verify_done: wait not synchronize

.github/workflows/pr-gate.yml

@@ -5,6 +5,10 @@ on:
         description: "Whether the PR must have the run-ci label"
         type: boolean
         default: true
+      require-run-ci-extra:
+        description: "Whether the PR must also have the run-ci-extra label (in addition to run-ci)"
+        type: boolean
+        default: false
       cool-down-minutes:
         description: "Cooldown period in minutes for low-permission users; 0 disables rate limiting"
         type: number
@@ -41,6 +45,7 @@ jobs:
           echo "PR Draft: ${{ steps.pr.outputs.draft }}"
           echo "PR User: ${{ steps.pr.outputs.user }}"
           echo "Require run-ci: ${{ inputs.require-run-ci }}"
+          echo "Require run-ci-extra: ${{ inputs.require-run-ci-extra }}"
           echo "Cool down minutes: ${{ inputs.cool-down-minutes }}"
           echo "==================="
 
@@ -59,6 +64,19 @@ jobs:
             exit 1
           fi
 
+      # Live-fetch label gate for pr-test-extra.yml. Workflow-level `if`
+      # checks would read the frozen event payload, which doesn't update on
+      # rerun — so /tag-and-rerun-ci extra (slash handler adds the label
+      # then reruns) could never un-skip. A runtime step using gh-api-fetched
+      # labels lets reruns pick up the new label set.
+      - name: Require run-ci-extra label (optional)
+        if: github.event_name == 'pull_request' && inputs.require-run-ci-extra == true
+        run: |
+          if [[ "${{ contains(fromJson(steps.pr.outputs.labels), 'run-ci-extra') }}" == "false" ]]; then
+            echo "Missing required label 'run-ci-extra'. Add the label (e.g. via /tag-and-rerun-ci extra) to opt this PR into the extra test workflow."
+            exit 1
+          fi
+
       - name: Enforce rate limit for low-permission actors (optional)
         if: github.event_name == 'pull_request' && inputs.cool-down-minutes > 0
         uses: actions/github-script@v7

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

@@ -1,9 +1,13 @@
 name: PR Test Extra
-# Label-gated CI for nightly-class tests opted into a per-PR run.
+# Label-gated extra CI workflow opted into per-PR via labels.
 #
-# Adds runtime to a PR only when the author asks for it: pull_request
-# events bail unless the PR carries the `run-ci-extra` label. The same job
-# graph runs unconditionally on workflow_dispatch / workflow_call so it
+# Adds runtime to a PR only when the author asks for it: the PR must carry
+# BOTH `run-ci` (basic-CI prerequisite) and `run-ci-extra` (explicit opt-in
+# to this workflow). The label check happens at runtime in pr-gate.yml via
+# live `gh pr view --json labels`, so reruns after adding the labels via
+# slash command pick up the new label set (workflow-level `if` checks would
+# read the frozen event payload, which never updates on rerun). The same
+# job graph runs unconditionally on workflow_dispatch / workflow_call so it
 # can be triggered manually or chained from another workflow.
 #
 # Stages: extra-a (1-/2-gpu) and extra-b (4-/8-gpu) caller stubs reuse
@@ -68,27 +72,24 @@ permissions:
 
 jobs:
   # =============================================== check changes ====================================================
-  # Label gate: pull_request events only proceed when the PR carries BOTH
-  # `run-ci` and `run-ci-extra` labels — `run-ci` is the basic-CI prerequisite
-  # (matching pr-test.yml's pr-gate `require-run-ci`) and `run-ci-extra` is the
-  # explicit opt-in to this workflow. Other event types
-  # (workflow_dispatch / workflow_call) always run. When this job is
-  # skipped by the gate, every downstream caller stub naturally skips
-  # because its needs do not resolve.
+  # The actual `run-ci` + `run-ci-extra` label gate now lives in the
+  # `call-gate` job below, which calls pr-gate.yml. pr-gate.yml live-fetches
+  # the PR's current labels at runtime, so a rerun (e.g. via
+  # /tag-and-rerun-ci extra, which adds labels via GITHUB_TOKEN — these
+  # don't cascade-trigger labeled events) picks up the new label set and
+  # passes. The previous workflow-level `if` checked the frozen event
+  # payload, which never updates on rerun and made the slash-add-label flow
+  # impossible to recover.
   #
-  # For `labeled` events we additionally require the just-added label to be
-  # one of the two gating labels — otherwise every unrelated label addition
-  # would dispatch a full CI run.
+  # We only keep one filter here: for `labeled` events, only proceed if the
+  # just-added label is one of the gating labels. This prevents every
+  # unrelated label addition from dispatching a full CI run.
   check-changes:
     if: |
       github.event_name != 'pull_request' ||
-      (
-        (github.event.action != 'labeled' ||
-         github.event.label.name == 'run-ci' ||
-         github.event.label.name == 'run-ci-extra') &&
-        contains(github.event.pull_request.labels.*.name, 'run-ci') &&
-        contains(github.event.pull_request.labels.*.name, 'run-ci-extra')
-      )
+      github.event.action != 'labeled' ||
+      github.event.label.name == 'run-ci' ||
+      github.event.label.name == 'run-ci-extra'
     uses: ./.github/workflows/_pr-test-check-changes.yml
     with:
       git_ref: ${{ inputs.git_ref || '' }}
@@ -97,11 +98,26 @@ jobs:
       pr_test_yml: '.github/workflows/pr-test-extra.yml'
     secrets: inherit
 
