Make GPU CUDA plugin require JAX #8919
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
tengyifei
force-pushed
the
yifeit/cuda-plugin
branch
from
e83d9b5 to
1b6d9ac
Compare
Some XLA GPU features require JAX. Rather than only installing the latest version of JAX in CI, we'll just make the CUDA plugin depend on a version of JAX that's the same as what's used by PyTorch/XLA on TPU. (Except the JAX CUDA wheels).
tengyifei
force-pushed
the
yifeit/cuda-plugin
branch
from
1b6d9ac to
47d6ea1
Compare
tengyifei
force-pushed
the
yifeit/cuda-plugin
branch
from
4f623f5 to
2e67486
Compare
|
This is not easily fixable because JAX 0.6.1 started requiring CuDNN 9.8 (see https://github.com/jax-ml/jax/blob/main/CHANGELOG.md?plain=1#L61), but CuDNN 9.8 requires Debian 12 (#8928). |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This fixes internal b/419277657.
Some PyTorch/XLA GPU features require JAX. The CI tests ad-hoc install the latest version of JAX, creating a skew with the JAX version pinned by PyTorch/XLA, thus causing test failures.
Rather than only installing the latest version of JAX in CI, we'll just make the CUDA plugin depend on a version of JAX that's the same as what's used by PyTorch/XLA on TPU.
Side note, there appears to be two ways of building PyTorch/XLA for GPU. One is by setting
XLA_CUDA=1, which will cause the PyTorch/XLA C++.soto be built with CUDA support. Another is by building a "PyTorch/XLA CUDA plugin" similar to thelibtpuplugin, thus factoring CUDA-specific functionality behind a backend.so. That shows up as the "Build PyTorch/XLA CUDA Plugin" job in CI. It appears that our cloud build jobs ref build the CUDA plugin but does not upload them to GCS (discussed earlier in #8876).In any case, our GPU CI infra uses the CUDA plugin path and don't bake in CUDA support into the PyTorch/XLA native
.so. Therefore, in order to fix GPU CI, I made the CUDA plugin depend on the right versions of JAX.