[jiangkuaixue123]

Purpose

Unused runtime helpers remained in the production package after the runner migrations. Remove the obsolete microbatch wrapper, DeepSeek attention-builder input isolation helper, unused configuration helpers, and unused connector key export. Move FP8 test-weight generation and runtime class resolution into test helpers.

Issue

No linked issue.

Scope

• Production helper cleanup, associated tests, design documentation, and review guidance.
• The DeepSeek V2/V3 Async CAM model execution path is outside this change.

Implementation Notes

• Keep the active MRV2 metadata installer and RoPE materialization functions. Use the MRV2 installer in the retained test that checks child contexts do not resend control messages.
• Delete the two tests dedicated to retired helpers; preserve native-stage isolation, FP8 dequantization, and runtime classpath coverage.
• Share test-weight generation through tests/unit/fakes/weights.py and class resolution through tests/unit/fakes/runtime.py, including the isolated NPU import script.
• Add a type-only cast to coerce_extra_int() to fix an existing mypy overload error reproduced on the base revision. Conversion and error handling remain unchanged.
• No new patches, native source changes, or intended serving behavior changes.

Test Plan

• .venv/bin/python -m pytest tests/unit -m 'not gpu and not vllm_runtime' -ra
• .venv/bin/python -m pytest tests/unit/config -o addopts= -q
• Run pre-commit on every changed file; check removed-symbol references and git diff --check.

Test Result

• CPU unit suite: 751 passed, 106 skipped, 8 deselected, 10 subtests passed.
• Configuration suite after the type-only correction: 135 passed.
• Moved class-resolution helper: dotted/colon paths and invalid-path errors checked directly.
• Pre-commit passed, including Ruff, mypy for Python 3.11, Markdown lint, typos, and SPDX checks. Commit hooks also passed.
• Removed helpers have no remaining source/documentation references; migrated helpers have no production imports.
• Environment: macOS, Python 3.12.13, without torch, vLLM, or vLLM-Ascend. Dependency- and build-tool-gated cases were skipped using existing guards.
• Target pins remain vLLM 0.32.0 (5df63a7203cdb9328d824a27b682e77c7b9928be) and vLLM-Ascend bd42eea01d8312c4df469fabaed45f77a12df4d3.
• GPU/NPU runtime tests and native builds were not run; native sources and ABI are unchanged.

Docs Impact

Update the module design index, Attention runtime and model integration pages, and the review skill's worker/engine references to describe the remaining production paths.

---

Essential PR Checklist

• Purpose is clear and linked to public context when possible.
• Scope is bounded.
• Compatibility with vLLM v0.32.0 and, for NPU changes, the pinned vLLM-Ascend source is considered, as applicable.
• No changes are made to the vLLM source checkout.
• Plugin-owned classes or explicit dotted class paths are preferred over monkey patches.
• Any compat shim or monkey patch is isolated, idempotent, version-guarded, documented, and tested. No new shims or patches.
• Imports remain CPU-safe; CUDA-heavy work is delayed or GPU-gated.
• Validation evidence is included, including skipped GPU tests when applicable.
• Documentation impact is stated.