Nanobind Static Build Free-Threaded PEP 703 GIL Slot Declaration in OpenAI Triton

Engineering objective

System challenge

In CPython 3.13t free-threaded builds (PEP 703), extension modules compiled with nanobind without explicit GIL slot declarations trigger runtime GIL reactivation warnings, degrading GPU compilation throughput in multi-threaded workflows.

Constraints

Working within real limits

• ✕Preserve compatibility across standard GIL and free-threaded CPython builds
• ✕Maintain ABI stability with nanobind bindings in static extension mode
• ✕Pass full upstream CI across Linux x86_64, CUDA, and AMD ROCm runner matrices

Approach

How it was built

1. 01

   Audited Triton nanobind static extension definitions in python/src/interpreter.cc and python/src/llvm.cc.

2. 02

   Explicitly declared the Py_MOD_GIL_NOT_USED slot in the module definition structure under Py_GIL_DISABLED.

3. 03

   Validated multi-threaded kernel compilation scaling without GIL re-acquisition penalties.

4. 04

   Verified zero compiler regressions across upstream test matrices.

Technical validation

Implementation evidence

Observable engineering evidence. Discipline: MERGED UPSTREAM

Public evidence

Review the upstream contribution

Outcome

What was delivered

• Eliminated GIL reactivation warnings on free-threaded Python 3.13t runs

• Unblocked true multi-threaded GPU kernel compilation pipelines

• Achieved 100% green CI on first submission to triton-lang/triton

Technologies

Stack and tools

Previous and Next Engineering Systems