Weekly GitHub Report for Xla: July 20, 2026 - July 27, 2026 (22:04:33)
Weekly GitHub Report for Xla
Thank you for subscribing to our weekly newsletter! Each week, we deliver a comprehensive summary of your GitHub project's latest activity right to your inbox, including an overview of your project's issues, pull requests, contributors, and commit activity.
Table of Contents
I. News
1.1 Recent Version Releases:
No recent version releases were found.
1.2 Version Information:
Please provide the version release information you would like me to analyze and summarize.
II. Issues
2.1 Top 5 Active Issues:
We consider active issues to be issues that that have been commented on most frequently within the last week. Bot comments are omitted.
- [BUG] collective-permute cycle decomposer miscompiles permutes with more than one cycle: This issue describes a miscompilation in the
CollectivePermuteCycleDecomposerwhere permutes containing more than one cycle are incorrectly handled, causing all but one back-edge receiver to output zeros instead of the expected permuted data. The root cause is that the code uses only the last back-edge target in a scalar equality predicate, which fails to correctly select all back-edge receivers in multi-cycle permutations, leading to incorrect routing and output.- The comment confirms the diagnosis and reports a fix that replaces the single back-edge target comparison with a disjunction over all back-edge targets, updating the test expectations accordingly; the fix is pending CI validation due to local build issues.
- Number of comments this week: 1
Since there were fewer than 5 open issues, all of the open issues have been listed above.
2.2 Top 5 Stale Issues:
We consider stale issues to be issues that has had no activity within the last 30 days. The team should work together to get these issues resolved and closed as soon as possible.
As of our latest update, there are no stale issues for the project this week.
2.3 Open Issues
This section lists, groups, and then summarizes issues that were created within the last week in the repository.
Issues Opened This Week: 7
Summarized Issues:
- XLA compilation failures with TensorFlow ops involving dynamic or zero-sized dimensions: Several issues report XLA compilation errors when TensorFlow operations are applied to outputs with dynamic or zero-sized dimensions, where eager execution succeeds but XLA fails due to rank inference errors, invalid argument errors, or unimplemented features. These failures occur in contexts such as tf.bitcast on tf.raw_ops.Where outputs, tf.linalg.matrix_transpose on zero-sized embeddings, tf.image.pad_to_bounding_box on SelfAdjointEigV2 outputs, and tf.linalg.svd on zero-sized embedding outputs.
- issues/45967, issues/45969, issues/45971, [issues/45972](https://github.com/issues/45972]
- XLA compilation inconsistency with tf.raw_ops.Cast using Truncate attribute: One issue highlights that using tf.raw_ops.Cast with the Truncate=True attribute to cast from float32 to int8 works in eager mode but fails in XLA compilation due to an UnimplementedError, showing a discrepancy between TensorFlow eager execution and XLA lowering regarding supported attributes. This reveals a limitation in XLA's handling of certain cast operations that are otherwise valid in eager execution.
- issues/45973
- Potential CUDA memory safety issue in CheckRedzones function: There is a reported potential undefined behavior in CUDA memory management where cuMemFreeHost is called on memory that may still be undergoing an asynchronous memset operation without proper stream synchronization, which could lead to memory safety problems. This indicates a need for improved synchronization to ensure safe memory operations in CUDA within the XLA codebase.
- issues/46093
- Proposal for portable autotunable GPU backend for XLA kRaggedDot operation: A feature proposal suggests adding a portable, autotunable GPU code-generation backend for the XLA kRaggedDot operation using the Triton/XTile tiling pipeline, aiming to unify CUDA and ROCm paths with full autotuning and scheduling integration. This enhancement would enable improved fusion, per-shape tuning, and future code generation improvements while keeping existing vendor library defaults as opt-in options.
- issues/46127
2.4 Closed Issues
This section lists, groups, and then summarizes issues that were closed within the last week in the repository. This section also links the associated pull requests if applicable.
Issues Closed This Week: 3
Summarized Issues:
- JIT Compilation TPU Runtime Errors: A bug during JIT compilation on TPUs causes a communication channel to be incorrectly reused for multiple host send/recv instructions, leading to an internal XLA runtime error. This error disrupts training workflows with JAX and Equinox across various TPU versions despite previously stable code and environments.
