Ahmed Taha

Standardized CUDA kernel annotations (#6532) by applying the existing CUML_KERNEL macro to the 16 remaining C++ kernels; six of them gain internal linkage, so libcuml.so no longer exports externally linked kernel symbols, as confirmed by RAPIDS detect-weak-linking.

Added GDB and LLDB pretty-printers for cuda::hierarchy in libcu++ (#10089) that report each level's bottom unit and dimensions, reading static extents from type information and dynamic extents from stored values without calling into the debugged program; tests cover partial hierarchies, derived descriptors, and live updates.

Made CUB's DeviceReduce::ReduceByKey run-to-run deterministic (#9995): the public dispatch now opts into the stable cross-tile ordering already supported by TilePrefixCallbackOp, so FP64 segments spanning multiple tiles combine their partials in a fixed order and identical calls return bit-identical aggregates, while RLE and the deprecated dispatcher keep their previous behavior; a regression test checks the reported reproducer bit-for-bit across repeated calls.

Made --stopping-criterion reject conflicting options (#441) that would silently discard parameters or overwrite a criterion set earlier in the same scope, while still letting benchmark-local options override inherited global settings.

Reworked per-run MSA output handling (#31) in OpenFold-3 so each run gets a unique temporary workspace: raw ColabFold records are saved as soon as the server response validates and OpenFold-formatted alignments once processing finishes, with new save_openfold_outputs, save_colabfold_outputs, and colabfold_output_dir settings, plus cleanup that can no longer delete explicit output directories.

Moved OpenFold-3 inference intermediates (#342) out of /tmp and under <output_dir>/openfold3_intermediates, so temporary MSA and template files left by a crashed cluster job stay next to that job's outputs; directories created by a run are removed after normal completion or a handled error.

Annotated OpenFold-3 inference configurations (#342), split out of #363 at a maintainer's request: typed InferenceDataModule.inference_config as InferenceJobConfig, narrowed InferenceExperimentRunner to InferenceExperimentConfig, and removed the 17 matching mypy-baseline.txt entries without changing runtime behavior.

Added an optional residue_name for SMILES ligands (#323) in OpenFold-3 without changing the LIG# defaults: explicit names are trimmed, uppercased, and validated as ASCII alphanumeric, rejected on collisions with CCD, non-canonical, or standard polymer residue names, and carried into the SMILES component mapping, with full names written to mmCIF and three-character truncation documented for legacy PDB output.

Fixed assigned C=C planarity on OpenFold-3's default inference path (#136): RDKit-assigned E/Z, non-aromatic sp2 C=C bonds are detected and their heavy-atom dihedrals pulled to the assigned cis/trans plane by a flat-bottom correction late in diffusion, with one more projection after rollout so the sampler cannot undo it, while queries without matching bonds are left unchanged; the worst targeted dihedral deviation fell from 101.13° to 15.01°.

Stopped compiled Domino cross entropy from materializing full-vocabulary logits (#190718): the AOTAutograd partitioner now recomputes the cheap cat/log-softmax path in backward while keeping small row statistics and expensive projections saved; on the reported BF16 shape, compiled time drops from 5.25 ms to 3.69 ms and temporary memory from 1,940 MiB to 0.1 MiB.

Let the Inductor scheduler fuse compatible sibling CUDA reductions (#190718) when shared direct reads cover at least half of the larger reduction's input traffic, with indirect reads and disjoint slices excluded; split out of #193121 at a maintainer's request, and on the reported BF16 shape it cuts the vocabulary kernels from three to two, taking compiled time from 3.69 ms to 2.64 ms.

Added complex-input support to jax.scipy.special.gamma using the Lanczos approximation and reflection formula in Google JAX.

Made image resizing a JAX primitive (#8452): resize operations now lower through a dedicated image_resize primitive that keeps identity resizes as no-ops and uses the existing implementation as its default lowering, while preserving batching, autodiff, explicit sharding, Checkify, Jet, key-reuse, and BCOO sparsify behavior.

Fixed an azure-data-tables bug (#46014) where unsupported TableClient.list_entities(...) kwargs leaked into the async transport and surfaced as an aiohttp TypeError; added early SDK-level validation for query_filter/parameters with clear guidance to use query_entities, plus sync/async regression tests.