Yukio Siraichi

@ezyang @jansel Here's a summary of the last changes: - Fixed the unique name generation - Previously thought that `sources` didn't have repeated sources (guarding against that, now) - Refactored...

@ezyang @jansel @voznesenskym Here's a summary of this week's changes: - Fixed CSE implementation - There were 2 problems with the implementation: 1. I was trying to count expressions that...

@voznesenskym It should be ready to benchmark (it's running all the benchmarks the vanilla PyTorch passes).

@ezyang @jansel I haven't checked whether these guards produce correct results. Should I try and run TorchBench with `--accuracy` flag?

@ezyang @jansel @voznesenskym @Lezcano Here are the results of re-running the C++ backend version, without CSE at all. Before that, a few observations: - I am not completely sure why...

I am still investigating the CI failures. That said, do you have any idea of why that is happening? ``` OSError: /tmp/torchinductor_jenkins/jf/cjfwzugaedyclbugrkjgubl7th5t4sp6yci4knbcctdgi2vju5mh.so: undefined symbol: _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_S2_ ```

As far as I've seen it, it should be linking against libtorch. Also, the linking process seems to be fine. The error is raised when the newly created library is...

Disregard my last comment. Jansel was completely right. I couldn't really repro on my machine because [this condition](https://github.com/pytorch/pytorch/blob/e0c24ec2a5db3581ce54ab5cbf49eee4bf1c2716/torch/_inductor/codecache.py#L391-L394) was always true (`vec_isa=avx2`). In the CI, that condition was false, i.e....

@ezyang @jansel As far as I can see, there are 3 types of errors in the CI tests: 1. [Lint] failed on "Calculate docker image" - Likely to be solved...

Update on `GLIBCXX` version error: I can't seem to reproduce the error. Here are the steps I went through: - Ssh into the `linux-focal-py3.9-clang7-asan / test (default, 2, 5, linux.4xlarge)`...