- issues/7904
- Incorrect Alpha Parameter Computation in GemmRewriter: The GemmRewriter incorrectly folds scalar multiplication into the cuBLASLt call's alpha parameter without considering an existing BIAS epilogue, resulting in a wrong computation of alpha * (A@B + bias) as alpha * (A@B) + bias. This causes a deterministic and silent offset in the output, affecting numerical correctness.
- issues/46110
- Performance Regression in GPU Scatter-Add: A regression in jaxlib versions 0.10.2 and 0.11.0 causes the f32 scatter-add operation on GPU to compile into a compare-and-swap retry loop instead of a native float atomic add. This results in a severe performance degradation of approximately 450x slower under index contention in common embedding-gradient patterns.
- issues/46225
2.5 Issue Discussion Insights
This section will analyze the tone and sentiment of discussions within this project's open and closed issues that occurred within the past week. It aims to identify potentially heated exchanges and to maintain a constructive project environment.
Based on our analysis, there are no instances of toxic discussions in the project's open or closed issues from the past week.
III. Pull Requests
3.1 Open Pull Requests
This section provides a summary of pull requests that were opened in the repository over the past week. The top three pull requests with the highest number of commits are highlighted as 'key' pull requests. Other pull requests are grouped based on similar characteristics for easier analysis. Up to 25 pull requests are displayed in this section, while any remaining pull requests beyond this limit are omitted for brevity.
Pull Requests Opened This Week: 27
Key Open Pull Requests
1. [GPU][Triton][RaggedDot] End-to-end integration tests (PR 8/8): This pull request adds comprehensive end-to-end integration tests for the Triton backend implementation of the ragged-dot operation in the XLA GPU compiler, covering backend selection, tile propagation, tiling space, and emission correctness across all ragged-dot variants and hardware targets, as the final step in a series enabling efficient Triton-based ragged-dot support.
- URL: pull/46002
2. [XLA:GPU][oneAPI] Integrate oneDNN matmul for SYCL backend: This pull request integrates oneDNN matrix multiplication support for the SYCL backend in the XLA GPU project, including the addition of relevant utility files and tests, and depends on a prior pull request to complete the sequence.
- URL: pull/46181
3. [NVIDIA GPU] Add a new hlo op CollectiveReduce: This pull request introduces a new High-Level Optimization (HLO) operation called CollectiveReduce to the NVIDIA GPU backend, with detailed motivation provided in the linked discussion, and notes that the runtime implementation will be added later while also integrating this operation into StableHlo.
- URL: pull/46104
Other Open Pull Requests
- Triton backend for ragged-dot operations: Multiple pull requests introduce and enhance the Triton backend for ragged-dot operations in XLA GPU, including adding new functions for processing ragged-dot tiling, implementing various ragged-dot modes (NonContracting, Contracting, Batch), and enabling autotuning and tile propagation strategies. These changes collectively improve performance and efficiency over existing backends by providing more generic and optimized handling of ragged-dot computations across different GPU targets.
- GPU backend performance improvements and bug fixes: Several pull requests address performance and correctness in the XLA GPU backend, including fixing dynamic-slice offset bugs in offloaded scan residuals, improving DotMerger canonicalization to preserve minor dimension order for better GPU fusion, and marking streamed-once global loads/stores as non-temporal to optimize cache usage on AMD GPUs. These changes enhance runtime efficiency and correctness, with some providing up to 10% faster execution on specific hardware.
- ROCm and AMD GPU integration and improvements: Multiple pull requests add and improve ROCm and AMD GPU support, including foundational AMD MORI collectives backend integration, enabling XF32 hipblaslt kernels for f32 GEMMs, and fixing ROCm XLA profiler stream_id tracking for accurate GPU event assignment. These contributions enhance collective communication, kernel performance, and profiling accuracy on AMD platforms.
- Collective communication and lowering improvements: Pull requests refactor collective lowering and async execution management in XLA:GPU, add stage-dependent context decoding for collective communication in XLA FFI, and improve handling of grouped collectives and device send/recv operations. These changes improve code structure, documentation, and functionality for collective operations in the GPU backend.
