[voltjia]

Summary

• Allow WITH_NINETOOTHED to select the Ascend backend in addition to the existing NVIDIA Triton backend.
• Pass the NineToothed backend and platform through the InfiniOps code generator.
• Generate Ascend RMSNorm FP32/FP16 rank-2/3/4 and block-size 256/512 variants.
• Link the generated standalone Ascend C++ launchers with the Ascend runtime.
• Register RmsNorm implementation index 9 for Ascend and share its launcher setup with the NVIDIA specialization.
• Explicitly skip BF16 for Ascend NineToothed RMSNorm because triton-ascend 3.2 currently crashes in triton-adapter-opt for that AOT configuration.

Motivation

InfiniOps already integrates NineToothed-generated RMSNorm on NVIDIA. This extends the same generated-native-kernel path to Ascend 910B4 instead of requiring a Python loader at inference time.

Type of Change

• feat — new feature / new operator / new platform
• Breaking change for callers that explicitly request Ascend index 9 with BF16 before that compiler limitation is fixed

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Ascend (WITH_ASCEND)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

# Ascend 910B4, final InfiniOps commit c6e2d25
$ python3.11 -m pytest -q tests/test_rms_norm.py --devices npu -k "npu-9" --tb=short
24 passed, 13 skipped, 37 deselected in 1.72s

# NVIDIA A100, final InfiniOps commit c6e2d25
$ python3 -m pytest -q tests/test_rms_norm.py --devices nvidia -k "cuda-9" --tb=short
37 passed, 37 deselected in 0.51s

# Generator
$ python3.11 -m pytest -q tests/test_generate_ninetoothed_ops.py --tb=short
2 passed in 0.07s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	WITH_CPU=ON, WITH_NVIDIA=ON, WITH_NINETOOTHED=ON build passed	RMSNorm NineToothed index 9: 37 passed
Iluvatar	No	N/A - not affected	N/A
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	Yes	WITH_CPU=ON, WITH_ASCEND=ON, WITH_NINETOOTHED=ON build passed	FP32/FP16: 24 passed; BF16: 13 skipped for the documented triton-ascend adapter crash

Full `pytest` output

Ascend:
24 passed, 13 skipped, 37 deselected in 1.72s

NVIDIA:
37 passed, 37 deselected in 0.51s

Generator:
2 passed in 0.07s

Benchmark / Performance Impact

N/A. This PR adds an Ascend integration path and does not claim a performance result.

Notes for Reviewers

This branch is intended to stack on these upstream changes:

• Feat(backend): Modular non-invasive Ascend backend integration ninetoothed#240
• Fix static bounds for Ascend kernels ssen0829/ninetoothed#1
• Support Ascend dtype promotion and C++ exports ssen0829/ninetoothed#2
• Cast RMSNorm epsilon for AOT ntops#108

The Ascend validation used NineToothed 9838a1e3, ntops fcc2ade5, and InfiniRT 5175a694 on an idle 910B4. The generated launchers embed kernel binaries, register them through rtDevBinaryRegister/rtFunctionRegister, and invoke them through rtKernelLaunch.

CI Notes

At final head c6e2d25, clang-format, Ruff, documentation build, matrix generation, and the NVIDIA unit job passed. The repository platform matrix also reported Ascend, Cambricon, and MetaX failures.

The latest master run at 5653c00 had already failed the same CI workflow in its Ascend and MetaX jobs. The PR Ascend log fails while building stock AscendC workspace kernels, before NineToothed code generation. The Cambricon errors are native Cambricon embedding/mul type-conversion compile errors and do not reference files changed by this PR. Direct affected-platform builds and tests above pass with the intended upstream dependency stack.