Debug

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

Author: huydhn

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

@@ -209,30 +209,20 @@ jobs:
             "${DOCKER_IMAGE}"
           )
 
-          # Run perf tests
-          docker exec -t "${container_name}" bash -c "
-            pushd vllm-benchmarks/vllm
-            cp vllm/benchmarks/lib/utils.py /app/vllm-os-mini/vllm/benchmarks/utils.py || true
-            if [[ $DEVICE_NAME != 'rocm' ]]; then
-              pip install -U openai transformers
-              pip install --pre vllm==0.10.1+gptoss \
-                --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
-                --extra-index-url https://download.pytorch.org/whl/nightly/cu128
-            fi
-            pip freeze
-            bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
-            popd
-          "
-
           # Run accuracy check
           docker exec -t "${container_name}" bash -c "
+            set -eux
+
+            echo $MODEL
+
             tp=0
             if [[ $MODEL == 'openai/gpt-oss-120b' ]]; then
               tp=4
             elif [[ $MODEL == 'openai/gpt-oss-20b' ]]; then
               tp=1
             fi
 
+            echo $tp
             # Prepare the accuracy test
             vllm serve $MODEL --tensor_parallel_size $tp &
             server_pid=$!
@@ -280,6 +270,21 @@ jobs:
             kill -9 $server_pid
           "
 
+          # Run perf tests
+          docker exec -t "${container_name}" bash -c "
+            pushd vllm-benchmarks/vllm
+            cp vllm/benchmarks/lib/utils.py /app/vllm-os-mini/vllm/benchmarks/utils.py || true
+            if [[ $DEVICE_NAME != 'rocm' ]]; then
+              pip install -U openai transformers
+              pip install --pre vllm==0.10.1+gptoss \
+                --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
+                --extra-index-url https://download.pytorch.org/whl/nightly/cu128
+            fi
+            pip freeze
+            bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+            popd
+          "
+
       - name: Authenticate with AWS
         # AWS CUDA runners already have access to the bucket via its runner IAM role
         if: env.DEVICE_NAME == 'rocm' || contains(env.DEVICE_TYPE, 'B200')