You're an odd one, ain't you?

Signed-off-by: Huy Do <huydhn@gmail.com>

Author: huydhn

.github/scripts/gpt-oss/run_accuracy_checks.sh

@@ -30,22 +30,29 @@ fi
 pushd vllm-benchmarks/gpt-oss
 mkdir -p /tmp/gpqa_openai
 
+# Not sure why this is needed on ROCm image
+if [[ "${DEVICE_NAME}" == "rocm" ]]; then
+  ls -la gpt_oss
+  ls -la gpt_oss/evals
+  export PYTHONPATH=$(pwd):$PYTHONPATH
+fi
+
 # Low
-OPENAI_API_KEY='' python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
+OPENAI_API_KEY="" python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
   --model $MODEL \
   --eval gpqa \
   --reasoning-effort low \
   --n-threads $(expr $(nproc) / 2)
 
 # Mid
-OPENAI_API_KEY='' python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
+OPENAI_API_KEY="" python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
   --model $MODEL \
   --eval gpqa \
   --reasoning-effort medium \
   --n-threads $(expr $(nproc) / 2)
 
 # High
-OPENAI_API_KEY='' python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
+OPENAI_API_KEY="" python3 -m gpt_oss.evals --base-url http://localhost:8000/v1 \
   --model $MODEL \
   --eval gpqa \
   --reasoning-effort high \

.github/workflows/gpt-oss-benchmark.yml

@@ -22,7 +22,7 @@ jobs:
           - runner: linux.dgx.b200.8
             model: openai/gpt-oss-120b
             docker-image: 'public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:6d8d0a24c02bfd84d46b3016b865a44f048ae84b'
-          - runner: linux.rocm.gpu.gfx942.4
+          - runner: linux.rocm.gpu.gfx942.8
             model: openai/gpt-oss-120b
             docker-image: rocm/vllm-dev:open-mi300-08052025
           # gpt-oss-20b