Weekly GitHub Report for Pytorch: July 06, 2026 - July 13, 2026 (13:52:20)
Weekly GitHub Report for Pytorch
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:
The current version of this repository is v2.6.0
1.2 Version Information:
Released on January 29, 2025, PyTorch 2.6 introduces significant enhancements including torch.compile support for Python 3.13, a new dynamic compilation control API torch.compiler.set_stance, and improved AOTInductor packaging and ABI compatibility. Notable highlights also include FP16 support on X86 CPUs, expanded Intel GPU support, FlexAttention for X86 CPUs targeting LLMs, and a backward-incompatible security improvement flipping the default of torch.load to weights_only=True, alongside the deprecation of official Conda package publishing.
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.
-
[MODULE: CUDA] [MODULE: CI] [MODULE: TESTS] [TRIAGED] [MODULE: REGRESSION] [ONCALL: PT2] [MODULE: VLLM] [BOT-TRIAGED] [vllm] vllm-x-pytorch-test multimodal pooling tests fail with EngineCore "CUDA driver initialization failed" since Triton 3.8 bump (#188251): This issue reports that vLLM multimodal pooling model tests fail to start the EngineCore due to a "CUDA driver initialization failed" error following an update to Triton 3.8, which likely requires a newer CUDA driver than what is available on the testing runners. The problem is suspected to be caused by the Triton 3.8 toolkit bump, as downgrading to Triton 3.7.1 resolves the errors, but further investigation is needed to confirm the root cause and whether a smaller reproducer exists without vLLM.
- Comments confirm that downgrading Triton to version 3.7.1 stops the errors, and there is a request to identify a smaller reproducer to isolate the issue, with a possible related issue in the vLLM project mentioned for further context.
- Number of comments this week: 4
-
[MODULE: CI] [TRIAGED] [MODULE: INFRA] [BOT-TRIAGED] [CI][B200] B200 Runner Maintenance Notification: This issue addresses the ongoing maintenance of six B200 runners, which may cause job failures requiring restarts during the maintenance period. The discussion focuses on identifying whether specific job errors are related to this maintenance and managing the affected runner to prevent further job failures.
- The comments confirm that some job failures are linked to the maintenance on a specific runner, with suggestions to retry jobs after maintenance completion and to temporarily stop the runner daemon to prevent additional job assignments.
- Number of comments this week: 3
-
[HIGH PRIORITY] [TRIAGE REVIEW] [MODULE: BINARIES] [MODULE: ROCM] [TRIAGED] [MODULE: VECTORIZATION] [MODULE: REGRESSION] [ONCALL: PT2] [MODULE: INDUCTOR] [BOT-TRIAGED] [ROCm] torch 2.13 wheel: "Can't detect vectorized ISA for CPU" in torch.compile smoke test on non-ROCm image (regression vs 2.12.1): This issue reports a regression in the
torch==2.13.0+rocm7.2wheel where thetorch.compilesmoke test fails with a "Can't detect vectorized ISA for CPU" error when run on a non-ROCm image, despite succeeding on ROCm images and in the previous 2.12.1 version. The problem appears to stem from a change in the CPU vectorized ISA detection mechanism or packaging in 2.13, causing the probe to fail in environments lacking ROCm-specific dependencies.- The comments confirm the issue persists in nightly builds and identify a specific pull request that removed an import critical for CPU ISA detection on non-ROCm images; reverting this change restores functionality and resolves the error.
- Number of comments this week: 3
-
[TRIAGE REVIEW] [MODULE: PICKLE] [MODULE: SERIALIZATION] [SECURITY] [BOT-TRIAGED] weights_only unpickler: _check_append_target loosening (draft PR #189353) re-enables invoking an allowlisted list subclass’s overridden append/extend during unpickling, bypassing _blocklisted_modules: This issue discusses a security concern introduced by a draft pull request that loosens type checks during unpickling, allowing allowlisted list subclasses with overridden append/extend methods to execute arbitrary code, potentially bypassing module blocklists and enabling remote code execution. The reporter demonstrates that this change re-enables a method-invocation gadget that was previously blocked, and suggests invoking base list methods directly to prevent subclass overrides from running during unpickling.
- The comments clarify that the remote code execution depends on malicious code being present in the user's codebase and occurs during unpickling, with the reporter confirming that the blocklist is bypassed and the dangerous code runs in the background; further discussion highlights the risk of loading models with overridden methods and the need for careful review.
