vllm.triton_utils

Modules:

• dispatcher –

  Registry for overriding Triton kernels with platform implementations.

• force_first_config –

  Skip Triton autotuning under VLLM_TRITON_FORCE_FIRST_CONFIG.

• tensor_descriptor –

Functions:

• use_tensor_descriptor –

  Tri-state VLLM_TRITON_USE_TD: unset=auto (on for XPU), 1/0=force on/off.

use_tensor_descriptor(override=None)

Tri-state VLLM_TRITON_USE_TD: unset=auto (on for XPU), 1/0=force on/off.

Source code in vllm/triton_utils/tensor_descriptor.py

def use_tensor_descriptor(override: bool | None = None) -> bool:
    """Tri-state VLLM_TRITON_USE_TD: unset=auto (on for XPU), 1/0=force on/off."""
    from vllm import envs
    from vllm.platforms import current_platform

    if override is None:
        override = envs.VLLM_TRITON_USE_TD
    if override is not None:
        return override
    return current_platform.is_xpu()