mirror of
https://github.com/NixOS/nixpkgs.git
synced 2026-11-01 09:39:56 +00:00
809eea01c7
The gpuCheck variant (doCheck = true) never got as far as running the tests: - It added CCCL to the CUDA_HOME used for the build, which makes <cuda/tma> visible to cuda/core/_cpp/tensor_map.cpp. That compiles in a CCCL code path that needs crt/host_defines.h, then fails to import with "undefined symbol: cudaGetErrorString". Upstream never compiles that path (the CUDA Toolkit keeps CCCL in include/cccl/, which the build doesn't add), and has since disabled it explicitly; backport that change. The build's CUDA_HOME is now the cuda_cudart headers alone, without a symlinkJoin. - The tests and two examples derive the include paths they pass to NVRTC from $CUDA_HOME/include and $CUDA_HOME/include/cccl, assuming a single toolkit root. Patch them to use the separate CCCL and cudart include directories instead. - The tests import cuda.bindings._test_helpers, which the packaged cuda-bindings release doesn't ship; import the copy in the source tree. - A trailing comma in a parametrize argument list makes pytest 9 unpack a NumPy array sample, failing collection of tests/test_utils.py. - The checkpoint scenarios run in a child interpreter started with -I, which ignores PYTHONPATH and can't import pytest. Also disable tests that fail because cuda-bindings 13.4 has enum members this cuda.core release doesn't map yet, and one that assumes green context SM partitions of at least 8 on GB10. Assisted-by: Claude Code (Claude Opus 5.5, claude-opus-5-5)
72 lines
2.6 KiB
Diff
72 lines
2.6 KiB
Diff
--- a/examples/thread_block_cluster.py
|
|
+++ b/examples/thread_block_cluster.py
|
|
@@ -70,18 +70,8 @@
|
|
print("This example requires NumPy 2.2.5 or later", file=sys.stderr)
|
|
sys.exit(1)
|
|
|
|
- cuda_path = get_cuda_path_or_home()
|
|
- if cuda_path is None:
|
|
- print("This example requires CUDA_PATH or CUDA_HOME to point to a CUDA toolkit.", file=sys.stderr)
|
|
- sys.exit(1)
|
|
- cuda_include = os.path.join(cuda_path, "include")
|
|
- if not os.path.isdir(cuda_include):
|
|
- print(f"CUDA include directory not found: {cuda_include}", file=sys.stderr)
|
|
- sys.exit(1)
|
|
- include_path = [cuda_include]
|
|
- cccl_include = os.path.join(cuda_include, "cccl")
|
|
- if os.path.isdir(cccl_include):
|
|
- include_path.insert(0, cccl_include)
|
|
+ # Nixpkgs installs CCCL and the CUDA runtime headers in separate prefixes, not under one toolkit root.
|
|
+ include_path = ["@cccl@", "@cudartInclude@"]
|
|
|
|
dev = Device()
|
|
arch = dev.compute_capability
|
|
--- a/examples/tma_tensor_map.py
|
|
+++ b/examples/tma_tensor_map.py
|
|
@@ -108,21 +108,8 @@
|
|
|
|
|
|
def _get_cccl_include_paths():
|
|
- cuda_path = get_cuda_path_or_home()
|
|
- if cuda_path is None:
|
|
- print("This example requires CUDA_PATH or CUDA_HOME to point to a CUDA toolkit.", file=sys.stderr)
|
|
- sys.exit(1)
|
|
-
|
|
- cuda_include = os.path.join(cuda_path, "include")
|
|
- if not os.path.isdir(cuda_include):
|
|
- print(f"CUDA include directory not found: {cuda_include}", file=sys.stderr)
|
|
- sys.exit(1)
|
|
-
|
|
- include_path = [cuda_include]
|
|
- cccl_include = os.path.join(cuda_include, "cccl")
|
|
- if os.path.isdir(cccl_include):
|
|
- include_path.insert(0, cccl_include)
|
|
- return include_path
|
|
+ # Nixpkgs installs CCCL and the CUDA runtime headers in separate prefixes, not under one toolkit root.
|
|
+ return ["@cccl@", "@cudartInclude@"]
|
|
|
|
|
|
def main():
|
|
--- a/tests/helpers/__init__.py
|
|
+++ b/tests/helpers/__init__.py
|
|
@@ -10,16 +10,9 @@
|
|
from cuda_python_test_helpers import *
|
|
|
|
CUDA_PATH = get_cuda_path_or_home()
|
|
-CUDA_INCLUDE_PATH = None
|
|
-CCCL_INCLUDE_PATHS = None
|
|
-if CUDA_PATH is not None:
|
|
- path = os.path.join(CUDA_PATH, "include")
|
|
- if os.path.isdir(path):
|
|
- CUDA_INCLUDE_PATH = path
|
|
- CCCL_INCLUDE_PATHS = (path,)
|
|
- path = os.path.join(path, "cccl")
|
|
- if os.path.isdir(path):
|
|
- CCCL_INCLUDE_PATHS = (path,) + CCCL_INCLUDE_PATHS
|
|
+# Nixpkgs installs CCCL and the CUDA runtime headers in separate prefixes, not under one toolkit root.
|
|
+CUDA_INCLUDE_PATH = "@cudartInclude@"
|
|
+CCCL_INCLUDE_PATHS = ("@cccl@", "@cudartInclude@")
|
|
|
|
|
|
@functools.cache
|