Checklist

• If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Motivation

On macOS, SGLang installs a mock triton module via _triton_stub.py so that code importing triton doesn't crash. However, with torch 2.9+, the stub is missing triton.compiler / triton.compiler.compiler sub-modules, causing bench_one_batch and server startup to fail.

Problem

The import chain that triggers the crash:

torchvision → torch._dynamo → torch._inductor.runtime.hints → import triton.compiler.compiler

Because the triton stub makes has_triton_package() return True (since triton exists in sys.modules), torch proceeds to import triton.compiler.compiler — which doesn't exist in the mock, causing ModuleNotFoundError.

ModuleNotFoundError: No module named 'triton.compiler'

The stub currently only registers:

• triton
• triton.backends
• triton.backends.compiler

But not triton.compiler or triton.compiler.compiler, which torch 2.9+ expects.

Additional Issue

The _TritonFinder meta-path finder uses the deprecated find_module/load_module protocol, which doesn't work reliably in Python 3.12+. Should be updated to find_spec (PEP 451).

Suggested Fix

1. Register triton.compiler and triton.compiler.compiler as mock sub-modules in _triton_stub.py
2. Update _TritonFinder to use the modern find_spec API

Environment

• macOS (Apple Silicon M3 Pro)
• Python 3.12.11
• torch 2.9.1
• torchvision 0.24.0
• SGLang main branch (commit e8d46f1)