Make tilelang a Linux-only dependency (like triton)

[Copilot]

Motivation

tilelang was added as an unconditional dependency in 0.12.2, but it is painful to build on Windows and requires a C++ compiler at runtime. It is only used by CUDA-specific SSM kernels (GatedDeltaRule, causal_conv1d) which are irrelevant on non-Linux platforms.

Modification

Add the same platform marker to tilelang in requirements/runtime_cuda.txt that already governs triton:

tilelang; sys_platform == "linux" and "aarch64" not in platform_machine and "arm" not in platform_machine

The Python call-sites already gate on has_tilelang(), so no runtime behaviour changes on Linux.

BC-breaking (Optional)

No. Windows and ARM Linux users were already unable to use tilelang-backed kernels; this just stops the install from failing.

Use cases (Optional)

N/A

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[Copilot]

Pull request overview

Updates CUDA runtime dependencies so tilelang is only installed on supported platforms (Linux, non-ARM), matching the existing triton platform marker behavior and preventing install/build failures on Windows and ARM.

Changes:

• Add a Linux/non-ARM PEP 508 environment marker to the tilelang dependency in requirements/runtime_cuda.txt.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.