mirror of
https://github.com/FlashML-org/FreeToken.git
synced 2026-11-21 01:52:30 +08:00
38 lines
1.5 KiB
Python
38 lines
1.5 KiB
Python
"""Version-gated ``cache_results`` for the pure-triton fallback kernels.
|
|
|
|
Without result caching, ``@triton.autotune`` re-benchmarks the whole config grid
|
|
on the first launch of every distinct key in every new process -- which lands
|
|
inside CUDA-graph capture / warmup and costs seconds at startup (worst on hybrid
|
|
models that touch many of these kernels). ``cache_results=True`` persists the
|
|
winning config per key to triton's on-disk cache and reloads it next time. The
|
|
on-disk key already folds in the triton version, the device/backend target hash,
|
|
and the kernel source hash, so a cached config is never reused across a different
|
|
GPU arch, triton build, or edited kernel.
|
|
|
|
``cache_results`` was only added to ``triton.autotune`` in newer releases; gate on
|
|
the signature so these fallbacks still import on older triton (there it is a
|
|
no-op). Mirrors kernel/fla/utils.py for the vendored FLA kernels.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import inspect
|
|
import os
|
|
|
|
import triton
|
|
|
|
_SUPPORTS_AUTOTUNE_CACHE = (
|
|
"cache_results" in inspect.signature(triton.autotune).parameters
|
|
)
|
|
|
|
# Escape hatch: FREETOKEN_TRITON_CACHE_RESULTS=0 forces a fresh sweep every run.
|
|
_CACHE_RESULTS = os.getenv("FREETOKEN_TRITON_CACHE_RESULTS", "1") == "1"
|
|
|
|
# Splat into each @triton.autotune(...): {"cache_results": True} on new triton,
|
|
# {} on old triton so the kwarg is simply absent.
|
|
autotune_cache_kwargs = (
|
|
{"cache_results": _CACHE_RESULTS} if _SUPPORTS_AUTOTUNE_CACHE else {}
|
|
)
|
|
|
|
__all__ = ["autotune_cache_kwargs"]
|