Jez Ng

Switch back to plain Coffeescript.

3

I've decided that I don't like IcedCoffeeScript all that much :( See reasons here: https://github.com/int3/doppio/issues/158#issuecomment-13928332

I've implemented the ability to 'dedent' the `where` clause by a half-shift-width, as described in these [style](https://github.com/tibbe/haskell-style-guide/blob/master/haskell-style.md#indentation) [guides](http://snapframework.com/docs/style-guide). If you're interested, I'll put it under an option setting and send...

Suggestion: (##) for setting element ID

3

I find it to be an operation at least as common as setting the CSS class. And `(##)` would be in obvious analogy to the CSS syntax plus the existing...

Better profile_callable

1

Summary: I extended profile_callable to accept a list of events of interest. In the case where this parameter is not specified, profile_callable will collect the timings of all events except...

Summary: Now that https://github.com/facebookincubator/AITemplate/pull/845 has landed, the backend supports softmax with `dim != -1` directly, and the fx converter no longer needs the workaround from https://github.com/facebookincubator/AITemplate/pull/395. Differential Revision: D48248330

Avoid highlighting if someone else has already done it.

1

This makes it play nice with other plugins.

Make `multi` accessible from outside Rendezvous

1

Use case: I'm writing a CPS interpreter. ICS has proven invaluable for simplifying the code, but the call-defer-once limitation means that I can't write generators that yield more than once....

Get cutlass_library import working under fbcode

5

Differential Revision: D56764089 cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire

Have do_bench_cudagraph create its own benchmarking stream

6

Per https://github.com/pytorch/pytorch/pull/9938, which fixes https://github.com/pytorch/pytorch/issues/9646, CUDA streams are now cheap to create under PyTorch. Let's have the benchmarking function create one per run instead of requiring its callers to do...

Remove fast_flush parameter from do_bench

2

The parameter was introduced in https://github.com/triton-lang/triton/pull/840, and it looks like it exists mainly to ease migration. In general there's no reason to use fast_flush=False, so let's remove it.