Jonatan Kłosko

@jalberto I updated to the latest XLA revision and EXLA main already uses that. I tried building with ROCm 5.7, but there were errors indicating that XLA already assumes 6.0...

@jalberto is it when loading the precompiled binary or during build?

As a sanity check, try without `XLA_ARCHIVE_URL`, which by default should download just the CPU-enabled binary. This way we will know if it is specific to the ROCm binary. Make...

I see, I have no idea where this LLVM error is coming from, I didn't find `x86-disable-avoid-SFB`, nor `X86AvoidStoreForwardingBlocks` in openxla/xla source mentioned explicitly. You can try building youtself with...

@monorkin it may not be related, but the only thing I can think of is to also set `export ROCM_PATH="/opt/rocm-6.0"` (or whatever the version is). > Is there another way...

`-Qunused-arguments` is a clang flag, which we now pass: https://github.com/elixir-nx/xla/blob/52ea2e8cffcbea0775b4a4b3056bb2e61bb8c727/lib/xla.ex#L349 Given `gcc: error:`, it looks like it still tries to use gcc, I'm not sure why. The stacktrace says `external/zlib/BUILD.bazel:5:11`,...

We already set those env vars: https://github.com/elixir-nx/xla/blob/52ea2e8cffcbea0775b4a4b3056bb2e61bb8c727/lib/xla.ex#L343-L344 The snippets mentions `$ ./build.sh rocm`, which builds inside Docker. We should fix the Docker build sooner or later, but unfortunately it's not...

Makes sense, `XLA_TARGET=rocm` is what you need!

@jschoch unfortunately this looks like a bug upstream (XLA). The best chance for getting those fixed is narrowing down the computation that causes the issue, then finding a reproduction in...

Axon provides a way to quantize a model (as described in the quantization section in [this article](https://dockyard.com/blog/2024/08/20/where-are-nx-axon-bumblebee-headed), however that requires loading the whole model first. Ideally we would be able...