[zhao-stack]

What this PR does / why we need it?

Advance the verified vLLM main dependency from ba15685ab7f4f5fbed8630bbb7441f8837e76fbe (Nov 6) to 3b95523ed904cf02a26b872f4a900ef17e6be8ad (Nov 12, 2026; 435 upstream commits). The Nov 10 checkpoint feb4cb019 is retained as an intermediate milestone; the newly added interval contains 149 commits. The release dependency remains v0.33.0.

Carry the unmerged #21005 commits and the already-reviewed official Ascend synchronization history without rewriting them. #21005 is unchanged. Official test-contract changes accepted in the earlier nine-commit and twelve-commit synchronization batches remain separately inventoried; this follow-up does not introduce another exception. The original active DSpark test, assertions, thresholds, and external-DP timeout remain preserved.

The inherited compatibility work covers compiler/config hashing (#65652); MRv1 removal, Mamba construction and MTP/310P interfaces (#63373); scheduler iteration details (#61424); Mooncake imports (#65877); PP auxiliary-state relay (#66177); DSpark sampler bias (#64476); Kimi configuration and weight-load accounting; shared-expert scaling; target-weight replacement callbacks (#66287); EPLB warmup (#66660); and the SFA constructor/cached DeepSeek decoder factory.

New Nov 12 compatibility fix: vLLM #66859, c9c36486c3f7c767c25bb37f087e38d101edf1ef, replaces routed-scale booleans with RoutedScaleMode. The previous Ascend runner passed the removed divide_shared_output_in_fp16 keyword and the factory forwarded the removed apply_routed_scale_to_output keyword. Before-edit execution of the actual native constructors passes release/Nov10 and fails Nov12 for both Ascend runners. The existing factory patch now translates Ascend-owned legacy callers and preserves native enum callers before router creation, so the scale is applied exactly once. Both Ascend and 310P runners forward the native mode. Explicit release branching preserves the v0.33.0 factory and native FP16 shared-output behavior. Patch lifecycle/removal guidance is included.

Does this PR introduce any user-facing change?

Yes: dependency alignment and the compatibility fixes above, documented in release_notes/fragments/21469.yaml.

The new max_async_kv_loads_per_step option is explicitly rejected with copied recompute, DyntraLB, profiling-chunk and balance schedulers: actual native enqueue evidence shows an unconsumed queue. Default/unset configurations and inherited native scheduling remain supported. Explicitly requesting the retired drafter-padding option on main also fails before model loading; release retains its existing options.

How was this patch tested?

Current published head: c67e962e87cdc1282a2cc0f1ff196a6198049b45. Draft; full dual-version CPU/NPU validation is pending.

Current E2E 43173866142 is terminal: pre-commit stopped before hooks/mypy at runner add-matcher, which could not find /home/runner/_work/_temp/mypy.json after the container copy step. CPU/NPU tests were skipped. Complete paginated audit: seven runs, 19 checks, two successful combined statuses, 18 all-attempt jobs. The synthetic checkout is 5db69d7d6398d96f5e366de7476b7b838b4c996e (c67e into b27049d); the log resolves test/lint pin to 3b95523e, but package/runtime identity is unvalidated. Docs and both Read the Docs builds pass. Infrastructure retry is recorded and deferred until at least Nov 12 09:16:29 UTC, followed by a fresh head/full-terminal audit; none requested. This failure does not close the historical numerical/startup/sleep issues.

The previous E2E 43169877136 is now terminal: pre-commit hooks and full mypy (Python 3.10/3.11/3.12) passed. Both CPU lanes and test selection failed before tests at the same historical 41d7f3f5c rebase conflict in Mamba source/test; NPU jobs were skipped. The later container-wrapper message is secondary. The failed runtime log does not print its fetched main SHA.

Local replay against official Ascend b27049dc118ccc07b55cd048fdc70f4690b932c8 reproduces the exact commit and both conflicting files. A normal signed merge incorporates six official commits after a4971912c, preserving the existing history and enabling the existing source-snapshot ancestor guard. No workflow/policy changes; both Mamba files, verified vLLM pin and existing DSpark tests remain unchanged. The new vLLM target remains fixed at 3b95523e.

Synchronization validation: 14 Python files parse on 3.10/3.11/3.12; all existing test functions, decorators and assertions in seven affected test files are preserved. Whole-tree Ruff lint/format passes 2,154 files; diff checks pass (official documentation trailing whitespace was removed). The ViT setup test passes, including six subtests. Local RFork test collection is blocked by absent regex/safetensors; the scoped three-version mypy check is blocked by absent safetensors. No dependency, ignore or environment workaround was added. These are not passing RFork/full mypy results for this head. The official RFork DSpark ownership fix is included, but it does not close the Qwen DSpark numerical failures, MiniMax initialization or historical sleep failures.

All previous-head runs were terminal before this necessary source synchronization push. No manual rerun or duplicate workflow trigger.

