Alexander Efimov
Alexander Efimov
> But, at a glance, this flex operation is not custom op this flex op is a custom op, but it is "special" kind of custom op, that is supported...
@chunseoklee Do we have some agreements on directory structures? I thought we could simply move `/compiler/luci-micro` to `/lite` (or whatever we choose). MCU interpreter do not have any dependencies from...
@ptillet > LGTM, but could you add a lit test to make sure that we don't regress this optimization in the future? Will do. I have a question: Do you...
@ThomasRaoux Sorry. I'll update this PR and fix error when #3081 is merged
Just noticed that https://github.com/openai/triton/pull/3099 removed this pass in Nvidia backend. I'll remove USE_ROCM in this file, since now there are nothing to compare pass from this PR to
@ptillet @ThomasRaoux This PR is ready for review, PTAL
+cc @zhanglx13
This is needed to test changes from #3060
@micmelesse PTAL
@ptillet I've wrapped code in `USE_ROCM` macro and added checks in CMake files. Another approach I can think is a separate tool, something like `triton-amd-opt`. What do you think?