[baominghelly]

Summary

• Install the declared build dependencies before non-isolated CI builds, including scikit-build-core>=1.1.1, and use python -m pip consistently.
• Add public and generated include directories to Iluvatar's custom dispatch compiler command.
• Query the Ascend runtime for the exact SOC name, then fall back to qualified npu-smi model tokens. Reject numeric fields such as 3103 and request an explicit SOC_VERSION when detection is ambiguous.
• Add 21 hardware-independent regression cases for build setup, wheel metadata, include lookup and SOC discovery.

Motivation

Follow-up to #1003. CI images can retain an older scikit-build-core when builds use --no-build-isolation, causing the sdist.force-include configuration to fail. Iluvatar dispatch compilation lacks header search paths, and the previous Ascend parser can interpret HBM statistics as a chip model.

This draft runs the fixes through the real accelerator CI environments. Local validation does not establish that hardware builds or operator tests pass.

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• Hygon (WITH_HYGON)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Accelerator smoke builds/tests: Not run locally — this Windows host lacks the accelerator SDKs and hardware. CI is the required hardware validation.

Local regression validation (PowerShell, Python 3.13, scikit-build-core 1.1.1, CMake 4.4.4 and MinGW g++):

$env:PATH = (Join-Path (Get-Location) '.venv/Scripts') + ';' + $env:PATH
.venv/Scripts/python.exe -m pytest --confcutdir=tests/build_config tests/build_config -q
# 21 passed in 11.62s
.venv/Scripts/ruff.exe check tests/build_config/test_ci_build.py
# All checks passed!
.venv/Scripts/ruff.exe format --check tests/build_config/test_ci_build.py
# 1 file already formatted
git diff --check
# No errors

These tests generate wheel metadata with the real build backend, compile a minimal C++ translation unit with the actual Iluvatar include flags, and invoke the actual SOC CMake detector with simulated runtime/SMI outputs. They do not compile accelerator kernels. --confcutdir avoids loading the operator test suite's torch/InfiniOps fixtures.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Not run locally — SDK/hardware unavailable	Pending CI
Iluvatar	Yes	Not run locally — CoreX SDK/hardware unavailable	Local include lookup regression passed; pending CI
MetaX	Yes	Not run locally — SDK/hardware unavailable	Pending CI
Cambricon	Yes	Not run locally — SDK/hardware unavailable	Pending CI
Moore	Yes	Not run locally — SDK/hardware unavailable	Pending CI; runner availability is outside this change
Ascend	Yes	Not run locally — CANN/hardware unavailable	Simulated SOC regressions passed; pending CI
Hygon	Yes	Not run locally — SDK/hardware unavailable	Setup updated; current automatic CI platform list omits Hygon

Benchmark / Performance Impact

N/A — build and CI fixes; no operator kernel changes.

Notes for Reviewers

• Verify both CI and CI v2 Shadow against this PR's head commit.
• Expected evidence: metadata configuration succeeds with the declared backend version; Iluvatar dispatch compiles past the missing public/generated headers; Ascend uses the exact SOC name and never Ascend3103.
• Ascend automatic detection now fails explicitly if neither source provides a qualified model. Cross-compilation and inaccessible devices require an explicit SOC_VERSION.
• Ascend device allocation and Moore runner availability remain outside scope. A failure there must be distinguished from a regression in these fixes.
• No changed file matches the repository's clang-format gate (.h, .cc, .cuh, .mlu).