Fix batched Cholesky OOM on ROCm by bypassing hipSolver's external hipMalloc#717
Open
FlemingH wants to merge 43 commits intoROCm:amd-mainfrom
Open
Fix batched Cholesky OOM on ROCm by bypassing hipSolver's external hipMalloc#717FlemingH wants to merge 43 commits intoROCm:amd-mainfrom
FlemingH wants to merge 43 commits intoROCm:amd-mainfrom
Conversation
…tignore (ROCm#563) When jaxlib was built in debug more, an assertion in LLVM code that lazy-loads VHLO dialect could fire, since the code path could execute in a multi-threaded environment, and LLVM dialect repositories aren't thread safe to modify. This patch applies the same changes that upstream makes to fix this: jax-ml@48c8762 (this includes disabling a call to `jax_mlir_ext.enter_multi_threaded_execution(context)` in `mlir.py`. Presumably, the whole functionality related to `enter_multi_threaded_execution()` multithreaded checks isn't ready yet, and it was prematurely rolled into the production code. Manual testing
(forgot this skip in the previous PR)
Co-authored-by: Daniel Suo <danielsuo@gmail.com> Co-authored-by: Jake VanderPlas <jakevdp@google.com>
hawkinsp
reviewed
Feb 27, 2026
jaxlib/gpu/solver_kernels_ffi.cc
Outdated
| return ffi::Error::Success(); | ||
| } | ||
|
|
||
| #ifdef JAX_GPU_HIP |
There was a problem hiding this comment.
Is it possible to just reuse the CUDA case with the appropriate defines? That's usually how it works since the APIs are almost identical.
Author
There was a problem hiding this comment.
Hi @hawkinsp Thanks for your suggestion.
Updated — platform difference is now absorbed in vendor.h and solver_interface, solver_kernels_ffi.cc has no #ifdef.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
11 participants
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.
jax.vmap(jnp.linalg.cholesky) with batch >= 2 crashes with OOM on ROCm.
Root cause: JAX calls hipsolverDnXpotrfBatched (dense API, no workspace parameter), which internally allocates workspace via hipMalloc. This bypasses XLA's BFC allocator, and since XLA preallocates ~75% of GPU VRAM by default, the external hipMalloc fails.
Fix: Switch to hipsolverXpotrfBatched (standard API), which accepts an external workspace buffer. The workspace is allocated through XLA's scratch allocator, keeping all GPU memory within XLA's control. CUDA path is unchanged.