[v0.19.1rc1] fix(npu): preserve AFD DBO config during auto worker selection

[jiaran-king]

Summary

This is the release/v0.19.1rc1 backport of the same focused Issue #188 fix proposed in #189.

• base: release/v0.19.1rc1@921ef6fc725011b4662ed2c3f4eb6942df1f2115
• source fix commit: 50a42b73083a54294c34751905b74a3c3de6302d
• backport commit: 3fde850
• the five resulting source/test files are byte-identical to the main-branch fix
• no main-only commits were merged into the release branch

Root cause and fix

With AFD NPU DBO enabled and worker_cls="auto", vLLM-Ascend v0.19.1rc1 rewrites the temporary validation backend from deepep_low_latency to flashinfer_all2allv. vLLM 0.19.1 then rejects that backend during native DBO validation before the AFD worker can be created.

The backport:

• installs the Ascend patch before engine config construction
• wraps the complete NPUPlatform.check_and_update_config normalization pass
• restores enable_dbo, ubatch_size, and all2all_backend in finally
• performs role-specific AFD worker selection after upstream resolves the automatic worker
• keeps repeated VllmConfig.__post_init__() validation safe
• retries patch installation after an early vLLM-Ascend import failure

Users should continue to omit --worker-cls. Explicit Attention/FFN workers remain a compatibility path, not a requirement. Validation temporarily uses deepep_low_latency; the existing worker initialization helper keeps the actual Ascend runtime backend at flashinfer_all2allv.

Validation on the release base

Environment: macOS arm64, Python 3.12.13. The release checkout was installed editable without runtime dependencies for CPU-safe tests.

• python -m pytest -q tests/unit/compat/patches/test_config_validation.py tests/unit/compat/test_runtime.py
  • 29 passed
• python -m pytest tests/unit -o addopts=
  • 148 passed, 23 skipped
• python -m ruff check .
  • passed
• python -m ruff format --check .
  • 128 files already formatted
• git diff --check upstream/release/v0.19.1rc1..HEAD
  • passed
• content comparison against fix(npu): preserve AFD DBO config during auto worker selection #189's commit for all five changed files
  • identical

The regression suite models the actual platform ordering and covers Attention/FFN automatic workers, DBO state preservation, repeated validation, exception restoration, explicit workers, non-AFD NPU behavior, DBO off, GPU behavior, 310P/Xlite restrictions, runtime backend selection, and ImportError retry.

Real NPU validation pending

This task did not access an NPU server or run hardware E2E. External validation should confirm:

• no explicit --worker-cls on either role
• no flashinfer_all2allv DBO assertion during startup
• correct automatic Attention/FFN worker selection
• DBO remains enabled and both ubatches execute
• FULL_DECODE_ONLY graph capture/replay succeeds
• /v1/models and completion requests succeed
• actual worker runtime backend is flashinfer_all2allv
• auto-worker output matches the explicit-worker compatibility control

No DBO performance or throughput claim is made.

Related to #188 and #189. This PR must not close Issue #188 by itself.

[jiaran-king]

@yujuancao07, this release/v0.19.1rc1 backport is ready for real NPU acceptance testing of #188. This task intentionally did not run NPU hardware E2E. Could you please validate commit 3fde8500fd51a63f444ce046252a2166f79ccc74 against the checklist in the PR body—especially startup without --worker-cls, Attention/FFN automatic worker selection, preserved DBO with two executed ubatches, graph/completion success, and runtime flashinfer_all2allv—and post the key results or logs here? The issue should remain open until both target PRs are merged and hardware validation is recorded.

[jiangkuaixue123]

Please add test result.

[yujuancao07]

================================================================================================== FAILURES ===================================================================================================
________________________________________________________________________ test_npu_afd_config_patch_retries_after_initial_import_error _________________________________________________________________________

monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0xfffeee256c50>

    def test_npu_afd_config_patch_retries_after_initial_import_error(monkeypatch):
        fake_package = ModuleType("vllm_ascend")
        fake_package.__path__ = []
        monkeypatch.setitem(sys.modules, "vllm_ascend", fake_package)
        monkeypatch.delitem(sys.modules, "vllm_ascend.platform", raising=False)
        monkeypatch.setattr(ascend_runtime, "_PATCHES_APPLIED", False)

        ascend_runtime.apply_afd_ascend_patches_if_needed()

>       assert ascend_runtime._PATCHES_APPLIED is False
E       assert True is False
E        +  where True = ascend_runtime._PATCHES_APPLIED

tests/unit/compat/test_runtime.py:216: AssertionError