@iotamudelta any advice on fixing

/var/lib/jenkins/workspace/aten/src/ATen/native/hip/RangeFactories.hip:61:35: error: no member named 'par' in namespace 'thrust::cuda'
22:30:07       auto policy = thrust::cuda::par.on(stream);

?
In other places (e.g. Embedding.cu) the following construct is used

auto policy = thrust::cuda::par(allocator).on(stream);

but I don't need allocator for the ranges, and thrust::cuda::par.on(stream) is valid syntax.

@ngimel looks to me like a shortcoming of the hipified Thrust. Could this be a construct that didn't exist in Thrust some two years ago? Either way, we are working on updating things in that realm, so could you work around it via __HIP_COMPILER_HCC__ ifdefs (either w/ allocator or serial execution, I guess) for now and leave a TODO in the code so that I can fix it once the bigger picture there is sorted out? Thanks!

rocm builds were fixed by including header, caffe2_onnx failure seems unrelated. BTW, what happened to cuda 10 ci?

BTW, what happened to cuda 10 ci?

It's run on master, but not on every commit. If you like, I can make it possible to explicitly request cuda 10 CI on pull requests, if you're futzing around with CUDA 10 specific code.

I'm going to go ahead and merge this, but ideally we'd also turn back on the failing tests when we make fixes like this. (This might be hard/annoying to do.)

But it does not fix the tests :-( Most of the tests fail because of backward running on the wrong stream. This is something that I noticed looking on the profile, and that has to be fixed anyway.

@ezyang merged this pull request in cf5c3bb.