update unit test in AMD CI (#3366)

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

@@ -56,6 +56,42 @@ jobs:
         run: |
           docker exec -w /sglang-checkout/test/srt ci_sglang python3 test_eval_accuracy_large.py
 
+  mla-test-1-gpu-amd:
+    if: github.repository == 'sgl-project/sglang' && github.event.pull_request.draft == false
+    runs-on: linux-mi300-gpu-1
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - 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:v0.4.2.post2-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:v0.4.2.post2-rocm630
+
+      - name: Install dependencies
+        run: |
+          docker exec ci_sglang pip install --upgrade pip
+          docker exec -w /sglang-checkout/sgl-kernel ci_sglang 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 ci_sglang python3 test_mla.py
+
   finish:
     needs: [
       accuracy-test-1-gpu-amd

test/srt/test_mla.py

@@ -1,6 +1,8 @@
 import unittest
 from types import SimpleNamespace
 
+import torch
+
 from sglang.srt.utils import kill_process_tree
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.run_eval import run_eval
@@ -58,16 +60,14 @@ class TestDeepseekV3(unittest.TestCase):
     def setUpClass(cls):
         cls.model = "lmzheng/sglang-ci-dsv3-test"
         cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = ["--trust-remote-code"]
+        if torch.cuda.is_available() and torch.version.cuda:
+            other_args.extend(["--enable-torch-compile", "--cuda-graph-max-bs", "2"])
         cls.process = popen_launch_server(
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=[
-                "--trust-remote-code",
-                "--enable-torch-compile",
-                "--cuda-graph-max-bs",
-                "2",
-            ],
+            other_args=other_args,
         )
 
     @classmethod