Fix rerun stage command with merged commit history (#17960)

.github/workflows/pr-test.yml

@@ -72,10 +72,16 @@ jobs:
   check-changes:
     runs-on: ubuntu-latest
     outputs:
-      main_package: ${{ steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }}
+      # Use API-based detection for target_stage mode (filter-api), otherwise use dorny/paths-filter (filter)
-      sgl_kernel: ${{ steps.filter.outputs.sgl_kernel }} # sgl-kernel tests only run when kernels are rebuilt
+      main_package: ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }}
-      jit_kernel: ${{ steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }}
+      # sgl_kernel is forced to false when target_stage is set, since sgl-kernel-build-wheels won't run
-      multimodal_gen: ${{ steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }}
+      # This prevents CUSTOM_BUILD_SGL_KERNEL=true when the wheel artifacts aren't available
+      # Note: If PR has kernel changes AND target_stage is set, the validate-target-stage step will fail
+      sgl_kernel: ${{ !inputs.target_stage && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }}
+      # Raw sgl_kernel value before target_stage override (used for validation)
+      sgl_kernel_raw: ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }}
+      jit_kernel: ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }}
+      multimodal_gen: ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }}
       max_parallel: ${{ steps.set-parallel.outputs.max_parallel }}
       b200_runner: ${{ steps.set-runner.outputs.b200_runner }}
       enable_retry: ${{ steps.set-retry.outputs.enable_retry }}
@@ -102,7 +108,9 @@ jobs:
       - name: Detect file changes
         id: filter
         uses: dorny/paths-filter@v3
-        if: steps.run-mode.outputs.run_all_tests != 'true'
+        # Only use paths-filter for pull_request events (where it works correctly)
+        # For workflow_dispatch with target_stage, we use GitHub API in the next step
+        if: steps.run-mode.outputs.run_all_tests != 'true' && !inputs.target_stage
         with:
           filters: |
             main_package:
@@ -124,6 +132,70 @@ jobs:
               - "python/pyproject.toml"
               - ".github/workflows/pr-test.yml"
 
+      # For /rerun-stage (workflow_dispatch with target_stage), dorny/paths-filter doesn't work
+      # correctly because it falls back to "last commit" detection which breaks for merge commits.
+      # Instead, we use the GitHub API to compare the PR commit against main.
+      - name: Detect file changes via API (for target_stage)
+        id: filter-api
+        if: inputs.target_stage && inputs.pr_head_sha
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          echo "Detecting file changes via GitHub API for target_stage mode..."
+          echo "PR head SHA: ${{ inputs.pr_head_sha }}"
+
+          # Get the list of changed files by comparing PR commit against main
+          # This correctly handles merge commits by looking at the actual PR diff
+          CHANGED_FILES=$(gh api "repos/${{ github.repository }}/compare/main...${{ inputs.pr_head_sha }}" \
+            --jq '[.files[].filename] | .[]' 2>/dev/null || echo "")
+
+          if [ -z "$CHANGED_FILES" ]; then
+            echo "Warning: Could not fetch changed files from API, assuming no changes"
+            echo "sgl_kernel=false" >> $GITHUB_OUTPUT
+            echo "main_package=false" >> $GITHUB_OUTPUT
+            echo "jit_kernel=false" >> $GITHUB_OUTPUT
+            echo "multimodal_gen=false" >> $GITHUB_OUTPUT
+            exit 0
+          fi
+
+          echo "Changed files:"
+          echo "$CHANGED_FILES" | head -20
+          echo "..."
+
+          # Check for sgl-kernel changes
+          if echo "$CHANGED_FILES" | grep -q "^sgl-kernel/"; then
+            echo "sgl_kernel=true" >> $GITHUB_OUTPUT
+            echo "Detected sgl-kernel changes"
+          else
+            echo "sgl_kernel=false" >> $GITHUB_OUTPUT
+          fi
+
+          # Check for main_package changes (excluding multimodal_gen)
+          # Note: Need to filter out multimodal_gen before checking, not pipe grep -q output
+          MAIN_PKG_FILES=$(echo "$CHANGED_FILES" | grep -E "^(python/sglang/|python/pyproject\.toml|scripts/ci/cuda/|scripts/ci/utils/|test/|\.github/workflows/pr-test\.yml)" | grep -v "^python/sglang/multimodal_gen/" || true)
+          if [ -n "$MAIN_PKG_FILES" ]; then
+            echo "main_package=true" >> $GITHUB_OUTPUT
+            echo "Detected main_package changes"
+          else
+            echo "main_package=false" >> $GITHUB_OUTPUT
+          fi
+
+          # Check for jit_kernel changes
+          if echo "$CHANGED_FILES" | grep -qE "^(python/sglang/jit_kernel/|python/pyproject\.toml|\.github/workflows/pr-test\.yml)"; then
+            echo "jit_kernel=true" >> $GITHUB_OUTPUT
+            echo "Detected jit_kernel changes"
+          else
+            echo "jit_kernel=false" >> $GITHUB_OUTPUT
+          fi
+
+          # Check for multimodal_gen changes
+          if echo "$CHANGED_FILES" | grep -qE "^(python/sglang/multimodal_gen/|python/sglang/cli/|python/pyproject\.toml|\.github/workflows/pr-test\.yml)"; then
+            echo "multimodal_gen=true" >> $GITHUB_OUTPUT
+            echo "Detected multimodal_gen changes"
+          else
+            echo "multimodal_gen=false" >> $GITHUB_OUTPUT
+          fi
+
       - name: Set max-parallel based on run type
         id: set-parallel
         run: |
@@ -142,8 +214,12 @@ jobs:
       - name: Set B200 runner tag
         id: set-runner
         run: |
-          sgl_kernel="${{ steps.filter.outputs.sgl_kernel || steps.run-mode.outputs.run_all_tests }}"
+          # Use kernel-build runner only when sgl_kernel changes are detected AND we're not in target_stage mode
-          if [[ "$sgl_kernel" == "true" ]]; then
+          # (target_stage skips wheel builds, so we can't use custom kernels)
+          # Use API-based detection (filter-api) for target_stage mode, otherwise use dorny/paths-filter (filter)
+          sgl_kernel="${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel || steps.run-mode.outputs.run_all_tests }}"
+          target_stage="${{ inputs.target_stage }}"
+          if [[ "$sgl_kernel" == "true" && -z "$target_stage" ]]; then
             echo "b200_runner=4-gpu-b200-kernel" >> $GITHUB_OUTPUT
           else
             echo "b200_runner=4-gpu-b200" >> $GITHUB_OUTPUT
@@ -166,6 +242,26 @@ jobs:
             echo "Filtered run, continue-on-error disabled"
           fi
 