Local checks for the unchanged MoE follow-up introduced in e90503024:

• Whole-tree Ruff lint/format: 2,150 Python files pass; git diff --check passes.
• Python 3.10/3.11/3.12 syntax checks pass for all six changed Python files. AST preservation audit confirms every original test function, decorator, assertion and threshold in both edited test files remains; new assertions/tests are additional.
• reproduce_moe_1112.py: actual native release/Nov10 constructors pass; Nov12 rejects the removed keyword in both Ascend/310P runners before editing.
• verify_moe_1112.py: actual native release/Nov12 constructors and inherited FP16/BF16 scaling matrices pass for both runners, including shared/no-shared outputs and all native modes.
• run_runner_test_bodies.py: 40 original/new published test-body cases pass across both versions using actual inherited native scaling, with scoped imports and parent initialization fixtures.
• verify_factory_1112.py: 14 published factory test-body cases pass, and normalized kwargs bind to the actual release/Nov12 native factory signatures for all four legacy flag combinations in each lane. Explicit native modes are retained.
• Three-version focused mypy passes the three changed production files plus the factory test. Adding the large runner test file produces three method-assignment diagnostics per version; the byte-identical parent test produces the same diagnostics under the same grouped local check. No ignores, assertion edits or dependency changes were added. This is not a full mypy pass.
• Required bash format.sh ci was attempted with a 45-second bound; hook-environment setup did not finish. The attempt/log is retained and is not a pass. Full CI remains required.
• These local CPU/source boundary checks do not establish full package initialization, device kernels, graph execution or NPU correctness.

The current c67e962e full fixed-interval scan (Nov6→Nov12) completed: 1,217 relations / 9,411 direct-call dependencies / 227 findings / 190 unresolved analyses. Raw reports are retained. Its three raw actionable entries concern two Ascend model calls to the patched MoE factory and the inherited local Score extension. The MoE translation is checked against actual native signatures above; monkey-patch behavior is outside this scanner's analyzed capabilities. Optional/positional override warnings and unresolved analyses are not counted as validated. The local Score method does not prove native MRv2 worker metadata transport.

Parent 7cdd7acbd results are historical, not current-head results. Complete paginated/all-attempt audit: six workflows, 64 checks, two combined statuses and 63 jobs; all workflows terminal. Both CPU jobs succeeded. All 42 NPU partitions terminated: main 19 success / 2 failure, release 20 success / 1 failure. Per-file assertion/skip and runtime/image provenance audits remain incomplete.

The three parent hardware failures have actual first-cause logs:

• Release Qwen3.6 DSpark graph replay: logprob difference 0.1472 > unchanged 0.1378.
• Main 35B DSpark acceptance: 3.575 vs 3.930, 9.03% > unchanged 8%.
• Main MiniMax TP4 dummy-load startup: KeyError: model.language_model.model.layers.0.block_sparse_moe.shared_experts.gate_proj.weight in ModelSlim quantization lookup. This has not been attributed to the pinned native interval or repaired.

Historical main sleep(level1) HTTP timeouts also remain open pending current-head per-file evidence and causal investigation. A previous lane becoming green does not by itself establish root-cause closure. No same-machine, same-configuration, version-only A/B conclusion is claimed for the numerical failures; none is dismissed as noise or infrastructure. Tests and tolerances remain active and unchanged.

Parent pre-commit checked a synthetic merge tree identical to 7cdd and passed hooks/full mypy on all three Python versions. Its selector selected 153 files / 21 groups per version. Those static/selection results and earlier seven sampler/identity hardware passes are historical. Skips, unreached files, cache hits, or one device's identity do not establish full coverage, a fresh image build or all hardware-family validation.

Completion requires current-head actual dual-version CPU/NPU execution, runtime/package/tag/import-branch and image provenance, per-file execution/skip review, and closure or causal exclusion of upgrade regressions. Keep Draft; do not merge automatically. The bot's moving-main line below is preserved and is not the verified pin.

Checklist

• Boundary / edge-case regression coverage

• Preserve existing assertions and thresholds

• Current-head full dual-version CPU/NPU validation

• Runtime/image provenance and executed/skip coverage audit

• Numerical/startup/sleep failure causality and closure

• vLLM main: vllm-project/vllm@24420b7

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

Suggested PR Title:\n\nmarkdown\n[Test][Misc] Update compatibility adapters and tests for vLLM main and v0.33.0 release support\n\n\nSuggested PR Summary:\n\nmarkdown\n### What this PR does / why we need it?\nThis pull request updates various compatibility adapters, model runners, and tests to ensure seamless support for both the vLLM main branch and the v0.33.0 release. Key changes include updating the verified commit, adding release notes, and adapting schedulers, model runners, and speculators to handle version-specific API differences (such as `vllm_version_is` checks). Additionally, several critical issues were identified during the review: missing imports of `vllm_version_is` in multiple files (`hccl_engine.py`, `patch_pp_mtp.py`, `patch_model_runner.py`, `model_runner.py`, and `speculator.py`) which would cause runtime `NameError`s; a potential `AttributeError` in `worker.py` when accessing `resolved_gpu_memory_utilization` on older vLLM versions; and potential `AttributeError`s in `scheduler_output.py` due to missing defensive checks on `observability_config` and `profiler_config`.\n\n### Does this PR introduce _any_ user-facing change?\nNo user-facing changes are introduced; these are internal compatibility and testing updates.\n\n### How was this patch tested?\nTested via existing and newly added unit tests under `tests/ut/`.\n

[gemini-code-assist]

The function vllm_version_is is used in receive_weights (line 158) but is not imported in this file, which will result in a NameError at runtime. We should import vllm_version_is from vllm_ascend.utils.

