feat(qy): enable QY CUDA runtime device backend

[Ceng23333]

Summary

Add QY (Denglin) CUDA runtime device backend so InfiniOps/InfiniLM can target QY for the MiniCPM5 FA/longctx/cudagraph stack.

Motivation

Required by ITW pin release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph and companion InfiniOps QY backend.

Changes

• QY native device/runtime headers and registration under src/native/cuda/qy/
• Device dispatch + install-consumer smoke test updates

Validation

• Configure / build / ctest (QY): not run here (no QY SDK on PR host). Request maintainer with QY access:

  cmake -S . -B build-qy -DCMAKE_INSTALL_PREFIX=$PWD/prefix-qy -DWITH_CPU=ON -DINFINI_RT_BUILD_TESTING=ON # plus QY backend flags per docs/backends.md
  cmake --build build-qy -j
  ctest --test-dir build-qy --output-on-failure

• Rebased onto current master; clang-format / ruff clean on touched files.
• Integration-validated via the ITW release pin above.

Checklist

• Conventional Commits title/branch (feat/qy-device)
• Minimal change scoped to QY backend enablement
• Smoke build+test on QY (blocked: no local QY SDK)
• Stated unrun checks and why

Related

• Companion InfiniOps feat/qy-backend PR
• Downstream InfiniCore pin + InfiniLM PR

Made with Cursor

Stack PR index (merge order)

1. InfiniOps: feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels InfiniOps#1002
2. InfiniRT: feat(qy): enable QY CUDA runtime device backend #49
3. InfiniCCL: feat(qy): QY-compatible CPU device header fix InfiniCCL#77
4. InfiniCore: chore(qy): pin InfiniRT/Ops/CCL for MiniCPM5 FA/longctx InfiniCore#1575
5. InfiniLM: feat(qy): MiniCPM5 FA, longctx MoE/RoPE, CUDA-graph capture safety InfiniLM#597 → base refactor/adopt-modern-infini-stack

ITW pin: https://github.com/Ceng23333/InfiniTensorWorktree/tree/release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph

[Ceng23333]

QY required smoke (PR head)

• Head SHA: 95d30fe80c7be6cbc8090e1596688709626865ea
• Host: Denglin QY (QY_ROOT=/usr/local/dlgpu in container), DENGLIN_DEVICES=3
• Command: ctest --test-dir …/prhead-install/build/infinirt --output-on-failure (with LD_LIBRARY_PATH preferring build prefix over /opt/infinicore)
• Result: PASS (12/12, incl. test_qy_runtime / test_qy_graph)
• Log: InfiniTensorWorktree-v20261009-post-refactor/logs/prhead-verify/rt-ctest-ldfix.log

100% tests passed, 0 tests failed out of 12