+  # =============================================== PR Gate ====================================================
+  # Runtime live-fetch gate. Mirrors pr-test.yml's call-gate, but additionally
+  # requires `run-ci-extra`. Failure here cascades down to every test job
+  # via `needs`, so the workflow ends with one red gate job (~30s) plus a
+  # row of skipped test jobs instead of consuming a runner.
+  call-gate:
+    needs: check-changes
+    if: github.event_name != 'schedule' && needs.check-changes.result == 'success'
+    uses: ./.github/workflows/pr-gate.yml
+    with:
+      require-run-ci: true
+      require-run-ci-extra: true
+    secrets: inherit
+
   # =============================================== sgl-kernel ====================================================
   sgl-kernel-build-wheels:
-    needs: check-changes
+    needs: [check-changes, call-gate]
     if: |
       needs.check-changes.result == 'success' &&
+      needs.call-gate.result == 'success' &&
       needs.check-changes.outputs.sgl_kernel == 'true'
     uses: ./.github/workflows/_pr-test-sgl-kernel-build.yml
     with:
@@ -113,8 +129,8 @@ jobs:
 
   # =============================================== extra-a (1-/2-gpu) ===============================================
   extra-a-test-1-gpu-small:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-a-test-1-gpu-small
@@ -126,8 +142,8 @@ jobs:
     secrets: inherit
 
   extra-a-test-1-gpu-large:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-a-test-1-gpu-large
@@ -140,8 +156,8 @@ jobs:
     secrets: inherit
 
   extra-a-test-2-gpu-large:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-a-test-2-gpu-large
@@ -154,8 +170,8 @@ jobs:
 
   # =============================================== extra-b (4-/8-gpu) ===============================================
   extra-b-test-4-gpu-h100:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-b-test-4-gpu-h100
@@ -167,8 +183,8 @@ jobs:
     secrets: inherit
 
   extra-b-test-4-gpu-b200:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-b-test-4-gpu-b200
@@ -181,8 +197,8 @@ jobs:
     secrets: inherit
 
   extra-b-test-8-gpu-h200:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-b-test-8-gpu-h200
@@ -194,8 +210,8 @@ jobs:
     secrets: inherit
 
   extra-b-test-deepep-8-gpu-h200:
-    needs: [check-changes, sgl-kernel-build-wheels]
-    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' }}
+    needs: [check-changes, call-gate, sgl-kernel-build-wheels]
+    if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && needs.call-gate.result == 'success' }}
     uses: ./.github/workflows/_pr-test-stage.yml
     with:
       self_name: extra-b-test-deepep-8-gpu-h200

python/sglang/srt/managers/schedule_batch.py

@@ -2451,10 +2451,14 @@ def prepare_for_decode(self):
             )
 
     def maybe_wait_verify_done(self):
+        # Use event.wait() (stream-level wait) instead of .synchronize()
+        # (CPU block). Schedule-stream prep ops following this call get
+        # ordered after the forward-stream verify via the wait; CPU is not
+        # blocked. Subsequent .cpu()/.item() naturally sync the stream.
         if self.is_spec_v2:
             draft_input: EagleDraftInput = self.spec_info
             if draft_input.verify_done is not None:
-                draft_input.verify_done.synchronize()
+                draft_input.verify_done.wait()
 
     def filter_batch(
         self,