fast_decode_enable.py

"""
fast_decode_enable.py — PERMANENT harness-side switch for FastGenerator decode.

One line at the top of serve_emoe.py (or any eval script), after other imports:

    import fast_decode_enable   # noqa: F401

Every HyperExpertRunner constructed afterwards decodes through FastGenerator
(sliding anchor) automatically. No edits to runtime_adapters.py; base and
hyper weights untouched.

Kill switch (instant native A/B, no code change):
    FAST_DECODE=0 python serve_emoe.py ...      -> native decode
    python serve_emoe.py ...                    -> fast decode (default on)

Every generation logs "generate[FAST]" when active. If you don't see that
tag, you are measuring native.
"""
import os

import runtime_adapters as RA
from runner_patch import install

_ENABLED = os.environ.get("FAST_DECODE", "1") != "0"
_orig_init = RA.HyperExpertRunner.__init__


def _patched_init(self, *a, **k):
    _orig_init(self, *a, **k)
    if _ENABLED:
        install(self)
        print("fast_decode_enable: FastGenerator decode ON "
              "(FAST_DECODE=0 to disable)")
    else:
        print("fast_decode_enable: DISABLED via FAST_DECODE=0 (native decode)")


RA.HyperExpertRunner.__init__ = _patched_init