Clean up AMD CI (#6365)

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

@@ -27,32 +27,15 @@ jobs:
 
       - 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
           touch github_summary.md
-          docker pull lmsysorg/sglang:v0.4.6.post3-rocm630
+          bash scripts/amd_ci_start_container.sh
-          docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
+        env:
-            -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            --cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
-            -w /sglang-checkout --name ci_sglang \
-            lmsysorg/sglang:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 ci_sglang pip install huggingface_hub[hf_xet]
 
       - name: Nightly Test
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" ci_sglang python3 run_suite.py --suite nightly-amd --timeout-per-file 7200
+          bash scripts/amd_ci_exec.sh -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" python3 run_suite.py --suite nightly-amd --timeout-per-file 7200
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY

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

@@ -35,37 +35,20 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Evaluate Accuracy
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_eval_accuracy_large.py
+          bash scripts/amd_ci_exec.sh 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
+          bash scripts/amd_ci_exec.sh 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
+          bash scripts/amd_ci_exec.sh python3 models/test_qwen_models.py
 
   accuracy-test-2-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -78,35 +61,18 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Evaluate accuracy (TP=2)
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_moe_eval_accuracy_large.py
+          bash scripts/amd_ci_exec.sh python3 test_moe_eval_accuracy_large.py
 
   mla-test-1-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -119,35 +85,18 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-rocm630
-          docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
-            -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
-            --cap-add=SYS_PTRACE -e HF_TOKEN=${{ secrets.AMD_HF_TOKEN }} --security-opt seccomp=unconfined \
-            -w /sglang-checkout --name ci_sglang \
-            lmsysorg/sglang:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: MLA TEST
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 test_mla.py
+          bash scripts/amd_ci_exec.sh python3 test_mla.py
 
   performance-test-1-gpu-part-1-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -160,56 +109,39 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Benchmark single latency
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1_small
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1_small
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1_default
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_bs1_default
 
       - name: Benchmark online latency
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_online_latency_default
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_online_latency_default
 
       - name: Benchmark offline throughput
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default
 
       - name: Benchmark offline throughput (Non-streaming, small batch size)
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_non_stream_small_batch_size
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_non_stream_small_batch_size
 
       - name: Benchmark online latency (EAGLE)
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_online_latency_eagle
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_online_latency_eagle
 
   performance-test-1-gpu-part-2-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -222,45 +154,28 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Benchmark offline throughput (w/o RadixAttention)
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_without_radix_cache
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_without_radix_cache
 
       - name: Benchmark offline throughput (w/ Triton)
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_with_triton_attention_backend
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_with_triton_attention_backend
 
       - name: Benchmark offline throughput (w/ FP8)
         timeout-minutes: 10
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_AMD_CI=1 -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default_fp8
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_offline_throughput_default_fp8
 
   bench-test-2-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -273,59 +188,38 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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: Benchmark dummy grok (TP=2)
         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
+          bash scripts/amd_ci_exec.sh python3 models/test_dummy_grok_models.py
 
       - name: Benchmark single latency (TP=2)
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_moe_tp2_bs1
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_moe_tp2_bs1
 
       - name: Benchmark single latency + torch.compile (TP=2)
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_torch_compile_tp2_bs1
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_torch_compile_tp2_bs1
 
       - name: Benchmark offline throughput (TP=2)
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_default
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_default
 
       - name: Benchmark offline throughput (w/o RadixAttention) (TP=2)
         timeout-minutes: 20
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_without_radix_cache
+          bash scripts/amd_ci_exec.sh python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_without_radix_cache
 
   unit-test-backend-1-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -338,35 +232,18 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Run test
         timeout-minutes: 30
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 -e SGLANG_AITER_MOE=1 ci_sglang python3 run_suite.py --suite per-commit-amd
+          bash scripts/amd_ci_exec.sh python3 run_suite.py --suite per-commit-amd
 
   unit-test-backend-2-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -379,35 +256,18 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Run test
         timeout-minutes: 30
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 -e SGLANG_AITER_MOE=1 ci_sglang python3 run_suite.py --suite per-commit-2-gpu-amd
+          bash scripts/amd_ci_exec.sh python3 run_suite.py --suite per-commit-2-gpu-amd
 
   unit-test-backend-8-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -420,35 +280,18 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup docker
+      - name: Start CI container
-        run: |
+        run: bash scripts/amd_ci_start_container.sh
-          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+        env:
-          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+          GITHUB_WORKSPACE: ${{ github.workspace }}
-            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-          else
-            DEVICE_FLAG="--device /dev/dri"
-          fi
-          docker pull lmsysorg/sglang:v0.4.6.post3-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:v0.4.6.post3-rocm630
 
       - name: Install dependencies
-        run: |
+        run: bash scripts/amd_ci_install_dependency.sh
-          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 .
 
       - name: Run test
         timeout-minutes: 30
         run: |
-          docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 -e SGLANG_AITER_MOE=1 ci_sglang python3 run_suite.py --suite per-commit-8-gpu-amd
+          bash scripts/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd
 
   finish:
     if: always()

scripts/amd_ci_exec.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+set -euo pipefail
+
+# Default working directory
+WORKDIR="/sglang-checkout/test/srt"
+ENV_ARGS=(
+  -e SGLANG_AMD_CI=1
+  -e SGLANG_IS_IN_CI=1
+  -e SGLANG_AITER_MOE=1
+)
+
+# Parse optional -w/--workdir and -e ENV=VAL flags
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -w|--workdir)
+      WORKDIR="$2"
+      shift 2
+      ;;
+    -e)
+      ENV_ARGS+=("-e" "$2")
+      shift 2
+      ;;
+    --)
+      shift
+      break
+      ;;
+    *)
+      break
+      ;;
+  esac
+done
+
+# Run docker exec
+docker exec \
+  -w "$WORKDIR" \
+  "${ENV_ARGS[@]}" \
+  ci_sglang "$@"

scripts/amd_ci_install_dependency.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+set -euo pipefail
+
+# Install the required dependencies in CI.
+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:/

scripts/amd_ci_start_container.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+set -euo pipefail
+
+# Set up DEVICE_FLAG based on Kubernetes pod info
+if [ -f "/etc/podinfo/gha-render-devices" ]; then
+  DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
+else
+  DEVICE_FLAG="--device /dev/dri"
+fi
+
+# Pull the image
+IMAGE="lmsysorg/sglang:v0.4.6.post3-rocm630"
+echo "Pulling Docker image: $IMAGE"
+docker pull "$IMAGE"
+
+# Run the container
+echo "Starting container: ci_sglang"
+docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
+  -v "${GITHUB_WORKSPACE:-$PWD}:/sglang-checkout" \
+  --ipc=host --group-add video \
+  --cap-add=SYS_PTRACE \
+  -e HF_TOKEN="${HF_TOKEN:-}" \
+  --security-opt seccomp=unconfined \
+  -w /sglang-checkout \
+  --name ci_sglang \
+  "$IMAGE"