python3Packages.cuda-tile: 1.4.0 -> 1.7.0

Changelog: https://docs.nvidia.com/cuda/cutile-python/generated/release_notes.html

- The C extension now includes nvvm.h (pynvvm.cpp); add libnvvm for the header.
- setup.py no longer passes dry_run to copy_file; drop that substitution.
- Follow the build-system setuptools pin (84.0.0) and the refactored
  tileiras lookup in _compile.py.

Assisted-by: Claude Code (Claude Opus 5.5)

pkgs/development/python-modules/cuda-tile/default.nix

@@ -20,6 +20,7 @@
 
   # tests
   cupy,
+  llvm,
   numba-cuda,
   numpy,
   pytest-benchmark,
@@ -40,7 +41,7 @@ let
     else
       "${addDriverRunpath.driverLink}/lib";
 
-  # https://github.com/NVIDIA/cutile-python/blob/v1.4.0/cmake/FetchXLAHeaders.cmake#L5-L6
+  # https://github.com/NVIDIA/cutile-python/blob/v1.7.0/cmake/FetchXLAHeaders.cmake#L5-L6
   xla = fetchFromGitHub {
     owner = "openxla";
     repo = "xla";
@@ -50,7 +51,7 @@ let
 in
 buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs: {
   pname = "cuda-tile";
-  version = "1.4.0";
+  version = "1.7.0";
   pyproject = true;
   __structuredAttrs = true;
 
@@ -58,7 +59,7 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
     owner = "NVIDIA";
     repo = "cutile-python";
     tag = "v${finalAttrs.version}";
-    hash = "sha256-R5V69nJLQ3/1995ezH1/WuueA6cm1vhKZdOECqbwPbU=";
+    hash = "sha256-mO5VOusQg6wYzrvkD9SB74O6g5fOnyEdCIIKBkjC7s4=";
   };
 
   patches = [
@@ -67,9 +68,7 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
 
   postPatch = ''
     substituteInPlace pyproject.toml \
-      --replace-fail "setuptools==80.10.2" "setuptools"
-    substituteInPlace setup.py \
-      --replace-fail ' dry_run=self.dry_run)' ' )'
+      --replace-fail "setuptools==84.0.0" "setuptools"
   ''
   # Otherwise fails to find libc
   # xla_ffi.cpp:(.text+0x308): undefined reference to `__stack_chk_fail'
@@ -79,6 +78,13 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
         "target_link_libraries(_cext_shared _cext_static \''${Python_LIBRARIES} \''${asan_library})" \
         "target_link_libraries(_cext_shared _cext_static \''${Python_LIBRARIES} \''${asan_library} c)"
   ''
+  # Don't build with -Werror: GCC 15 flags variables in cext/tile_kernel.cpp as set but not used
+  + ''
+    substituteInPlace CMakeLists.txt \
+      --replace-fail \
+        'treat warnings as errors" OFF)' \
+        'treat warnings as errors" ON)'
+  ''
   # Get rid of the vendored broken logic for finding the CUDA toolkit
   + ''
     rm cmake/FindCUDAToolkit.cmake
@@ -99,8 +105,8 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
   + lib.optionalString cudaPackages.cuda_tileiras.meta.available ''
     substituteInPlace src/cuda/tile/_compile.py \
       --replace-fail \
-        "for toolkit_path in _get_default_cuda_toolkit_paths():" \
-        "for toolkit_path in [\"${lib.getBin cudaPackages.cuda_tileiras}\"]:"
+        "res = _find_file(_get_default_cuda_toolkit_paths(), [\"bin\"], [binary_name]," \
+        "res = _find_file([\"${lib.getBin cudaPackages.cuda_tileiras}\"], [\"bin\"], [binary_name],"
   '';
 
   build-system = [
@@ -120,6 +126,7 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
 
   buildInputs = [
     cudaPackages.cuda_cudart # cuda.h
+    cudaPackages.libnvvm # nvvm.h (libnvvm itself is loaded at run time, by the experimental cuda.lang only)
   ];
 
   dependencies = [
@@ -130,6 +137,7 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
 
   nativeCheckInputs = [
     cupy
+    llvm # FileCheck, for the MLIR tests
     numba-cuda
     numpy
     pytest-benchmark
@@ -143,6 +151,29 @@ buildPythonPackage.override { stdenv = cudaPackages.backendStdenv; } (finalAttrs
   # Run the benchmarks once as tests, without timing them
   pytestFlags = [ "--benchmark-disable" ];
 
+  # The benchmarks look up their tuned configurations by GPU name, which they otherwise get from nvidia-smi
+  preCheck = ''
+    export CUDA_TILE_BENCHMARK_GPU_NAME="$(python -c 'import torch; print(torch.cuda.get_device_name(0))')"
+  '';
+
+  disabledTestPaths = [
+    # Separate packages (task_scheduling, tile_preview, cuda-lang), not installed here; upstream's CI ignores them too
+    "experimental"
+  ];
+
+  disabledTests = [
+    # Deselected by upstream's CI too
+    "int64_index"
+    # These need more memory than a 24 GiB GPU has: a 65536x16384 layer norm, and float32 vector adds of 2^30 elements
+    # (in each layout). Failed tests keep their tensors alive (through their tracebacks), so they also made later
+    # benchmarks run out of memory.
+    "64K-16K"
+    "1G-1G-f32"
+    "1G-1Mx1K-f32"
+    "1G-8Mx128-f32"
+    "1G-32Mx32-f32"
+  ];
+
   # Tests require access to a physical GPU
   doCheck = false;