ci: key E2E workflow_dispatch concurrency on head_sha

.github/workflows/ci-e2e.yml

@@ -39,12 +39,17 @@ on:
 # comments get a group of their own (keyed by comment id) so they cancel nothing. The
 # `/retest` predicate is kept identical to the one in `resolve.if`, so exactly the
 # comments that can start a run are the comments that can cancel one.
+#
+# workflow_dispatch has no pull_request.number / issue.number, so the group used
+# to fall through to github.ref (refs/heads/main for every dispatch) and
+# cancel-in-progress killed the previous GPU smoke. Key on inputs.head_sha
+# (then head_ref) so two fork-PR dispatches can run without cancelling each other.
 concurrency:
   group: >-
     ci-e2e-${{
       (github.event_name == 'issue_comment' && !contains(github.event.comment.body, '/retest'))
       && format('noop-comment-{0}', github.event.comment.id)
-      || (github.event.pull_request.number || github.event.issue.number || github.ref)
+      || (github.event.pull_request.number || github.event.issue.number || inputs.head_sha || inputs.head_ref || github.ref)
     }}
   cancel-in-progress: true
 

.github/workflows/forge-e2e.yml

@@ -190,7 +190,7 @@ jobs:
     # Concurrency starts only after resolve has authenticated and parsed the
     # event. A no-op comment therefore cannot cancel an in-flight GPU job.
     concurrency:
-      group: forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_ref || github.ref }}
+      group: forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_sha || needs.resolve.outputs.head_ref || github.ref }}
       cancel-in-progress: true
     runs-on: ubuntu-latest
     steps:
@@ -240,7 +240,7 @@ jobs:
     needs: resolve
     if: needs.resolve.outputs.run == 'true'
     concurrency:
-      group: forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_ref || github.ref }}
+      group: forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_sha || needs.resolve.outputs.head_ref || github.ref }}
       cancel-in-progress: true
     runs-on: ${{ vars.FORGE_E2E_RUNNER_LABEL || 'Hyperloom-e2e-ci' }}
     timeout-minutes: ${{ fromJSON(vars.FORGE_E2E_JOB_TIMEOUT_MIN || '240') }}

scripts/tests/test_forge_e2e_gate.py

@@ -107,7 +107,7 @@ def test_only_resolved_events_can_cancel_an_in_flight_forge_run() -> None:
     assert "github.event.comment.author_association == 'MEMBER'" in jobs
     assert "github.event.comment.author_association == 'COLLABORATOR'" in jobs
 
-    group = "forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_ref || github.ref }}"
+    group = "forge-e2e-${{ needs.resolve.outputs.pr_number || needs.resolve.outputs.head_sha || needs.resolve.outputs.head_ref || github.ref }}"
     assert workflow.count(group) == 2
     assert workflow.count("cancel-in-progress: true") == 2
 

src/hyperloom/inference_optimizer/tests/test_ci_e2e_concurrency_guard.py

@@ -91,6 +91,11 @@ def test_every_trigger_still_resolves_to_a_group(concurrency_group: str) -> None
     for key in (
         "github.event.pull_request.number",  # pull_request
         "github.event.issue.number",  # issue_comment (/retest)
-        "github.ref",  # workflow_dispatch
+        "inputs.head_sha",  # workflow_dispatch (fork PR smoke)
+        "inputs.head_ref",
+        "github.ref",  # last-resort fallback
     ):
         assert key in concurrency_group
+    # Dispatch used to share refs/heads/main and cancel-in-progress the
+    # previous GPU run. head_sha must win over github.ref.
+    assert concurrency_group.index("inputs.head_sha") < concurrency_group.index("github.ref")