Skip to content

CI: 04/25/25 upstream sync #385

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1,013 commits into
base: rocm-main
Choose a base branch
from

Conversation

rocm-repo-management-api-2[bot]
Copy link

Daily sync with upstream

marksandler2 and others added 30 commits April 14, 2025 19:11
PiperOrigin-RevId: 747663692
PiperOrigin-RevId: 747782436
Not all of these have replacements, but I want to mark these as deprecated now so they can be removed in a future release.

PiperOrigin-RevId: 747867544
Shift ops expect RHS to have the same type as RHS, but a scalar was being used for RHS.

PiperOrigin-RevId: 747934619
Reverts b336daf

PiperOrigin-RevId: 747988862
* A garbage collected type bound by nanobind is allocated with `PyType_GenericAlloc` (https://github.com/wjakob/nanobind/blob/8f245bd6e5544ff828beb46435af24b8769bfcdc/src/nb_type.cpp#L81).
* `PyType_GenericAlloc` calls `_PyObject_GC_TRACK` (https://github.com/python/cpython/blob/884df116d79b05d9342e05e50484d61c684ecb8b/Objects/typeobject.c#L2357), and as soon as an object is tracked by the GC, it may be traversed (https://docs.python.org/3/c-api/gcsupport.html#c.PyObject_GC_Track).
* However, at this point, the nanobind-bound object has been allocated, but its C++ constructor need not yet have run. The constructor runs when `__init__` is called.
* In tp_traverse methods of nanobind-bound classes, we should test whether the instance is ready before visiting any C++ state. Instance readiness will be set at the end of the dispatch of `__init__` (https://github.com/wjakob/nanobind/blob/8f245bd6e5544ff828beb46435af24b8769bfcdc/src/nb_func.cpp#L857).

In addition, all nanobind-bound classes have heap-allocated types. Make sure we visit `Py_TYPE(self)`.

PiperOrigin-RevId: 747998575
…Array` data

Generally, we want to maintain that key data backing a `PRNGKeyArray` is a `jax.Array`. This change converts NumPy arrays on construction.

Co-authored-by: Yash Katariya <[email protected]>
PiperOrigin-RevId: 748077900
The first letter was inadvertently made lower-case in the previous re-naming CL.

PiperOrigin-RevId: 748086763
When we run the program with "--xla_jf_bounds_check=true", we can selectively disable bounds checks for pallas kernels now.

PiperOrigin-RevId: 748193719
yashk2810 and others added 28 commits April 24, 2025 12:07
…pjit_test.py and array_test.py so we keep testing it until we fully delete it.

PiperOrigin-RevId: 751083672
… kernel. It's implemented as split dq and dkv kernels to have enough SMEM for double-buffering.

The compute throughput is better than in the forwards pass, which is expected because there's less vector ops since we're not computing the softmax.

PiperOrigin-RevId: 751084141
And do other things...

Co-authored-by: Yash Katariya <[email protected]>
…g so we don't get the internal name for it in errors

PiperOrigin-RevId: 751113771
PiperOrigin-RevId: 751139503
PiperOrigin-RevId: 751166905
…slice_p with Element block shapes

PiperOrigin-RevId: 751173626
This is like the scan version of jax-ml#27970, which applied to while_loop.

Fixes google/flax#4709
Co-authored-by: Matthew Johnson <[email protected]>
PiperOrigin-RevId: 751197878
PiperOrigin-RevId: 751212769
* Also add Python pipeline emitter support

PiperOrigin-RevId: 751228049
@rocm-repo-management-api-2 rocm-repo-management-api-2 bot requested a review from a team as a code owner April 25, 2025 06:02
@rocm-repo-management-api-2 rocm-repo-management-api-2 bot enabled auto-merge (rebase) April 25, 2025 06:02
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.