update grok test (#5171)

.github/workflows/pr-test-amd.yml

@@ -56,8 +56,8 @@ jobs:
           docker exec -w /human-eval ci_sglang pip install -e .
 
           docker exec -w / ci_sglang mkdir -p /dummy-grok
-          mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -P dummy-grok
-          docker cp ./dummy-grok ci_sglang:/dummy-grok/
+          mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
+          docker cp ./dummy-grok ci_sglang:/
 
       - name: Evaluate Accuracy
         timeout-minutes: 20
@@ -65,7 +65,6 @@ jobs:
           docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_eval_accuracy_large.py
           docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_eval_fp8_accuracy.py
           docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 models/test_qwen_models.py
-          docker exec -w /sglang-checkout -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m sglang.bench_one_batch --batch-size 32 --input 1024 --output 8 --model /dummy-grok --tokenizer-path Xenova/grok-1-tokenizer --load-format dummy --tp 8 --quantization fp8
 
   mla-test-1-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -105,6 +104,48 @@ jobs:
         run: |
           docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_mla.py
 
+  bench-test-2-gpu-amd:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
+      github.event.pull_request.draft == false
+    runs-on: linux-mi300-gpu-2
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup docker
+        run: |
+          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
+          else
+            DEVICE_FLAG="--device /dev/dri"
+          fi
+          docker pull lmsysorg/sglang:v0.4.5-rocm630
+          docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
+            -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
+            --cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
+            -w /sglang-checkout --name ci_sglang \
+            lmsysorg/sglang:v0.4.5-rocm630
+
+      - name: Install dependencies
+        run: |
+          docker exec ci_sglang pip install --upgrade pip
+          docker exec ci_sglang pip uninstall sgl-kernel -y || true
+          docker exec -w /sglang-checkout/sgl-kernel ci_sglang bash -c "rm -f pyproject.toml && mv pyproject_rocm.toml pyproject.toml && python3 setup_rocm.py install"
+          docker exec ci_sglang pip install -e "python[dev_hip]"
+
+          docker exec -w / ci_sglang git clone https://github.com/merrymercy/human-eval.git
+          docker exec -w /human-eval ci_sglang pip install -e .
+
+          docker exec -w / ci_sglang mkdir -p /dummy-grok
+          mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
+          docker cp ./dummy-grok ci_sglang:/
+
+      - name: Evaluate Benchmark
+        timeout-minutes: 20
+        run: |
+          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 models/test_dummy_grok_models.py
+
   finish:
     if: always()
     needs: [