# Changelog

## Baseline

- First survey — no prior run to compare against. CAI 51.

## Lenses

- Code Health 100
- Maturity 73
- Readiness 32
- Security 58

## Changes since last survey

- 300 commits — 257 feature/other, 43 fixes

## By area

- jax/_src — 144 commits
- (root) — 33 commits
- .github/workflows — 31 commits
- tests/pallas — 18 commits
- jaxlib/mosaic — 16 commits
- jax/experimental — 15 commits
- docs/new_docs — 6 commits
- ci/run_bazel_test_tpu.sh — 3 commits
- jaxlib/pytree.cc — 3 commits
- jaxlib/tools — 3 commits
- build/requirements_lock_3_12.txt — 2 commits
- build/rocm — 2 commits
- jax/numpy — 2 commits
- jax_plugins/rocm — 2 commits
- jaxlib/gpu — 2 commits
- tests/api_test.py — 2 commits
- tests/custom_api_test.py — 2 commits
- tests/key_reuse_test.py — 2 commits
- tests/lax_numpy_indexing_test.py — 2 commits
- docs/notebooks — 1 commit

## Notable commits

- fix: Add back `wrap_compute_type_in_place` cause `ffi_call`'s need to be choose the correct platform (is my guess). This will unblock the breakages and I'll figure out a nicer fix forward
- fix: Fix checksum mismatch.
- fix: Fix cuDNN packed-attention kv_offsets for GQA/MQA
- fix: Fix cudnn attention batching rules for operands without the vmap axis
- fix: Fix cudnn fp8 attention batching rules
- fix: Fix link in doc comment.
- fix: Fix wrapped_fun interacts with update_wrapper. Lead to failures because of jaxtyping
- fix: Fixed as_torch_kernel and as_torch_gpu_kernel Description: - Call EnsureLLVMNVPTXTargetIsRegistered in MosaicGpuCompile otherwise GetCudaComputeCapability is failing with `Unable to find target for this triple (no targets are registered)` - Recoded (with the help of AI) the compiled kernel launch as C function following compiled host launcher signature and without using three-entry wrapper argument array. Otherwise, data pointers passed into the C functions were incorrect: ``` mosaic_gpu_init_tma_desc, call cuTensorMapEncodeTiled - rank=1 - tma_sizes=128 - tma_strides=1 - tma_window_shape=128 - element_strides=1 - swizzle=0 - tma_desc=-745308928 - data_type=2 - base_addr=0x1   <----- HERE ```
- fix: Merge pull request #39413 from kodlan:fix/issue-20243-bidi-matmul-pad-add
- fix: Merge pull request #39427 from kodlan:fix/issue-13027-pytree-unhashable-metadata
- fix: Merge pull request #39450 from kodlan:fix/issue-7465-vmap-kwargs-in-axes
- fix: Merge pull request #39451 from kodlan:fix/issue-38495-cudnn-attention-jacobian
- fix: Merge pull request #39452 from kodlan:fix/issue-38495-fp8-attention-batchers
- fix: Merge pull request #39508 from kodlan:fix/issue-7222-slice-index-error-message
- fix: Merge pull request #39512 from kodlan:fix/issue-34685-cudnn-dbias
- fix: Merge pull request #39521 from bzantium:fix-packed-kv-offsets-gqa
- fix: Merge pull request #39557 from kodlan:fix/issue-2303-grad-scalar-error-message
- fix: Merge pull request #39596 from kodlan:fix/issue-37391-remat-pass-disable-flags
- fix: Merge pull request #39640 from vfdev-5:fix-_compile_as_torch_gpu_kernel
- fix: Merge pull request #39655 from ashahba:fix/oneapi-workflow-permissions-it
- …and 280 more