[CI] Isolate CUDA coredump dir per run to fix tracker mis-attribution (#27344)

.github/actions/upload-cuda-coredumps/action.yml

@@ -29,7 +29,23 @@ runs:
       id: check
       shell: bash
       run: |
-        dir="${SGLANG_CUDA_COREDUMP_DIR:-/tmp/sglang_cuda_coredumps}"
+        # Mirror get_dump_dir() in python/sglang/srt/debug_utils/cuda_coredump.py:
+        # explicit override > per-job RUNNER_TEMP > /tmp default, then a
+        # per-(run, attempt) subdir so dumps are never mis-attributed across CI
+        # jobs that share a self-hosted runner's filesystem.
+        if [ -n "$SGLANG_CUDA_COREDUMP_DIR" ]; then
+          base="$SGLANG_CUDA_COREDUMP_DIR"
+        elif [ -n "$RUNNER_TEMP" ]; then
+          base="$RUNNER_TEMP/sglang_cuda_coredumps"
+        else
+          base="/tmp/sglang_cuda_coredumps"
+        fi
+        if [ -n "$GITHUB_RUN_ID" ]; then
+          dir="$base/${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT:-1}"
+        else
+          dir="$base"
+        fi
+        echo "dump_dir=$dir" >> "$GITHUB_OUTPUT"
         if [ -d "$dir" ] && [ -n "$(ls -A "$dir" 2>/dev/null)" ]; then
           echo "has_dumps=true" >> "$GITHUB_OUTPUT"
         else
@@ -41,7 +57,7 @@ runs:
       uses: actions/upload-artifact@v4
       with:
         name: cuda-coredumps-${{ github.job }}${{ inputs.artifact-suffix && format('-{0}', inputs.artifact-suffix) }}
-        path: ${{ env.SGLANG_CUDA_COREDUMP_DIR || '/tmp/sglang_cuda_coredumps' }}/
+        path: ${{ steps.check.outputs.dump_dir }}/
         retention-days: ${{ inputs.retention-days }}
 
     - name: Signal coredump to tracker issue
@@ -82,4 +98,4 @@ runs:
 
     - name: Cleanup CUDA coredumps
       shell: bash
-      run: rm -rf "${{ env.SGLANG_CUDA_COREDUMP_DIR || '/tmp/sglang_cuda_coredumps' }}"
+      run: rm -rf "${{ steps.check.outputs.dump_dir }}"

python/sglang/srt/debug_utils/cuda_coredump.py

@@ -29,7 +29,24 @@ def is_enabled() -> bool:
 
 
 def get_dump_dir() -> str:
-    return envs.SGLANG_CUDA_COREDUMP_DIR.get()
+    # Resolve the base dir the same way as the uploader
+    # (.github/actions/upload-cuda-coredumps/action.yml) so they agree; an empty
+    # SGLANG_CUDA_COREDUMP_DIR counts as unset, like the action's `[ -n ... ]`.
+    explicit = envs.SGLANG_CUDA_COREDUMP_DIR.get()
+    runner_temp = os.getenv("RUNNER_TEMP")
+    if explicit:
+        base = explicit
+    elif runner_temp:
+        base = os.path.join(runner_temp, "sglang_cuda_coredumps")
+    else:
+        base = "/tmp/sglang_cuda_coredumps"
+    # Isolate dumps per (run, attempt): on a shared self-hosted runner a leftover
+    # dump from one job must not be picked up and mis-attributed by a later one.
+    run_id = os.getenv("GITHUB_RUN_ID")
+    if run_id:
+        attempt = os.getenv("GITHUB_RUN_ATTEMPT", "1")
+        return os.path.join(base, f"{run_id}-{attempt}")
+    return base
 
 
 def _inject_env():

python/sglang/srt/environ.py

@@ -221,7 +221,9 @@ class Envs:
     SGLANG_IS_IN_CI = EnvBool(False)
     SGLANG_IS_IN_CI_AMD = EnvBool(False)
     SGLANG_CUDA_COREDUMP = EnvBool(False)
-    SGLANG_CUDA_COREDUMP_DIR = EnvStr("/tmp/sglang_cuda_coredumps")
+    # None = unset, letting get_dump_dir() resolve the base (RUNNER_TEMP in CI,
+    # else /tmp); see debug_utils/cuda_coredump.py.
+    SGLANG_CUDA_COREDUMP_DIR = EnvStr(None)
     SGLANG_TEST_MAX_RETRY = EnvInt(None)
 
     # Constrained Decoding (Grammar)