ci: allow deploy jobs to be cancelled by concurrency groups (#7169)

Signed-off-by: Tushar Sharma <tusharma@nvidia.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.github/workflows/pr.yaml

@@ -253,7 +253,7 @@ jobs:
     runs-on: prod-default-small-v2
     # Run when any deploy test will run: if any framework or deploy files changed
     if: |
-      always() &&
+      !cancelled() &&
       (needs.changed-files.outputs.vllm == 'true' ||
        needs.changed-files.outputs.sglang == 'true' ||
        needs.changed-files.outputs.trtllm == 'true' ||
@@ -301,13 +301,13 @@ jobs:
    # End-to-end tests for each framework with various deployment profiles
    # ============================================================================
   deploy-test-vllm:
-    # Run if core, vllm, or deploy is changed
-    # always() is needed because reusable workflows with skipped internal jobs
-    # (e.g. multi-gpu tests, arm64 copy) cause GitHub Actions to skip dependent
-    # jobs unless always() is present. !failure() ensures we still skip on real failures.
+    # !cancelled() && !failure() is required because reusable workflows with skipped
+    # internal jobs (e.g. multi-gpu tests) propagate non-success through `needs`,
+    # auto-skipping dependents. See: https://github.com/orgs/community/discussions/189172
     if: |
-      always() && !failure() &&
-      (needs.changed-files.outputs.vllm == 'true' || needs.changed-files.outputs.deploy == 'true')
+      !cancelled() && !failure() &&
+      (needs.changed-files.outputs.vllm == 'true' || needs.changed-files.outputs.deploy == 'true') &&
+      (needs.vllm-pipeline.result == 'success')
     runs-on: prod-default-small-v2
     needs: [changed-files, deploy-operator, vllm-pipeline]
     timeout-minutes: 25
@@ -344,10 +344,10 @@ jobs:
 
   deploy-test-sglang:
     runs-on: prod-default-small-v2
-    # Run if core, sglang, or deploy is changed (see deploy-test-vllm for always() rationale)
     if: |
-      always() && !failure() &&
-      (needs.changed-files.outputs.sglang == 'true' || needs.changed-files.outputs.deploy == 'true')
+      !cancelled() && !failure() &&
+      (needs.changed-files.outputs.sglang == 'true' || needs.changed-files.outputs.deploy == 'true') &&
+      (needs.sglang-pipeline.result == 'success')
     needs: [changed-files, deploy-operator, sglang-pipeline]
     timeout-minutes: 25
     permissions:
@@ -381,10 +381,10 @@ jobs:
 
   deploy-test-trtllm:
     runs-on: prod-default-small-v2
-    # Run if core, trtllm, or deploy is changed (see deploy-test-vllm for always() rationale)
     if: |
-      always() && !failure() &&
-      (needs.changed-files.outputs.trtllm == 'true' || needs.changed-files.outputs.deploy == 'true')
+      !cancelled() && !failure() &&
+      (needs.changed-files.outputs.trtllm == 'true' || needs.changed-files.outputs.deploy == 'true') &&
+      (needs.trtllm-pipeline.result == 'success')
     needs: [changed-files, deploy-operator, trtllm-pipeline]
     timeout-minutes: 25
     permissions: