Fix nan value generated after custom all reduce (#8663)

Co-authored-by: wunhuang <wunhuang@amd.com>

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

@@ -290,6 +290,25 @@ jobs:
         run: |
           bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --timeout-per-file 3600
 
+  unit-test-backend-8-gpu-CAR-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]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Start CI container
+        run: bash scripts/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/amd_ci_install_dependency.sh
+
       - name: Run CustomAllReduce test
         timeout-minutes: 20
         run: |

python/sglang/srt/distributed/device_communicators/custom_all_reduce.py

@@ -398,7 +398,7 @@ class CustomAllreduce:
             else:
                 # If warm up, mimic the allocation pattern since custom
                 # allreduce is out-of-place.
-                return torch.empty_like(input)
+                return torch.zeros_like(input)
         else:
             if _is_hip:
                 # note: outside of cuda graph context,