[bugfix]: stop red CI checks from stacked PRs, Comfy republish, and GB10 cuDNN dispatch

- Trigger Merge Gate: push and label events on stacked PRs (base is another
  PR branch) hit a hard setFailed on every push. Skip the job for them; a
  /merge call still reaches the base check and fails loudly.
- Publish to Comfy registry: any pyproject.toml edit re-published the same
  version and the registry answered 400 'The node version already exists'.
  Publish only when [project].version changes.
- test_cudnn_sdpa_dense: enabling cuDNN SDPA only makes it eligible; on GB10
  (Spark 4 fastcheck) the dispatcher picks another backend, so the
  bit-identity reference was not cuDNN. Pin the reference to
  SDPBackend.CUDNN_ATTENTION.

.github/workflows/ci-trigger-full-suite.yml

@@ -20,10 +20,14 @@ permissions:
 
 jobs:
   trigger:
+    # Stacked PRs (base is another PR branch) skip here instead of failing on
+    # every push; the merge gate only applies to the default branch. An explicit
+    # `/merge` call still reaches the base check below and fails loudly.
     if: >-
       inputs.pr_number > 0
-      || (github.event.action == 'labeled' && github.event.label.name == 'ready')
-      || github.event.action == 'synchronize'
+      || (github.event.pull_request.base.ref == github.event.repository.default_branch
+          && ((github.event.action == 'labeled' && github.event.label.name == 'ready')
+              || github.event.action == 'synchronize'))
     runs-on: ubuntu-latest
     # Job-level concurrency: only this guarded job acquires the group, so an
     # unrelated `labeled` event (which skips the job) cannot cancel an in-flight

.github/workflows/publish-comfyui.yml

@@ -12,10 +12,43 @@ permissions:
   issues: write
 
 jobs:
-  publish-node:
-    name: Publish Custom Node to registry
+  # Any pyproject.toml edit triggers this workflow, but the registry rejects a
+  # version it already has ("The node version already exists"). Publish only
+  # when [project].version changed, as the PyPI workflow does.
+  check-version-change:
     runs-on: ubuntu-latest
     if: ${{ github.repository_owner == 'hao-ai-lab' }}
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+      - name: Compare [project].version with the previous commit
+        id: check
+        shell: python
+        run: |
+          import os
+          import subprocess
+          import tomllib
+
+          def version(text: str) -> str | None:
+              return tomllib.loads(text).get("project", {}).get("version")
+
+          new = version(open("pyproject.toml", encoding="utf-8").read())
+          previous = subprocess.run(["git", "show", "HEAD~1:pyproject.toml"], capture_output=True, text=True)
+          old = version(previous.stdout) if previous.returncode == 0 else None
+          changed = new is not None and new != old
+          print(f"previous version: {old}, current version: {new}, changed: {changed}")
+          with open(os.environ["GITHUB_OUTPUT"], "a", encoding="utf-8") as output:
+              output.write(f"changed={str(changed).lower()}\n")
+
+  publish-node:
+    name: Publish Custom Node to registry
+    needs: check-version-change
+    runs-on: ubuntu-latest
+    if: ${{ github.repository_owner == 'hao-ai-lab' && (needs.check-version-change.outputs.changed == 'true' || github.event_name == 'workflow_dispatch') }}
     steps:
       - name: Check out code
         uses: actions/checkout@v4
@@ -25,4 +58,4 @@ jobs:
         uses: Comfy-Org/publish-node-action@v1
         with:
           ## Add your own personal access token to your Github Repository secrets and reference it here.
-          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

fastvideo/tests/attention/test_cudnn_sdpa_dense.py

@@ -49,14 +49,11 @@ def test_cudnn_route_presents_bhsd_strided_views():
 def test_cudnn_route_is_bit_identical_to_cudnn_dispatched_sdpa():
     query, key, value = _qkv(2048)
     scale = 1.0 / math.sqrt(query.shape[-1])
-    previous = torch.backends.cuda.cudnn_sdp_enabled()
-    torch.backends.cuda.enable_cudnn_sdp(True)
-    try:
-        with torch.no_grad():
-            reference = torch.nn.functional.scaled_dot_product_attention(
-                query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2), scale=scale).transpose(1, 2)
-    finally:
-        torch.backends.cuda.enable_cudnn_sdp(previous)
+    # Pin the reference to cuDNN: enabling it only makes it eligible, and the dispatcher
+    # prefers another backend on some GPUs (GB10), which is not bit-identical.
+    with torch.no_grad(), torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.CUDNN_ATTENTION):
+        reference = torch.nn.functional.scaled_dot_product_attention(
+            query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2), scale=scale).transpose(1, 2)
 
     with torch.no_grad():
         output = cudnn_sdpa_bshd(query, key, value, scale=scale)

fastvideo/tests/contract/test_ci_test_collection.py

@@ -187,6 +187,10 @@ def test_merge_comment_has_one_change_aware_trigger_path():
     assert "pr.state !== 'open'" in ready_workflow
     assert "pr.base.repo.full_name !== context.payload.repository.full_name" in ready_workflow
     assert "pr.base.ref !== context.payload.repository.default_branch" in ready_workflow
+    # Push/label events on stacked PRs skip the job rather than failing it.
+    trigger_if = yaml.safe_load(ready_workflow)["jobs"]["trigger"]["if"]
+    assert "github.event.pull_request.base.ref == github.event.repository.default_branch" in trigger_if
+    assert trigger_if.startswith("inputs.pr_number > 0 ||")
     assert "core.setOutput('head_sha', pr.head.sha)" in ready_workflow
     assert "core.setOutput('base_sha', pr.base.sha)" in ready_workflow
     assert "api.buildkite.com" in ready_workflow