[Core][Kernel][SM70] Default validated PP2 TP4 paths

[yangzhuxinyzx]

Purpose

Promote the two measured inference-engine wins from the original campaign while removing superseded or unproven experiments.

• Default-on SM70 block-FP8 QPN8 for exact serialized PP2 x TP4, B1, no-spec operator/tensor contracts. Admission uses device, topology, concurrency, operator role, tensor shape/dtype/layout, and workspace availability only—never model name, checkpoint, model_type, or architecture identity. Missing operators/OOM fall back to TurboMind; VLLM_SM70_FP8_QPN8=0 or VLLM_SM70_FP8_QPN8_PP2_TP4=0 rolls back.
• Default-on metadata-free PP transfer for the exact replicated FP16 [1,4,4096] B1 schema. The engine validates the model-provided intermediate tensor schema on both stages and rejects DBO, ubatching, speculation, sequence parallelism, non-SM70, and non-CUDA-Graph configurations. VLLM_SM70_PP_STATIC_HIDDEN_TRANSFER=0 rolls back.
• Strict paired quality artifacts: dataset hashes and evaluation contracts must match, artifacts must be self-consistent, and aggregate HumanEval/LongBench/GSM8K/needle quality must not regress. Individual flips are reported but greedy/token identity is not a gate.

The generic PP profiling repair and compact MXFP4 work were already merged through #295. The short-indexer experiment and unrelated large patch surface are not retained.

Performance and numerical result

• QPN8: 59.248 -> 64.359 token/s (+8.63%), mean TPOT 16.878 -> 15.538 ms.
• Static PP on top: 64.359 -> 64.502 token/s (+0.22%), about 0.034 ms/token.
• Paired GSM8K-64: 63/64 -> 63/64, zero invalid answers; one regression is balanced by one improvement. No greedy identity requirement is imposed.
• Operator screen: finite outputs, relative L2 <= 6.05e-4, cosine >= 0.9999997, max absolute difference <= 0.00390625. All 365 audited block scales are finite positive powers of two and transform exactly.

Test Plan

• Exercise default/rollback and exact QPN8 shape/topology/concurrency admission.
• Exercise default weight preparation and grouped dispatch.
• Exercise static metadata-free send/receive, worker admission/schema, and self-copy avoidance.
• Exercise aggregate quality, contract/hash mismatch, malformed artifacts, and directional flips.
• Run changed-file pre-commit and focused CPU suites; do not repeat broad model E2E.

Test Result

• Focused combined suite: 47 passed.
• Changed-file pre-commit: passed, including Ruff, formatting, typos, Markdown lint, mypy, forbidden imports, SPDX, and repository policy hooks.
• git diff --check: passed.
• Added-line identity/private-path audit: no findings.
• Current GPU rerun was not started because GPUs 0-3 host an unrelated live service and GPUs 4-7 are reserved by unrelated processes. No process was interrupted. The existing PR evidence contains the V100 source-integration, CUDA Graph, three-run endpoint, and paired quality results above.

[yangzhuxinyzx]

Audit requirement from the current review: DeepSeek V4 may name the matched performance workload, but must not become a runtime activation identity. Any implementation should admit optimized routes only through inference-engine contracts such as SM70 capability, PP/TP topology, operator and collective support, tensor shape/dtype/layout, cache state, and CUDA Graph/concurrency conditions—never model name, checkpoint path, model_type, or architecture identity. Keep this PR Draft until source changes plus focused route, correctness/quality, cleanup, and matched pure-decode evidence are present. The current documentation-only scaffold is not merge-ready.

[yangzhuxinyzx]

本 PR 已完成拆分审计。可证明安全且与模型身份无关的引擎核心子集，已由 #295 在最新 main 上重新实现、验证并合并：PP profiling 按 semantic hidden_states 解包，以及两个精确张量形状的默认关闭 SM70 MXFP4 compact 路径。

本 PR 剩余实验不合并：QPN8 候选已有成对基线正确/候选错误的质量回归证据；静态 PP 传输仅约 0.034 ms/token 且现有记录已否定该方向；其余大范围实验缺少足够的通用安全收益闭环。保留本分支和证据供后续独立研究，避免把互不相关且存在风险的实验一次性进入 main。