[Perf][SM70] Integrate AWQ decode defaults with safe capability fallback

[yangzhuxinyzx]

Purpose

Owner-authorized integration of #519/#520/#521/#522 on main, preserving all original commits and author attribution. This is not a competing kernel implementation. Adds default-on M1 admission with fallback for unsupported layer shapes and old native builds, plus the #509 loading-test interaction fix.

Changes

• Keep the reviewed QPN kernels, dequantization/FP16 boundaries, grouped dispatch and cooperative metadata implementation unchanged.
• Enable QPN M1 automatically only for the supported TP4/E512/native-g32 geometry and available native operator. No checkpoint-name/path identity admission.
• Unset flag + unsupported layer/native build retains the old path; explicit 1 fails closed; explicit 0 restores the old path. Restart is required for graph changes.
• Repair the loading lifecycle test stub for the new grouped flag. Preserve [Bugfix][SM70] Release unused AWQ/NVFP4 MoE conversion cache between layers #509 cleanup in production.
• Include main's [Bugfix][SM70] Stabilize QSA page4 order under KV relocation #494 logical-order QSA repair, which was a baseline dependency of the submitted model comparisons.

Test Plan and Result

At head d01b371, integrated main 96954be:

• Targeted CPU quantization/admission/loading suites: 118 passed. Run .venv/bin/python -m pytest --confcutdir=tests/quantization -q on test_awq_qpn_sm70.py, test_sm70_awq_active_grouped_decode.py, test_sm70_awq_compact_metadata.py, test_sm70_moe_load_cache_release.py, test_awq_sm70_fallback.py, test_sm70_awq_indexed_prefill.py, test_awq_sm70_persistent_cap.py and test_sm70_awq_prefill_exact_dense.py.
• Unmodified integration had 67 passes and one failure in the [Bugfix][SM70] Release unused AWQ/NVFP4 MoE conversion cache between layers #509 env stub; fixed, not suppressed.
• Applicable local pre-commit hooks pass.
• Fresh CUDA 12.8 / SM70 compile of awq_sm70_gemm.cu and production torch_bindings.cpp passes with ENABLE_SM70_TURBOMIND.
• Fresh reviewed awq_qpn_m1_sm70.cu compiles and links in an isolated test fragment using the production declaration/schema. This is not a complete release wheel. Initial test-binding header/macro omissions were corrected without changing production CUDA source.
• Fresh GPU execution is pending available GPU leases. Two CUDA-only cases skipped with CUDA hidden are NOT counted as passes. Existing contributor GPU/graph and independent FP64 evidence remains separately attributed in docs.
• No full E2E rerun, model deployment, GPU preemption or wheel release.

Performance and numerical decision

Reuse, do not relabel, original same-contract evidence: C1 x 64K pure decode 50.0326 -> 59.0313 tok/s; grouped-route C4/C8 gains are separate measurements and not additive. Compact metadata saves approximately 0.88 GiB/card and its default-change model comparisons preserve throughput within recorded noise. The recorded 65-case task screen and 1152 independently checked local samples support bounded-shape promotion under the owner's policy, not greedy equality or statistical noninferiority. Preserve both the GSM loss and IFEval gain; neither is hidden or declared proof of zero regression. No new speed measurement is claimed.

AI assistance: OpenAI Codex. Integration and merges are explicitly requested by the repository owner; independent human execution of these new tests is not asserted. Draft until final CI and base/head verification.

[yangzhuxinyzx]

已通过完整静态 CI 34019173319 并合并，main 合并提交 a4567d7。最终集成树与本地验证树一致。#520/#521/#522 自动标记 merged。更正提交祖先说明：#519 的等价 kernel 已由后继分支的 9ae1a29 提交携带，但原 PR 的独立提交 5b4135f 不在祖先链，所以 #519 继续 open；没有把它误记成已合并。118 项 CPU 回归和编译结果见正文，不宣称重复运行 GPU/E2E。