- Security and correctness fixes: A pull request adds explicit validation in
BufferAllocation::AddAssignmentto prevent negative buffer offsets and sizes, addressing a security and correctness issue that could cause out-of-bounds memory accesses during LLVM IR code generation. This fix ensures safer deserialization and memory handling in the compiler pipeline.
- Bug fix for DecomposeCollectivePermuteCycle: One pull request fixes the
DecomposeCollectivePermuteCyclefunction to correctly handle multiple disjoint cycles by combining back-edge target comparisons with a logical OR, ensuring correct data permutation for all devices involved. This resolves a bug where only the last back edge target was considered, preventing incorrect zero data returns.
- oneDNN library upgrade: A pull request upgrades the oneDNN library to version 3.12.3 for the XLA CPU backend, including improvements in asynchronous support. This update enhances CPU backend capabilities and performance.
- Triton TMEM scale padding workaround: A pull request restores a temporary workaround for the Triton TMEM scale padding issue to maintain CUDA 12.8 compatibility, addressing a test failure on Blackwell (sm_100) GPUs until the upstream fix can be re-enabled.
- GPU specification enhancements: One pull request adds a new Python module
_gpu_specand exposes theget_gpu_specmethod along withGpuTargetConfigandGpuDeviceInfoclasses to improve GPU specification capabilities.
- Bazel server shutdown fix: A pull request implements a cleanup to ensure the bazel server is properly shut down after job completion, addressing docker termination issues potentially caused by the bazel server running during tmpfs unmount.
- SYCL matmul utility addition: A pull request adds a SYCL matmul utility file to provide utility functions for GEMM operations on the SYCL platform using oneDNN, as part of a sequence of related changes.
3.2 Closed Pull Requests
This section provides a summary of pull requests that were closed in the repository over the past week. The top three pull requests with the highest number of commits are highlighted as 'key' pull requests. Other pull requests are grouped based on similar characteristics for easier analysis. Up to 25 pull requests are displayed in this section, while any remaining pull requests beyond this limit are omitted for brevity.
Pull Requests Closed This Week: 64
Key Closed Pull Requests
1. [XLA:GPU] Add collective-reduce HLO and lowering for Layered FSDP backward: This pull request introduces a new reduce-to-root HLO collective and its corresponding SPMD lowering and GPU runtime support to efficiently implement the backward pass communication pattern for Layered-FSDP in Muon training, enabling gradient reductions that materialize only on the dynamically determined owner rank per layer, thereby optimizing multi-GPU distributed training by avoiding unnecessary all-reduce operations.
- URL: pull/44799
- Associated Commits: 42deb, 77668, cf5ab, 70b3e, f44fb, 2987e, 76f28, b9089, 77679, 2fbb9, 7d3ca, 2d8ba, 41a7f, a97dd
- Associated Commits: 42deb, 77668, cf5ab, 70b3e, f44fb, 2987e, 76f28, b9089, 77679, 2fbb9, 7d3ca, 2d8ba, 41a7f, a97dd
2. SPMD: Use collective-broadcast for replicated dynamic-slice from sharded operands: This pull request introduces a specialized SPMD lowering in XLA that replaces the inefficient all-gather-then-slice pattern for replicated dynamic-slice operations on operands sharded along the sliced dimension with a more efficient per-owner collective-broadcast approach, significantly reducing communication overhead in Layered-FSDP workloads by broadcasting only the selected layer slice from its owning partition to all others.
- URL: pull/44796
- Associated Commits: 82c7d, ead5f, 9b27a, ef83c, 726bd, 6d3c5, 9bbb6, 8fc79, 529e3, 985ae, 74187, 3fe5b
- Associated Commits: 82c7d, ead5f, 9b27a, ef83c, 726bd, 6d3c5, 9bbb6, 8fc79, 529e3, 985ae, 74187, 3fe5b
3. [XLA:GPU] Add device-initiated ragged all-to-all kernel using NCCL LSA/GIN: This pull request introduces an opt-in device-initiated ragged all-to-all GPU kernel for XLA using NCCL 2.29 primitives (LSA memcpys and GIN put+signal) to perform the entire collective communication in a single CUDA launch, eliminating host coordination, improving bandwidth through vectorization, and supporting both local-only and mixed local/remote topologies.
- URL: pull/41903
Other Closed Pull Requests
- GPU command buffer virtual address remapping improvements: Multiple pull requests enhance GPU command buffer VA remapping by introducing adaptive policies, allocation ID tracking, persistent mappings, and reuse of stale allocations to reduce overhead and improve efficiency. These changes collectively optimize memory management, reduce driver calls, and simplify command-buffer thunk responsibilities.
- pull/43869, pull/41626, pull/43249, pull/43400, pull/45597
- GPU latency hiding and asynchronous D2D memory copy scheduling: Several pull requests introduce and improve GPU scheduling rules that selectively overlap asynchronous device-to-device memory copies with compute-bound kernels to enhance latency hiding and avoid memory bandwidth contention. These include new scheduling policies, thunk emitter fixes, and pre-scheduling HLO passes to enable asynchronous D2D copies with feature flags for safe rollout.
- pull/46151, pull/45983, pull/46106, pull/45082
- Collective communication and backend enhancements: Pull requests add initial support for oneCCL collective communicators, implement runtime dispatch for the Triton backend in All-Gather operations, and introduce stable memory address features for collective buffers to prevent corruption. These changes improve collective operation performance and memory stability in the GPU backend.
- pull/44406, pull/44560, pull/46029
- XLA GPU backend code refactoring and utility improvements: Multiple pull requests refactor GPU output buffer management, extract common cuDNN fusion utilities, and improve command buffer runtime handling to unify policies and enhance maintainability. These efforts consolidate logic and improve code reuse without affecting performance.
- pull/45478, pull/44288, pull/45707
- oneAPI and SYCL backend support enhancements: Pull requests add a utility file for oneDNN integration with SYCL, implement the oneAPI compute capability interface, enable the ScaledDotRewriter pass unconditionally, and add DMA Map support on Intel GPUs to improve performance and feature completeness. These changes advance oneAPI backend capabilities and GPU memory management.
- pull/44915, pull/46158, pull/45684, [pull/45954](https://github.com/openxla/xla/pull/45954], pull/45693
- XLA GPU plugin and API improvements: A pull request adds support for the
max_inflight_computationsoption in the PJRT C API GPU plugin, enabling configuration from Python and preventing deadlocks. This brings GPU plugin functionality closer to parity with the TPU plugin. - pull/45684
- Bug fixes and correctness improvements: Pull requests fix the
InferDynamicUpdateSliceShapefunction to correctly handle unbounded dynamic operand dimensions and skip compilation/loading of empty constants LLVM modules to avoid unnecessary backend work. These fixes improve correctness and efficiency in the GPU backend. - pull/45263, pull/45877
3.3 Pull Request Discussion Insights
This section will analyze the tone and sentiment of discussions within this project's open and closed pull requests that occurred within the past week. It aims to identify potentially heated exchanges and to maintain a constructive project environment.
Based on our analysis, there are no instances of toxic discussions in the project's open or closed pull requests from the past week.
IV. Contributors
4.1 Contributors
Active Contributors:
We consider an active contributor in this project to be any contributor who has made at least 1 commit, opened at least 1 issue, created at least 1 pull request, or made more than 2 comments in the last month.
If there are more than 10 active contributors, the list is truncated to the top 10 based on contribution metrics for better clarity.
| Contributor | Commits | Pull Requests | Issues | Comments |
|---|---|---|---|---|
| shawnwang18 | 132 | 14 | 0 | 8 |
| alekstheod | 60 | 4 | 0 | 3 |
| akuegel | 0 | 0 | 0 | 47 |
| mfrancepillois | 26 | 8 | 1 | 0 |
| EylonKrause | 18 | 5 | 1 | 0 |
| ezhulenev | 13 | 6 | 0 | 3 |
| pemeliya | 15 | 1 | 0 | 5 |
| akhilgoe | 12 | 2 | 0 | 5 |
| kanvi-nervana | 14 | 5 | 0 | 0 |
| nurmukhametov | 10 | 7 | 0 | 0 |