vllm.v1.attention.ops.flydsl_kernels

Modules:

• buffer_utils –

  Buffer-resource (V#) views shared by the FlyDSL decode kernels.

• mma_utils –

  MMA-atom calls on register vectors for the FlyDSL decode kernels.

• tq_decode –
• ultraquant_decode_hd256 –

_run_compiled(launcher, *args)

Reuse the compiled callable of a factory-style host launcher.

Source code in vllm/v1/attention/ops/flydsl_kernels/__init__.py

def _run_compiled(launcher, *args) -> None:
    """Reuse the compiled callable of a factory-style host launcher."""
    import flydsl.compiler as flyc
    from flydsl.utils import env

    if env.compile.compile_only:
        flyc.compile(launcher, *args)
        return

    compiled = getattr(launcher, "_cf", None)
    if compiled is not None:
        compiled(*args)
        return

    # flyc.compile executes once; do not launch its result again on this call.
    launcher._cf = flyc.compile(launcher, *args)