Jez Ng
Code is a lot easier to review if you ignore whitespace: https://github.com/triton-lang/triton/pull/4392/files?w=1
Updated. It looks like `test_performance.py` has bitrotted though so I couldn't test it: ``` python/test/regression/test_performance.py:6: in import triton.ops E ModuleNotFoundError: No module named 'triton.ops' ```
oops, didn't mean to close this
Bump. Just rebased
I have no intention of changing the current behavior wrt cudagraphs; it's orthogonal to making the autotuner device-agnostic.
They are `do_bench`-specific parameters, so if we are parameterizing `autotune` over `do_bench`, we can just pass them directly to `do_bench`: ``` @autotune( ..., do_bench=lambda fn: triton.testing.do_bench(fn, warmup=25, rep=100) ) ......
Bump on this -- I'm particularly interested in whether there are objections to outright removing the `warmup` and `rep` parameters, or if we need a gradual deprecation process. @ptillet @Jokeren...
ISTR Triton having quite a few breaking changes in the past, though. E.g. the relocation of the libdevice code (and more I can't recall). PyTorch pins the version of Triton...
> I can't believe I need to say it but the fact that you've made the HW vendors' life miserable on multiple occasions in the past should push you to...
cc @albanD, per this discussion: https://github.com/pytorch/pytorch/pull/113610#discussion_r1392922602 I too would like Module fields to all be properly statically typed, but that seems to conflict with the desire to allow people to...