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
  • 1.2. Other Noteworthy Updates
• II. Issues
  • 2.1. Top 5 Active Issues
  • 2.2. Top 5 Stale Issues
  • 2.3. Open Issues
  • 2.4. Closed Issues
  • 2.5. Issue Discussion Insights

• III. Pull Requests
  • 3.1. Open Pull Requests
  • 3.2. Closed Pull Requests
  • 3.3. Pull Request Discussion Insights

• IV. Contributors
  • 4.1. Contributors

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.

1. [BUG] [STAT:AWAITING OPENXLA-ENG] [Unit Test][Python] socket_bulk_transport_test is flaky: This issue reports that the socket_bulk_transport_test in the Python unit tests is flaky and intermittently fails with a specific error related to message reception in the SendQueue test. The failure occurs during the SendAndRecvQueuesArtificialLimit test case, and the user provides a Bazel command to reproduce the problem.

   • The comments discuss the possibility that the test failure might be related to running on an ARM architecture, which is a known issue, but the original reporter clarifies that they are running the test on an x86 Linux CPU, indicating the problem occurs outside the ARM-specific context.
   • Number of comments this week: 2

2. [BUG] [NVIDIA-GPU] Crash on GPU: host_offload_utils.cc:98 !tmp_shape_index.empty() Expected shape index to be non-empty.: This issue reports a crash occurring on the GPU platform when running a StableHLO module with the error message indicating an expected non-empty shape index failed in host_offload_utils.cc:98. The module runs successfully on CPU but fails on GPU during both execution and compile-only modes, suggesting a problem specific to GPU handling of host offload instructions in the XLA compiler.

   • 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

3. [BUG] [NVIDIA-GPU] Crash on GPU: triton_tiling_propagation.cc:627 subdim_size_accumulator == dim_size: This issue reports a crash occurring on the GPU platform when running a specific StableHLO module, triggered by a failed check in the Triton tiling propagation code related to dimension size mismatches. The module runs successfully on CPU and interpreter platforms, indicating a GPU-specific problem likely tied to the fusion and optimization passes in the GPU compiler.

   • An internal reproducer for the crash has been submitted, confirming the issue is being actively investigated.
   • Number of comments this week: 1

4. [BUG] [NVIDIA-GPU] Crash on GPU: shape_util.cc:238 array_shape.IsArray() TrueNumDimensions called on non-array 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 in shape_util.cc related to calling TrueNumDimensions on a non-array shape. The module runs successfully on CPU but fails on GPU with a SIGABRT, and the crash has been reproduced with detailed commands and environment information provided for debugging.

   • 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

5. [BUG] [NVIDIA-GPU] Crash on GPU: all_reduce_thunk.cc:87 reduction_kind.has_value(): This issue reports a crash occurring on the GPU platform when running a StableHLO module, specifically triggered by a failed check in the file all_reduce_thunk.cc at line 87 related to reduction_kind.has_value(). The problem reproduces consistently during execution and compilation steps on the GPU, despite successful conversion from StableHLO to HLO proto, indicating a potential bug in the GPU backend handling of collective operations.

   • A 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: 2

Summarized Issues:

• Test Flakiness and Message Reception Errors: The socket_bulk_transport_test in the Python codebase is flaky and intermittently fails with a specific error related to message reception during the SendAndRecvQueuesArtificialLimit test case. This instability affects the reliability of the test outcomes and complicates debugging efforts.
• issues/48239

• GpuLayoutAssignment Pass Layout Inconsistencies: The GpuLayoutAssignment pass leaves __cudnn$fusion fusion bodies internally inconsistent by not properly constraining the fusion's layout to match its interior instructions. This mismatch leads to verification failures during layout assignment in the XLA compiler, causing compilation issues.
• issues/48241

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:

• Crashes during StableHLO compilation on CPU and other backends: Multiple issues report crashes during the compilation of StableHLO modules on the CPU backend, triggered by failed type casts and unsafe fusion instruction removals. These crashes also manifest differently on GPU and other backends, indicating a broader instability in handling certain HloInstructions and fusion operations.
• [issues/47384, issues/47385]

• Fusion computation ownership and verification failures: The computation-deduplicator merges identical fusion computations into a single shared computation, violating the one-to-one ownership invariant. This causes the verifier to reject the module due to mismatched fusion instructions, breaking expected module correctness guarantees.
• [issues/48147]

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: 18

Key Open Pull Requests

1. Add mi450 backend and replace gfx1250 with mi450: This pull request adds the mi450 backend to the project and replaces the existing gfx1250 backend with mi450, including updates to documentation and tests.

• URL: pull/48355

• Associated Commits: 4ce04, e2231, 9ce61

2. [ROCm] Add ProfileOptions.advanced_configuration into the ROCm tracer (PR3): This pull request introduces the integration of the ProfileOptions.advanced_configuration into the ROCm GpuTracer, enabling layered configuration of tracing options that override hardcoded defaults and existing flags, while enhancing diagnostics reporting and adding comprehensive tests to validate ROCm profiler caps and behaviors without affecting existing CUDA code.

• URL: pull/48503

• Associated Commits: c83ae, 7a518, bdd96

3. [ROCM] MORI enable CCO-style initialization and code cleanup: This pull request enables CCO-style initialization for the MORI communicator and collectives in XLA by wiring them to the new mori::collective::CollectivesFacade, replaces the previous unimplemented stubs with fully implemented collective operations, updates the vendored MORI dependency and Bazel overlays to match the new source layout, introduces a new HIP translation unit for device-side kernels, cleans up obsolete code and dependencies, and adds a multi-process all-reduce test to validate the integration.

• URL: pull/48226

• Associated Commits: 68005, 86f2b

Other Open Pull Requests

• Bug fixes in XLA CPU backend and sort comparator handling: Multiple pull requests address stability and correctness issues in the XLA CPU backend. One fixes a crash caused by nested computations in sort comparators by pre-emitting transitively called computations, while another prevents reductions with non-trivial reducers from being sent to the YNNPACK emitter, avoiding runtime aborts.
• [pull/48160, pull/48378]

• ROCm architecture and tracer improvements: Several pull requests enhance ROCm support by adding detailed per-architecture core information tables and improving ROCm tracer configuration. These include updates to fpus_per_core and unit descriptions for CDNA architectures, refactoring tracer options for better host-buildability, and adding advanced configuration parsing with error handling.
• [pull/48195, pull/48501, pull/48502]

• Fixes for GPU compilation and target resolution: A pull request fixes a crash when compiling for SM107 GPUs on CUDA 12.9 by updating target resolution logic to fallback safely to compatible targets based on PTX version constraints, ensuring successful LLVM NVPTX compilation and program execution.
• [pull/48156]

• Loop trip-count calculation correction: One pull request fixes the trip-count calculation for integer while loops in OpenXLA, ensuring correct handling of loops where the positive step does not evenly divide the distance to the bound. This prevents undercounting and maintains arithmetic overflow safety for loop simplifications.
• [pull/48454]

• Handling of reshape and slice operations: A pull request corrects the handling of reshape and slice operations when the reshape operand has leading unit dimensions by fixing the bail condition. This prevents incorrect transpositions and regressions related to dynamic slicing along non-zero dimensions, with updated tests included.
• [pull/48404]

• HostOffloader pass variadic reduce support: One pull request fixes a compile-time crash in the XLA HostOffloader pass by properly supporting variadic (tuple-shaped) reduce operations on host-offloaded paths. Adjustments include shape dimension checks, successor tracking, and memory space validation to handle tuple outputs correctly.
• [pull/48445]

• oneDNN custom calls disabled for YNN tests: A pull request explicitly disables oneDNN custom calls for YNN-specific tests to prevent unit test failures when oneDNN is disabled by default.
• [pull/48307]

• BF16 exponential and logarithm lowering improvements: A pull request improves lowering of bf16 exponential and logarithm functions on AMD GPUs by using f32 transcendentals, reducing instruction counts and fixing subnormal input errors. This enhances performance and accuracy for bf16 workloads like softmax, attention, and GELU.
• [pull/48369]

• ROCm tarball download hash verification update: One pull request updates the ROCm tarball download process to pass an empty hash when the 'no_hash' string is provided, allowing hash verification to be skipped for frequently changing nightly and development releases. This facilitates internal testing and cleanup until hashes are published.
• [pull/48188]

• Static linking enforcement under dynamic_mode=fully: A pull request enforces static linking for specific targets even when the build is configured with dynamic_mode=fully, fixing an issue where the linkstatic=True setting was previously ignored.
• [pull/48489]

• Infrastructure OOM error simulation: One pull request introduces a simulation of an infrastructure out-of-memory error to test system behavior under failure conditions.
• [pull/48491]

• Benchmarking script cleanup: A pull request removes a leftover 60-second sleep command from the XLA run_hlo_module benchmarking script that was incorrectly included in measured runtime, improving accuracy of reported execution time without changing compiled output or behavior.
• [pull/48497]

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: 27

Key Closed Pull Requests

1. [ROCm][Triton] StableHLO lowering & end-to-end tests (PR 4/4): This pull request is the final part of a four-PR series that implements StableHLO lowering for the All-Gather collective operation using the Triton backend on AMD/ROCm, adding comprehensive end-to-end correctness and regression tests to demonstrate significant performance improvements over the RCCL backend.

• URL: pull/44564

• Associated Commits: 3e800, 72643, 1dac0, 64f81, e7db1, 28b19, 4e74e, 73ad5

2. [GPU][Triton] Reject tiles exceeding shared memory during Triton tile selection: This pull request adds a shared-memory constraint to the Triton tile-selection process to reject tile sizes that exceed the device's shared memory capacity before compilation, enabling the search to fallback to smaller tiles and preventing RESOURCE_EXHAUSTED errors during kernel compilation, with support for both legacy and experimental tiling paths and accompanying tests.

• URL: pull/46764

• Associated Commits: c613e, 21080, 584b3, 9c617, 16894

3. Fix integer overflow in IFRT proxy byte-strides size check: This pull request fixes an integer overflow vulnerability in the IFRT proxy's byte-strides size check by replacing unchecked int64_t multiplication with a safe overflow-checked multiplication in ArrayMemRegion::FromZerothElementPointer, preventing out-of-bounds reads and heap buffer overflows in gRPC handlers like MakeArrayFromHostBuffer and CopyToHostBuffer, and adds appropriate error handling and build dependencies.

• URL: pull/46604

• Associated Commits: 36967, 12451, 58c42, 439c9

Other Closed Pull Requests

• ROCm and Triton GPU Collectives Enhancements: Multiple pull requests focus on improving GPU collective operations and backend support for ROCm and Triton. These include implementing an experimental all-gather backend using collective memory, refactoring MORI GPU collectives into a unified facade for better modularity, and enabling the Triton backend for All-Gather on AMD/ROCm with performance improvements over RCCL.
• pull/46707, pull/47176, pull/44562

• Build and CI Infrastructure Improvements for ROCm and Bazel: Several pull requests address build system enhancements including dynamically selecting ROCm repository runfiles for bzlmod support in lit tests, adding infrastructure error detection with notification in ROCm pipelines, and avoiding LLVM symbol clashes in ROCm CI by using the xla_test symbol version from rules_ml_toolchain.
• pull/48140, pull/48400, pull/48341

• PowerPC64 and PowerPC64LE Build Fixes: Two pull requests fix compilation errors on PowerPC64 and PowerPC64LE architectures by including the header and refining conditional compilation guards to correctly detect _Float16 support. These changes enable successful builds on these platforms.
• pull/48221, pull/48240

• OneAPI Backend Fixes and Feature Adjustments: Pull requests address issues in the oneAPI backend including adding F64 support for matrix multiplication, temporarily disabling command buffer conversion to improve test pass rates, and fixing compile-time hangs caused by recursive walks over arith.trunci users narrowing to sub-byte types.
• pull/47656, pull/47936, pull/48202

• Triton Tile Selection and ThreadDim Extraction Enhancements: Improvements to Triton include adding a shared-memory constraint to reject tile sizes exceeding device capacity to prevent RESOURCE_EXHAUSTED errors and adding a target check for XPU in ThreadDim extraction for Intel-XPU Triton targets lacking warp specialization support.
• pull/47769, pull/47666

• Dynamic Mode and Lit Test Fixes for CI Builds: One pull request fixes lit tests and fully switches to dynamic mode during CI builds under the --dynamic_mode=fully setting to reduce network load by generating more granular cached binaries. Another related pull request proposes fully utilizing dynamic mode for CI builds to improve the build process, though it has not been merged.
• pull/48191, pull/48189

• StableHLO Regression Tests for Export Path Stability: Two pull requests add regression tests for the StableHLO module to verify correct export of stablehlo.while constructs with private function conditions containing stablehlo.case. These tests ensure the StableHLO-to-HLO export path used by hlo-translate continues to work without reintroducing previous crashes.
• pull/48198, pull/48200

• Bug Fixes in Algebraic Simplifier and Loop Analysis: Fixes include correcting the AlgebraicSimplifier’s folding of nested dynamic-slice and dynamic-update-slice operations by clamping indices properly, and improving while loop range analysis by requiring constant induction-variable steps and adding regression tests to prevent crashes.
• pull/48024, pull/47846

• Symbol and API Fixes for GPU and LLVM: Pull requests fix the PJRT C API GPU test for Intel GPUs and propose updating the cuDNN frontend to version 1.26 to support tensorIR integration. Another PR gates unsafe exponential algebraic rewrites behind a fast math flag to preserve strict floating-point semantics.
• pull/48416, pull/48168, pull/48107

• Memory Safety and Overflow Fixes: A pull request fixes an integer overflow in the ArrayMemRegion::FromZerothElementPointer function by adding checked multiplication to prevent out-of-bounds reads and heap buffer overflows in downstream gRPC handlers.
• pull/48286

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.

Access Last Week's Newsletter:

• Link