Python 3.14t: C extension missing Py_MOD_GIL declaration causes GIL re-enable warning

[kdt2014]

Description

Triton 3.5.0 provides binary wheels for Python 3.14t (free-threaded build), but the C extension module triton._C.libtriton does not declare GIL compatibility, causing Python to re-enable the GIL with a warning.

Environment

• Python version: 3.14.0 free-threading build
• Triton version: 3.5.0
• Platform: Linux x86_64
• Installation: pip install triton (cp314t wheel)

Current Behavior

When importing Triton in Python 3.14t:

import triton._C.libtriton

Output:

RuntimeWarning: The global interpreter lock (GIL) has been enabled to load module 
'triton._C.libtriton', which has not declared that it can run safely without the GIL.

Expected Behavior

The C extension should declare its GIL compatibility status using the Py_mod_gil slot in the module definition, as described in PEP 703.

Reproduction Steps

python3.14t -c "import triton._C.libtriton"

Suggested Fix

Add the appropriate Py_mod_gil declaration in the C extension module initialization:

static PyModuleDef_Slot module_slots[] = {
    {Py_mod_gil, Py_MOD_GIL_NOT_USED},  // or Py_MOD_GIL_USED if needed
    {0, NULL}
};

Additional Context

• Since Triton already builds and distributes cp314t wheels, adding the GIL declaration would complete the free-threaded support
• This affects users trying to leverage true parallel execution in Python 3.14t
• The warning can be suppressed with PYTHON_GIL=0, but proper declaration is the correct solution

References

• PEP 703: https://peps.python.org/pep-0703/
• Python 3.14 free-threading docs: https://docs.python.org/3.14/howto/free-threading-extensions.html

[Qubitium]

@kdt2014 You also need my PR to even remotely make Triton autotuner kernels launch under nogil without segfaults.

#8437