Move the logic to a script

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

Author: huydhn

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

@@ -0,0 +1,57 @@
+#!/bin/bash
+
+set -eux
+
+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 $tp &
+server_pid=$!
+
+wait_for_server() {
+  timeout 1200 bash -c '
+    until curl -X POST localhost:8000/v1/completions; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+if wait_for_server; then
+  echo "vLLM server is up and running"
+else
+  echo "vLLM failed to start within the timeout period"
+fi
+
+pushd vllm-benchmarks/gpt-oss
+mkdir -p /tmp/gpqa_openai
+
+# Low
+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 \
+  --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 \
+  --model $MODEL \
+  --eval gpqa \
+  --reasoning-effort high \
+  --n-threads $(expr $(nproc) / 2)
+
+mv /tmp/gpqa_openai .
+popd
+
+kill -9 $server_pid

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

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+set -eux
+
+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

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

@@ -210,80 +210,10 @@ jobs:
           )
 
           # 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 $tp &
-            server_pid=$!
-
-            wait_for_server() {
-              timeout 1200 bash -c '
-                until curl -X POST localhost:8000/v1/completions; do
-                  sleep 1
-                done' && return 0 || return 1
-            }
-
-            if wait_for_server; then
-              echo 'vLLM server is up and running'
-            else
-              echo 'vLLM failed to start within the timeout period'
-            fi
-
-            pushd vllm-benchmarks/gpt-oss
-            mkdir -p /tmp/gpqa_openai
-
-            # Low
-            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 \
-              --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 \
-              --model $MODEL \
-              --eval gpqa \
-              --reasoning-effort high \
-              --n-threads $(expr $(nproc) / 2)
-
-            mv /tmp/gpqa_openai .
-            popd
-
-            kill -9 $server_pid
-          "
+          docker exec -t "${container_name}" bash .github/scripts/gpt-oss/run_accuracy_checks.sh
 
           # 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 [[ $DEVICE_NAME != 'rocm' ]]; then
-              pip install -U openai transformers
-              pip install --pre vllm==0.10.1+gptoss \
-                --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
-                --extra-index-url https://download.pytorch.org/whl/nightly/cu128
-            fi
-            pip freeze
-            bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
-            popd
-          "
+          docker exec -t "${container_name}" bash .github/scripts/gpt-oss/run_benchmarks.sh
 
       - name: Authenticate with AWS
         # AWS CUDA runners already have access to the bucket via its runner IAM role