Skip to content

Commit 0f580e8

Browse files
sbryngelsonclaude
andcommitted
Disable Frontier AMD bench: IBM case hits zero-size GPU transfer bug
The IBM benchmark case crashes with the AMD OpenMP offloading compiler due to a zero-size host-to-device memory copy that the HSA runtime rejects (HSA_STATUS_ERROR_INVALID_ARGUMENT). This works fine with CCE. Disable the AMD bench entry until the underlying issue is fixed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1 parent eed3af5 commit 0f580e8

1 file changed

Lines changed: 9 additions & 8 deletions

File tree

.github/workflows/bench.yml

Lines changed: 9 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -74,14 +74,15 @@ jobs:
7474
device: gpu
7575
interface: omp
7676
build_script: "bash .github/workflows/frontier/build.sh gpu omp bench"
77-
- cluster: frontier_amd
78-
name: Oak Ridge | Frontier (AMD)
79-
group: phoenix
80-
labels: frontier
81-
flag: famd
82-
device: gpu
83-
interface: omp
84-
build_script: "bash .github/workflows/frontier_amd/build.sh gpu omp bench"
77+
# frontier_amd bench disabled: IBM case hits zero-size GPU transfer bug with AMD OpenMP offloading
78+
# - cluster: frontier_amd
79+
# name: Oak Ridge | Frontier (AMD)
80+
# group: phoenix
81+
# labels: frontier
82+
# flag: famd
83+
# device: gpu
84+
# interface: omp
85+
# build_script: "bash .github/workflows/frontier_amd/build.sh gpu omp bench"
8586
runs-on:
8687
group: ${{ matrix.group }}
8788
labels: ${{ matrix.labels }}

0 commit comments

Comments
 (0)