the pinned submodule, never the CTK's CCCL: cuda.core's feature set
+    must not vary with whatever the build environment happens to ship, so a
+    missing submodule is a build error rather than a fallback.
+
+    Only libcu++ is listed: a CCCL checkout keeps one include root per
+    component (see the "GitHub" section of CCCL's README) and cuda.core
+    includes nothing from cub or thrust.
+    """
+    libcudacxx_include = _CCCL_SUBMODULE_DIR / "libcudacxx" / "include"
+
+    # <cuda/tma> is where _tensor_map.pyx gets cuda::make_tma_descriptor, and
+    # it postdates the rest of the layout, so it doubles as the version check.
+    if not (libcudacxx_include / "cuda" / "tma").is_file():
+        raise RuntimeError(
+            f"The CCCL submodule at {_CCCL_SUBMODULE_DIR} is missing or not initialized. Run\n"
+            "    git submodule update --init cuda_core/third_party/cccl"
+        )
+
+    return [str(libcudacxx_include)]
+
+
 def _build_cuda_core(debug=False):
     # Customizing the build hooks is needed because we must defer cythonization until cuda-bindings,
     # now a required build-time dependency that's dynamically installed via the other hook below,
@@ -453,7 +490,8 @@ def module_names():
                 continue
             yield mod
 
-    all_include_dirs = [os.path.join(cuda_path, "include")]
+    # The pinned CCCL must precede the toolkit's bundled headers.
+    all_include_dirs = [*_get_cccl_include_dirs(), os.path.join(cuda_path, "include")]
 
     # Resolve the C/C++ toolchain (CUDA_PYTHON_TOOLCHAIN). The default (gnu on
     # Linux, msvc on Windows) reproduces the previous build behavior and does