+      - name: Validate target_stage with kernel changes
+        # Use API-based detection (filter-api) for target_stage mode, otherwise use dorny/paths-filter (filter)
+        if: inputs.target_stage && (steps.filter-api.outputs.sgl_kernel == 'true' || steps.filter.outputs.sgl_kernel == 'true')
+        run: |
+          echo "::error::Cannot use /rerun-stage when PR has sgl-kernel changes."
+          echo "::error::The sgl-kernel-build-wheels job is skipped in target_stage mode, but this PR modifies sgl-kernel/ files."
+          echo "::error::Please use /tag-and-rerun-ci to run the full workflow including kernel builds."
+          echo ""
+          echo "ERROR: Cannot use /rerun-stage when PR has sgl-kernel changes."
+          echo ""
+          echo "This PR modifies files in sgl-kernel/, which requires building custom kernel wheels."
+          echo "The /rerun-stage command skips the wheel build job, so the test would run against"
+          echo "the wrong (PyPI) version of sgl-kernel instead of your changes."
+          echo ""
+          echo "To properly test your kernel changes, use one of these commands instead:"
+          echo "  /tag-and-rerun-ci           - Re-run the full workflow including kernel builds"
+          echo "  /rerun-ci                   - Re-run the full workflow"
+          echo ""
+          exit 1
+
       - name: Show filter results in summary (table)
         run: |
           {
@@ -173,10 +269,13 @@ jobs:
             echo ""
             echo "| Component         | Changed |"
             echo "|-------------------|---------|"
-            echo "| main_package      | ${{ steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| main_package      | ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }} |"
-            echo "| sgl_kernel        | ${{ steps.filter.outputs.sgl_kernel || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| sgl_kernel (raw)  | ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }} |"
-            echo "| jit_kernel        | ${{ steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| sgl_kernel (used) | ${{ !inputs.target_stage && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }} |"
-            echo "| multimodal_gen    | ${{ steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| jit_kernel        | ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| multimodal_gen    | ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| target_stage      | ${{ inputs.target_stage || '(none)' }} |"
+            echo "| detection_method  | ${{ inputs.target_stage && 'GitHub API' || 'dorny/paths-filter' }} |"
             echo "| max_parallel      | ${{ steps.set-parallel.outputs.max_parallel }} |"
             echo "| b200_runner       | ${{ steps.set-runner.outputs.b200_runner }} |"
             echo "| enable_retry      | ${{ steps.set-retry.outputs.enable_retry }} |"