Enable unit tests for AMD CI. (#6283)

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

@@ -325,11 +325,135 @@ jobs:
         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
 
+  unit-test-backend-1-gpu-amd:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
+      github.event.pull_request.draft == false
+    strategy:
+      matrix:
+        runner: [linux-mi300-gpu-1, linux-mi325-gpu-1]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup docker
+        run: |
+          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
+          else
+            DEVICE_FLAG="--device /dev/dri"
+          fi
+          docker pull lmsysorg/sglang:v0.4.6.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: |
+          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
+
+  unit-test-backend-2-gpu-amd:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
+      github.event.pull_request.draft == false
+    strategy:
+      matrix:
+        runner: [linux-mi300-gpu-2, linux-mi325-gpu-2]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup docker
+        run: |
+          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
+          else
+            DEVICE_FLAG="--device /dev/dri"
+          fi
+          docker pull lmsysorg/sglang:v0.4.6.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: |
+          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
+
+  unit-test-backend-8-gpu-amd:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
+      github.event.pull_request.draft == false
+    strategy:
+      matrix:
+        runner: [linux-mi300-gpu-8, linux-mi325-gpu-8]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup docker
+        run: |
+          # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
+          if [ -f "/etc/podinfo/gha-render-devices" ]; then
+            DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
+          else
+            DEVICE_FLAG="--device /dev/dri"
+          fi
+          docker pull lmsysorg/sglang:v0.4.6.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: |
+          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
+
   finish:
     if: always()
     needs: [
       accuracy-test-1-gpu-amd, mla-test-1-gpu-amd, bench-test-2-gpu-amd,
-      accuracy-test-2-gpu-amd, performance-test-1-gpu-part-1-amd, performance-test-1-gpu-part-2-amd
+      accuracy-test-2-gpu-amd, performance-test-1-gpu-part-1-amd, performance-test-1-gpu-part-2-amd,
+      unit-test-backend-1-gpu-amd, unit-test-backend-2-gpu-amd, unit-test-backend-8-gpu-amd
     ]
     runs-on: ubuntu-latest
     steps:

test/srt/run_suite.py

@@ -85,6 +85,12 @@ suites = {
         TestFile("test_w8a8_quantization.py", 46),
         TestFile("models/lora/test_lora_cuda_graph.py", 250),
     ],
+    "per-commit-amd": [
+        TestFile("test_mla.py", 242),
+        TestFile("test_mla_deepseek_v3.py", 221),
+        TestFile("test_torch_compile.py", 76),
+        TestFile("test_torch_compile_moe.py", 172),
+    ],
     "per-commit-2-gpu": [
         TestFile("models/lora/test_lora_tp.py", 116),
         TestFile("test_data_parallelism.py", 73),
@@ -95,6 +101,9 @@ suites = {
         TestFile("test_update_weights_from_distributed.py", 103),
         TestFile("test_verl_engine.py", 64),
     ],
+    "per-commit-2-gpu-amd": [
+        TestFile("test_mla_tp.py", 170),
+    ],
     "per-commit-8-gpu": [
         # Disabled deepep tests temporarily because it takes too much time.
         # TODO: re-enable them after reducing the test time with compilation cache and smaller models.
@@ -107,6 +116,9 @@ suites = {
         TestFile("test_full_deepseek_v3.py", 250),
         TestFile("test_pp_single_node.py", 150),
     ],
+    "per-commit-8-gpu-amd": [
+        TestFile("test_full_deepseek_v3.py", 250),
+    ],
     "nightly": [
         TestFile("test_nightly_gsm8k_eval.py"),
     ],

test/srt/test_full_deepseek_v3.py

+import os
 import unittest
 from types import SimpleNamespace
 
@@ -66,7 +67,10 @@ class TestDeepseekV3(CustomTestCase):
             write_github_step_summary(
                 f"### test_bs_1_speed (deepseek-v3)\n" f"{speed=:.2f} token/s\n"
             )
-            self.assertGreater(speed, 75)
+            if os.getenv("SGLANG_AMD_CI") == "1":
+                self.assertGreater(speed, 12)
+            else:
+                self.assertGreater(speed, 75)
 
 
 class TestDeepseekV3MTP(CustomTestCase):
@@ -144,8 +148,14 @@ class TestDeepseekV3MTP(CustomTestCase):
                 f"{acc_length=:.2f}\n"
                 f"{speed=:.2f} token/s\n"
             )
-            self.assertGreater(acc_length, 2.9)
-            self.assertGreater(speed, 105)
+            if os.getenv("SGLANG_AMD_CI") == "1":
+                self.assertGreater(acc_length, 2.8)
+            else:
+                self.assertGreater(acc_length, 2.9)
+            if os.getenv("SGLANG_AMD_CI") == "1":
+                self.assertGreater(speed, 15)
+            else:
+                self.assertGreater(speed, 105)
 
 
 if __name__ == "__main__":