Suggested change

	from contextlib import nullcontext, suppress
	from contextlib import nullcontext, suppress
	from vllm_ascend.utils import vllm_version_is

[gemini-code-assist]

The function vllm_version_is is used here but is not imported in this file, which will raise a NameError at runtime. We should import it locally.

Suggested change

	if vllm_version_is("0.33.0"):
	_patch_scheduler_make_cached_request_data()
	from vllm_ascend.utils import vllm_version_is
	if vllm_version_is("0.33.0"):
	_patch_scheduler_make_cached_request_data()

[gemini-code-assist]

The function vllm_version_is is used here but is not imported in this file, which will raise a NameError at runtime. We should import it locally.

Suggested change

	if vllm_version_is("0.33.0"):
	breakable_cg = upstream.envs.VLLM_USE_BREAKABLE_CUDAGRAPH
	graph_kwargs = {"decode_query_len": self.decode_query_len}
	from vllm_ascend.utils import vllm_version_is
	if vllm_version_is("0.33.0"):
	breakable_cg = upstream.envs.VLLM_USE_BREAKABLE_CUDAGRAPH
	graph_kwargs = {"decode_query_len": self.decode_query_len}

[gemini-code-assist]

The function vllm_version_is is used in this file (lines 204 and 1613) but is not imported, which will result in a NameError at runtime. We should import vllm_version_is from vllm_ascend.utils.

Suggested change

	from collections.abc import Mapping
	from collections.abc import Mapping
	from vllm_ascend.utils import vllm_version_is

[gemini-code-assist]

The function vllm_version_is is used here but is not imported in this file, which will raise a NameError at runtime. We should import it locally.

Suggested change

	if step == 0 and vllm_version_is("0.33.0"):
	# #60989 includes the correction token before calling this builder
	# on main. Only release still needs the extra prefill token here.
	# Keep the input batch's bounds unchanged.
	draft_seq_lens_cpu_upper_bound = (seq_lens_cpu_upper_bound[:num_reqs] + 1).clamp(max=self.max_model_len)
	if step == 0:
	from vllm_ascend.utils import vllm_version_is
	if vllm_version_is("0.33.0"):
	# #60989 includes the correction token before calling this builder
	# on main. Only release still needs the extra prefill token here.
	# Keep the input batch's bounds unchanged.
	draft_seq_lens_cpu_upper_bound = (seq_lens_cpu_upper_bound[:num_reqs] + 1).clamp(max=self.max_model_len)

[gemini-code-assist]

On older vLLM versions (like v0.33.0), CacheConfig does not have the resolved_gpu_memory_utilization attribute. If gpu_memory_utilization is explicitly set to None or not resolved, accessing self.cache_config.resolved_gpu_memory_utilization will raise an AttributeError. We should use getattr with a fallback to DEFAULT_GPU_MEMORY_UTILIZATION to ensure compatibility across versions.

Suggested change

	def _get_gpu_memory_utilization(self) -> float:
	utilization = self.cache_config.gpu_memory_utilization
	# vLLM #56492 defers the default until the KV allocation mode is known.
	if utilization is None:
	return self.cache_config.resolved_gpu_memory_utilization
	return utilization
	def _get_gpu_memory_utilization(self) -> float:
	utilization = self.cache_config.gpu_memory_utilization
	# vLLM #56492 defers the default until the KV allocation mode is known.
	if utilization is None:
	return getattr(
	self.cache_config,
	"resolved_gpu_memory_utilization",
	self.cache_config.DEFAULT_GPU_MEMORY_UTILIZATION,
	)
	return utilization

[gemini-code-assist]

To prevent potential AttributeError if observability_config or profiler_config are None (e.g., in certain test environments or custom configurations), we should use defensive checks before accessing their attributes.

Suggested change

	if (
	scheduler.log_stats and scheduler.observability_config.enable_logging_iteration_details
	) or scheduler.vllm_config.profiler_config.profiler is not None:
	logging_enabled = (
	scheduler.log_stats
	and getattr(scheduler, "observability_config", None) is not None
	and getattr(scheduler.observability_config, "enable_logging_iteration_details", False)
	)
	profiler_config = getattr(scheduler.vllm_config, "profiler_config", None)
	profiler_enabled = (
	profiler_config is not None
	and getattr(profiler_config, "profiler", None) is not None
	)
	if logging_enabled or profiler_enabled: