Weekly GitHub Report for Xla: September 07, 2026 - September 14, 2026 (20:09:04)
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.
-
[CPU] [STAT:AWAITING RESPONSE FROM CONTRIBUTOR] Multiplication by max fp64 is not elementwise: This issue reports that multiplying an array by the maximum double-precision floating-point value on the CPU runtime results in infinities, whereas multiplying a single element by the same value produces a large but finite number. The user has not tested other runtimes and is seeking clarification on whether this behavior is expected or a bug.
- The maintainers requested a reproducible example to investigate the problem, but the user indicated difficulty providing one due to their use of a different language. The maintainers then tested the scenario themselves and found no issue, showing valid finite results for both scalar and vector multiplications, and asked the user to verify with the latest version before closing the issue.
- Number of comments this week: 1
-
[BUG] [NVIDIA-GPU] Crash on GPU: hlo_instruction.cc:5794 parent_ != nullptr Instruction must have a parent in order to have a unique…: This issue reports a crash occurring on the GPU platform when running a StableHLO module with the
run_hlo_modulecommand, resulting in a SIGABRT due to a failed check that an instruction must have a parent to have a unique ID. The problem reproduces consistently on GPU with both execution and compile-only modes, while the StableHLO to HLO-proto conversion succeeds without errors, indicating a potential bug in the GPU backend's handling of the HLO instructions.- The single comment notes that an internal reproducer has been submitted for further investigation, indicating that the issue is being actively worked on.
- Number of comments this week: 1
-
[BUG] [NVIDIA-GPU] Crash on GPU: hoist_fused_bitcasts.cc:168 ShapeUtil::ArrayDataSize(…) == ShapeUtil::ArrayDataSize(instruction->shape())…: This issue reports a crash occurring on the GPU platform when running a specific StableHLO module with
run_hlo_module, triggered by a failed check inhoist_fused_bitcasts.ccrelated to mismatched array data sizes during shape manipulation. The problem does not occur on the Interpreter backend and the module compiles cleanly on both CPU and GPU when run in compile-only mode, indicating a runtime-specific failure on the GPU.- The single comment notes that an internal reproducer for the issue has been submitted, indicating that the problem is being actively investigated.
- Number of comments this week: 1
-
[BUG] [CPU] Crash on CPU: abstract_cpu_buffer.cc:92 literal.size_bytes() == size: This issue reports a crash occurring on the CPU platform when running a specific StableHLO module with the
run_hlo_moduletool, triggered by a size mismatch check failure inabstract_cpu_buffer.ccat line 92. The module runs successfully on other platforms and configurations, including GPU and Interpreter backends, indicating a platform-specific problem with CPU execution of this StableHLO input.- The single comment notes that an internal reproducer for the issue has been submitted, indicating that the problem is being actively investigated.
- Number of comments this week: 1
-
[BUG] hlo-translate failure: segv:mlir::stablehlo::ReduceOp::build: This issue reports a segmentation fault occurring in the
hlo-translatetool when attempting to convert a StableHLO module to an HLO proto, specifically during the build of aReduceOp. The same module runs successfully using therun_hlo_modulecommand on various platforms, indicating the failure is isolated to the translation step rather than the module itself.- The single comment notes that an internal reproducer has been submitted for further investigation, indicating the issue is being actively worked on.
- Number of comments this week: 1
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: 8
Summarized Issues:
- XLA CPU Backend Convolution and Padding Bugs: Several issues in the XLA CPU backend involve incorrect handling of padding in convolution and reduce_window operations, leading to out-of-bounds reads, incorrect results, and memory safety violations. These bugs include improper handling of negative low padding, ignoring base_dilation parameters, and incorrect folding of negative padding into window padding, all causing incorrect outputs or crashes.
- [issues/48713, issues/48714, issues/48715, issues/48719]
- Scatter Operation Bugs in XLA: Multiple scatter-related bugs cause incorrect results due to improper handling of updates and indices. These include applying only the last update instead of combining all, dropping valid in-bounds updates due to narrow index element types, and misinterpreting unsigned indices as signed, leading to invalid index clamping.
- [issues/48716, issues/48717, issues/48718]
- Performance Table Update for GB200 Matmul: This issue involves updating the GB200 matmul performance table for the sol cost model and removing the workaround that used the B200 matmul performance table for GB200, as part of tracking a TODO in the codebase.
- [issues/48627]
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:
- Performance regressions due to buffer alignment: Increasing the minimum buffer alignment from 16 to 64 bytes in XLA causes significant CPU performance regressions for BERT and BART model inference. This change leads to additional memory copies because of unaligned numpy inputs, highlighting the trade-offs and challenges in matching Eigen's buffer alignment requirements.
- issues/18611
- GPU runtime crashes from shape handling errors: Running a specific StableHLO module on the GPU platform causes a crash due to a failed check in
shape_util.ccwhereTrueNumDimensionsis called on a non-array shape, resulting in a SIGABRT error. The same module runs successfully on the CPU, indicating platform-specific shape handling issues. - issues/47382
- Layout assignment inconsistencies in GPU fusion passes: The GpuLayoutAssignment pass fails to properly constrain fusion layouts to match interior instructions, leaving __cudnn$fusion fusion bodies internally inconsistent. This causes verification failures during layout assignment in the XLA compiler, affecting GPU fusion correctness.
- issues/48241
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: 20
Key Open Pull Requests
1. Test dpx mode with isolation: This pull request introduces and tests a new DPX mode with CPU and GPU isolation by adjusting parallel job counts, enforcing test re-execution, and modifying the test runner to ensure isolated execution environments, thereby improving test reliability and performance measurement accuracy.
- URL: pull/48649
- Associated Commits: 05889, e2d7a, 73eb6, 818f0, b9649, faddc, 8ad52, 907b4, c8d0c, e9db2, 52161, f1b95, 3d992
2. [ROCm] Read peak memory bandwidth from amd-smi: This pull request introduces a new feature that adds the QueryPeakMemoryBandwidthGbps function to the smi_util module, making it the primary method for retrieving ROCm memory bandwidth by reading the peak memory bandwidth directly from the amd_smi firmware, thereby eliminating the need for per-GPU table entries and improving accuracy for new hardware.
- URL: pull/48722
3. [DO NOT MERGE] Test rbe remote building: This pull request is focused on testing remote building using the Remote Build Execution (RBE) system, including switching to dynamic RBE configurations to validate the build process.
- URL: pull/48776
Other Open Pull Requests
- Dependency and build fixes in XLA project: This topic covers fixes to the DWYU check by resolving external repository dependencies and module labels to avoid false missing dependency errors. It also addresses issues with multiple exporters of the same header and ensures proper attribution of includes to maintain build integrity.
- Overflow and allocation bug fixes: These pull requests fix overflow bugs in the IFRT ArrayMemRegion byte-stride span and the BFC allocator by adding overflow checks and rejecting allocation requests that would overflow. Regression tests were added to verify correct handling of these overflow boundaries.
- Bug fix in XLAConfigOptions method: This fix prevents in-place mutation of the frozen dataclass's
compiler_optionslist in theto_bazelrc_lines()method, ensuring that repeated calls do not accumulate duplicate compiler flags and that the method behaves as a pure function without side effects.
- Enhancements and bug fixes in XLA CPU backend: This set of changes enables support for scatter operations and small while loops inside sort comparators by expanding scatter lowering, adding a pass to unroll small embedded while loops, modifying thread-local buffer allocation, and replacing fatal checks with error returns to prevent compiler aborts.
- Documentation grammar corrections: This pull request fixes several grammatical errors related to the use of "a" and "an" in the project documentation.
- ROCm target collective feature support and test adjustments: This update introduces an explicit
UnimplementedErrorfor cross-host one-shot collective features on ROCm targets and marks related multiprocess tests as CUDA-only to prevent silent fallback and clarify unsupported behavior due to lack of symmetric memory support.
- Async support for oneDNN custom path on AArch64: This work-in-progress pull request adds asynchronous support for the oneDNN custom path on AArch64 within XLA, including new build wiring and plans for performance benchmarking and testing to ensure compatibility with XLA's async runtime.
- Unit test updates: This pull request updates unit tests as indicated by the title and commit message.
- Unification of oneDNN rewrite eligibility criteria: This change unifies the eligibility criteria for oneDNN dot and convolution rewrites across both primitive and graph-based paths in the XLA CPU backend to ensure consistent rewrite logic.
- oneAPI backend device-side batch-pointers kernel: This implementation adds a device-side batch-pointers kernel for the oneAPI backend, registering it as a free-function SYCL kernel to compute batch pointers on the GPU, and updates kernel loading in the SYCL executor without adding new tests.
- Triton GEMM fusion multi-CTA support: This pull request enables support for multiple CTAs in Triton GEMM fusions by fixing launch grid configuration, adding a cluster size parameter to the search space, and including unit and execution tests to ensure correctness and performance improvements.
- Median execution time reporting in hlo_runner_main: This feature adds the ability to report median execution time alongside the mean when using the
--append_profile_to_csv_fileoption, allowing users to select the statistic via a flag or environment variable to better handle outliers and random spikes.
- SPIR-V backend cube root lowering: This pull request implements the use of driver intrinsics to lower the cube root (cbrt) function for single, double, and 16-bit floating point precisions in the SPIR-V backend of the XLA GPU code generator.
- Fix out-of-bounds read in XLA CPU spatial convolution: This fix updates the
hasPadding()function to correctly handle negative padding cases, preventing out-of-bounds reads and crashes while preserving optimized vectorized paths, with added unit tests verifying the fix.
- VMM allocator overlap lookup optimization: This optimization replaces hash maps with ordered
absl::btree_mapindexes keyed by virtual address to enable more efficient exact and partial overlap checks, preserving existing range handling and adding corresponding tests and build dependencies.
- GitHub Actions runner test with dpx mode: This pull request tests the GitHub Actions runner configured with dpx mode to validate the setup and execution environment without merging into the main codebase.
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: 21
Key Closed Pull Requests
1. Add UT for AllGather Group Call: This pull request proposes adding a unit test for the AllGather group call with large inputs to detect and prevent memory access issues in certain RCCL versions when multiple AllGather operations are submitted together.
- URL: pull/48558
2. [ROCm] Migrate SMI device queries from rocm_smi to amd_smi: This pull request migrates the SMI device queries in the XLA project from the deprecated rocm_smi library to the newer amd_smi library, implementing conditional compilation to select the appropriate library based on ROCm version, restructuring the backend code for clarity, and ensuring compatibility with recent ROCm changes to improve PCIe bandwidth and xGMI topology reporting.
- URL: pull/47495
3. [ROCm] Use dpx nodes with isolated gpus: This pull request proposes switching the continuous integration (CI) system to use ROCm nodes with DPX mode enabled, which utilize isolated GPUs to increase the number of available runners and thereby improve CI response times.
- URL: pull/48663
Other Closed Pull Requests
- ROCm multi-GPU execution fix: This pull request reverts a previous change that caused ROCm multi-GPU executions involving
CollectiveKernelThunkto abort due to improper handling ofMemorySpace::kCollectiveallocations. The revert restores the original behavior to prevent silent crashes in continuous integration.
- AMX_FP8 CPU feature addition: This pull request adds the AMX_FP8 CPU feature definition and CPUID detection to the TensorFlow Support Library (TSL) to resolve build errors caused by missing AMX_FP8 references. This step enables subsequent integration of related upstream changes.
- GPU kernel tiling test enablement on theRock: This pull request enables the GpuKernelTilingTest.ReductionInputTooLarge test on theRock platform by aligning reduction dimension behavior with CUDA. It addresses a bug and adds relevant tests to ensure correctness.
- oneAPI 2026 compiler bug workaround: This pull request adds a guard to disable SYCL free functions in oneCCL to avoid a bus error caused by the oneAPI 2026 compiler. This temporary workaround facilitates migration to oneAPI 2026.
- ROCm test script update: This pull request replaces the deprecated
rocm-smiCLI with theamd-smicommand in ROCm test scripts to ensure compatibility and accurate GPU enumeration inrun_xla.shandrun_xla_multi_gpu.sh.
- NVPTX backend PTX version enforcement: This pull request fixes a fatal LLVM error by enforcing a minimum PTX version of 8.7 for the Blackwell sm_120 architecture in the NVPTX backend. This ensures compatibility with older CUDA toolkits that previously limited PTX to version 8.5.
- Peer-visible atomics support and fallback: This pull request introduces a
has_peer_visible_atomics()predicate to gate XLA's hand-written collective kernels on devices supporting peer-visible atomics. It prevents hangs on gfx90a hardware by falling back to library collectives and removes obsolete fine-grained device allocation paths in RocmExecutor.
- Bypass hash verification for ROCm tarballs: This pull request introduces a 'no_hash' string to bypass hash verification when downloading frequently changing nightly and development ROCm tarballs. This allows internal runs to proceed without waiting for published hashes.
- Fix compile-time crash in HostOffloader pass: This pull request fixes a compile-time crash by properly handling variadic (tuple-shaped) reduce operations on host-offloaded paths. It ensures correct shape and memory space checks and adds tests covering these cases.
- Static linking enforcement under dynamic_mode=fully: This pull request enforces static linking for specific targets even when the build is configured with
dynamic_mode=fully. It addresses an issue where static linking was ignored after switching ROCm to this mode to reduce CI network load.
- Infrastructure error simulation for OOM: This pull request introduces a simulation of an out-of-memory (OOM) infrastructure error to help test and improve system error handling.
- Benchmark script sleep removal: This pull request removes a leftover 60-second sleep command from the XLA
run_hlo_modulebenchmarking script that was incorrectly included in the measured runtime. This improves benchmark timing accuracy without affecting execution or output.
- FFT operations treated as elementwise for partitioning: This pull request proposes treating FFT operations as elementwise when all transformed dimensions have a tile count of one. This preserves batch sharding along non-transformed dimensions and avoids unnecessary data replication and collectives in the SPMD partitioner.
- Fix Triton scaled-dot operation bug: This pull request fixes a bug where scales on certain operand types were silently dropped in the Triton scaled-dot operation, leading to incorrect results. It introduces validation and decomposition logic to ensure only provably all-one scales are allowed, particularly fixing ROCm environment issues.
- XPU Triton hook update to upstream Triton commit: This pull request updates the XPU Triton hook in XLA to align with a specific upstream commit from the Triton repository, ensuring consistency between implementations.
- Unmerged Docker failure simulation: This pull request proposes changes to simulate a Docker failure scenario but was not merged into the project.
- Unmerged XPU Triton hook update to LLVM commit: This pull request proposes updating the XPU Triton hook in the XLA GPU oneAPI codebase to align with a specific upstream LLVM commit but has not been merged.
- Unmerged remote build test template: This pull request is a test for remote building processes and is marked as not to be merged, containing a template for summarizing changes, justifications, contribution types, and testing requirements.
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 |
|---|---|---|---|---|
| alekstheod | 51 | 8 | 0 | 3 |
| draganmladjenovic | 13 | 0 | 0 | 17 |
| mfrancepillois | 15 | 2 | 0 | 7 |
| kodlan | 19 | 3 | 0 | 2 |
| nurmukhametov | 15 | 4 | 0 | 0 |
| cj401-amd | 9 | 3 | 0 | 6 |
| ScXfjiang | 15 | 1 | 0 | 1 |
| hsharsha | 9 | 4 | 0 | 0 |
| akhilgoe | 7 | 4 | 0 | 0 |
| gaurides | 9 | 2 | 0 | 0 |
Access Last Week's Newsletter: