Skip to content

Commit 6dae0d7

Browse files
mapatel-amdclaude
andcommitted
ci: use venv amdclang++ and bump stable pin to 7.14.0a20260616
whl-multi-arch converts symlinks to hardlinks during install. After rocm-sdk init, _rocm_sdk_devel/bin/amdclang++ is a hardlink whose exec stub resolves siblings via /proc/self/exe — it looks for _rocm_sdk_devel/bin/amdllvm, which was absent in all whl packages before 7.14.0a20260616 (TheRock PR #5810 / issue #5726). Switch CXX from {rocm}/bin/amdclang++ to {venv}/bin/amdclang++. The venv path is a pip entry point that calls os.execv directly on lib/llvm/bin/amdclang++, where amdllvm has always existed. This makes the CI immune to future regressions of the same class. Also bump the stable pin to 7.14.0a20260616, the first version that includes the upstream amdllvm fix in the whl package itself. Co-Authored-By: Claude Sonnet 4 <noreply@anthropic.com>
1 parent 326f6ef commit 6dae0d7

2 files changed

Lines changed: 2 additions & 2 deletions

File tree

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1 +1 @@
1-
7.14.0a20260613
1+
7.14.0a20260616

.github/workflows/build-rocm-examples-reusable.yml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -164,7 +164,7 @@ jobs:
164164
f.write(f"LIBRARY_PATH={rocm}/lib:{rocm}/lib64\n")
165165
f.write(f"LD_LIBRARY_PATH={core}:{libs}:{rocm}/lib:{rocm}/llvm/lib\n")
166166
f.write("ENABLE_OPENMP=OFF\n") # whl-multi-arch omits amdllvm needed for OpenMP GPU offloading
167-
f.write(f"CXX={rocm}/bin/amdclang++\n") # hipDNN headers require C++20; system g++ on AlmaLinux 8 is too old
167+
f.write(f"CXX={venv}/bin/amdclang++\n") # hipDNN headers require C++20; system g++ on AlmaLinux 8 is too old
168168
rocm_version = subprocess.check_output(["rocm-sdk", "version"], text=True).strip()
169169
with open(github_output, "a") as f:
170170
f.write(f"rocm_version={rocm_version}\n")

0 commit comments

Comments
 (0)