Skip to content

Conversation

@olupton
Copy link
Collaborator

@olupton olupton commented Feb 13, 2025

Previously build-jax.sh --sm all would cause XLA to emit 9.0a, which cannot be PTX JIT compiled to 10.0 and beyond.

This is due to logic in XLA:
https://github.com/openxla/xla/blob/6b470af69810d93411ac8c83e476eeee84031432/third_party/tsl/third_party/gpus/cuda/hermetic/cuda_configure.bzl#L129-L193
that it not aware of the semantics of the arch-conditional (a suffix) architectures (and a lexicographic sort where a numerical one is needed).

This means that --sm all will produce binaries that do not support pre-Turing architectures, see https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#deprecated-architectures.

nouiz
nouiz previously approved these changes Feb 13, 2025
Copy link
Collaborator

@nouiz nouiz left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe update the description that this also drop arch before and including V100 that we don't support anymore with this container?

Arch-conditional PTX cannot be JIT compiled to newer architectures.
These lists are only applies to a few kernels built into JAX/XLA, so
emitting arch-conditional code should not be important.
Also, remove deprecated support for pre-Turing architectures.
@jaro-sevcik
Copy link
Collaborator

jaro-sevcik commented Feb 14, 2025

Looks good, but it looks like I do not have permissions to approve.

Update: Scratch that, I have the permissions, I just keep getting confused about Git* UI.

@olupton olupton merged commit eac7ae4 into main Feb 14, 2025
78 of 83 checks passed
@olupton olupton deleted the olupton/tune-build-arch branch February 14, 2025 09:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants