fix(npu): preserve AFD DBO config during auto worker selection

[jiaran-king]

Summary

• install the AFD Ascend compatibility patch before upstream constructs the NPU VllmConfig
• preserve enable_dbo, ubatch_size, and the validation backend across the complete vLLM-Ascend platform normalization pass
• select the role-specific AFD NPU worker only after upstream resolves worker_cls="auto"
• keep repeated VllmConfig.__post_init__() validation safe and allow a failed early vLLM-Ascend import to be retried

Root cause

For an AFD NPU config with DBO enabled, AFD temporarily presents deepep_low_latency so vLLM 0.19.1's native DBO backend assertion can run. vLLM-Ascend v0.19.1rc1 then normalizes an automatic, non-SP worker by replacing that backend with flashinfer_all2allv. The native assertion consequently fails before the AFD worker is created.

The previous compatibility patch wrapped only NPUPlatform._fix_incompatible_config. It did not cover the later backend rewrite in the complete NPUPlatform.check_and_update_config pass.

Behavior after this change

Users continue to omit --worker-cls:

• Attention resolves to afd_plugin.v1.worker.npu.AFDNPUAttentionWorker
• FFN resolves to afd_plugin.v1.worker.npu.AFDNPUFFNWorker
• DBO and two-ubatch settings survive platform normalization
• deepep_low_latency is only the validation-safe backend
• the existing worker initialization helper still selects the actual Ascend runtime backend, flashinfer_all2allv

Explicit role-specific workers remain compatible, but are not required by this fix. Non-AFD NPU and GPU platform behavior is unchanged. Existing 310P/Xlite automatic-selection restrictions remain in place.

Validation performed

Environment: macOS arm64, Python 3.12.13. The package 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
  • proves the real normalization order is modeled: DBO reset, auto-worker backend rewrite, native backend assertion, state restoration, and role-specific worker selection
• 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
  • passed

The regression coverage also verifies repeated __post_init__, exception-safe restoration, explicit workers, non-AFD NPU behavior, DBO-off behavior, GPU automatic workers, 310P/Xlite restrictions, the runtime flashinfer_all2allv helper, and retry after an initial vLLM-Ascend import failure.

Real NPU validation pending

This task did not log in to an NPU server or execute hardware E2E. External validation should confirm:

• Attention and FFN commands omit --worker-cls
• startup no longer raises the flashinfer_all2allv DBO assertion
• both processes select the correct role-specific AFD NPU workers
• DBO remains enabled with two ubatches and both ubatches execute
• FULL_DECODE_ONLY graph capture/replay completes
• /v1/models and a completion request succeed
• the actual worker runtime backend is flashinfer_all2allv
• auto-worker output matches the explicit-worker compatibility control

This change makes no DBO performance or throughput claim.

Related to #188. The issue should remain open until both target branches are merged and hardware validation is recorded.