You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Harden multi-version CUDA toolchain selection on top of rebased #422
- Make redist component repo names unique per redist tag to avoid collisions
when multiple `cuda.redist_json` entries resolve to the same version
- Deduplicate generated version selectors and namespace version config_settings
per component to prevent `config_setting` conflicts
- Add platform-aware redist selection in repo rules and align deliverable tool
labels/aliases (`nvcc`, `nvlink`, `bin2c`, `fatbinary`, `ptxas`, `cicc`)
- Improve clang CUDA path handling for deliverable layouts to avoid invalid
`--cuda-path` resolution
- Extend dummy/tool alias coverage and templates for missing components/files
- Add multi-version integration test coverage and skip linux-redist tests on
Windows runners; guard CI redist override to NVIDIA-source cases
fail("Multiple conflicting toolkits declared for name {} ({} and {}".format(toolkit.name, toolkit.toolkit_path, registrations[toolkit.name].toolkit_path))
185
+
fail("Multiple conflicting toolkits declared for name {} ({} and {}".format(
186
+
toolkit.name,
187
+
toolkit.toolkit_path,
188
+
registrations[toolkit.name].toolkit_path,
189
+
))
179
190
else:
180
191
registrations[toolkit.name] =toolkit
181
192
182
-
iflen(registrations) >1:
183
-
fail("multiple cuda.toolkit is not supported")
184
-
185
193
for_, toolkitinregistrations.items():
186
194
ifcomponents_mapping!=None:
187
195
# Always use the maximum version so the toolkit includes all components.
0 commit comments