[KernelGen][Nvidia] Add _cholesky_solve_helper operator with Triton kernel

[Yukun-Cui]

PR Category

Operator

Type of Change

New Feature

Description

Adds a two-stage Triton triangular solve for _cholesky_solve_helper, including broadcasted batch handling and upper/lower factor support.

This is a reviewed and updated reintroduction of the implementation previously discussed in #3354.

Issue

Related prior PR: #3354

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Testing

• pytest -q tests/test_cholesky_solve_helper.py — PASS (6 cases) on NVIDIA H20
• strict operator checker — PASS (0 errors, 0 warnings)
• pre-commit hooks — PASS

Performance

Command: pytest -q benchmark/test_cholesky_solve_helper.py --level core -s on NVIDIA H20.

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float32	B [8, 2], A [8, 8]	0.031744	0.138880	0.229x
float32	B [64, 8], A [64, 64]	0.034464	0.158992	0.217x

Core suite: 5 cases, geometric mean speedup 0.217x, measured range 0.200x–0.229x.

This initial implementation prioritizes correctness and operator coverage.

Multi-backend Testing

Backend	Accuracy	Benchmark	Notes
NVIDIA H20	PASS (6 cases)	PASS (5 cases)	Primary validation platform
Other backends	Not run	Not run	CI coverage requested

Files Changed

• Operator kernel or dispatcher wrapper and registrations
• Accuracy tests and core benchmark
• Operator metadata and tuning configuration where applicable