- Number of comments this week: 3
-
[MODULE: PERFORMANCE] [MODULE: BUILD] [FEATURE] [TRIAGED] [BOT-TRIAGED] [RFC] Build profiling for PyTorch: This issue proposes adding build profiling instrumentation to the PyTorch build system using recent CMake capabilities to monitor and analyze build times, aiming to detect and prevent significant build time regressions. The initial implementation shared includes detailed timing metrics for various build targets and commands, providing insights into compile and link durations.
- The comments include a mention to another contributor and a follow-up showing build profiling results with a specific configuration flag enabled, accompanied by a screenshot illustrating the profiling output.
- Number of comments this week: 2
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: 82
Summarized Issues:
- CI and Runner Failures Due to GPU and Cluster Issues: Multiple issues report persistent failures and timeouts in CI jobs and runners caused by GPU availability problems, long compilation times, and cluster outages. These include GPU state problems on B200 runners causing NCCL timeouts, MI300 cluster outages reducing capacity and causing queueing, and maintenance on B200 runners leading to job failures requiring restarts.
- Performance Regressions in CPU and GPU Kernels: Several issues describe significant slowdowns and regressions in CPU operator benchmarks and GPU kernels, including slowdowns in transcendental unary CPU ops, quantization operators, embedding functions, and Inductor convolution fusion regressions. These regressions affect both x86 and aarch64 architectures and impact model latency and throughput.
- PrivateUse1 Device and Backend Integration Proposals: A cluster of sub-RFCs propose refactoring device identity, tracing, and Inductor codegen to support PrivateUse1 (out-of-tree) backends via registry-based device interfaces. These aim to replace hardcoded device lists with extensible registries for better backend integration and dynamic device capability queries.
- Bugs and Crashes in CUDA and GPU Operations: Multiple issues report crashes, illegal memory accesses, and assertion failures in CUDA kernels and GPU operations, including CUDA graph replay crashes, caching allocator assertion failures, and illegal memory reads/writes due to integer overflows or kernel bugs. These affect various CUDA kernels such as max-pooling, bincount, im2col, embedding backward, and LU solve.
- Bugs and Limitations in PyTorch Inductor Backend: Several issues describe bugs in the Inductor backend including incorrect results from unrolled sum reduction on transposed inputs, layout optimization not applied to convolution backward causing regressions, and deadlocks due to communication operation reordering. These issues cause numerical errors, performance regressions, and hangs during compilation or execution.
- Graph Caching and Compilation Cache Issues: Problems with graph caching and compilation cache reuse are reported, including failures to generate stable cache keys due to unpicklable inputs, and near-total cache misses in large-scale distributed training due to embedded ProcessGroup names. These cause excessive recompilation and reduced performance.
- issues/189330, [issues/189541](https://github.com/issues/189541]
- Security Vulnerabilities in Unpickling and Checkpoint Loading: Security concerns are raised about unsafe deserialization in
torch.distributed.checkpointusing rawpickle.loadwithout warnings, and asymmetrical blocklists intorch._weights_only_unpicklerallowing dangerous code execution. Another issue discusses a bypass in unpickling that can enable remote code execution via overridden list methods.
- Test Failures and Disabled Tests on XPU Platform: Multiple tests related to XPU support have been disabled due to failures on the main branch, including tests for pinned memory copying and comprehensive arange operations. These indicate ongoing instability or incomplete support for the XPU platform in CI.
- Incorrect Behavior and Bugs in Tensor Operations and Serialization: Issues include a bug where subtraction of complex64 tensors fails due to serialization argument mismatches, a bug in
torch.ldexpcausing illegal memory access on fresh CUDA devices, and a bug where copying from MPS to CPU silently corrupts the source tensor under certain conditions.
- Documentation and Warning Noise Issues: Some issues report misleading warnings such as the
register_full_backward_pre_hookwarning when inputs do not require gradients, and documentation mismatches where sparse tensor index outputs are incorrectly documented as int32 instead of int64. Additionally, persistent dead or unreachable URLs cause lint failures.
- Proposals for New Features and Enhancements: Proposals include adding build profiling instrumentation to CMake, a CustomLoggerRegistry for custom profiler trace formats, a new 'symmetric' padding mode, enhanced provenance tracking in Inductor, and new
ctxAPI methods for autograd.Function to declare gradient dtypes. These aim to improve observability, extensibility, and precision.
- Deadlocks and Synchronization Issues in Distributed and CUDA Operations: Deadlocks occur in all-reduce operations issued from multiple CUDA streams due to missing synchronization channels, and stale lock files cause infinite waits during compilation with
torch.utils.cpp_extension.load(). These issues cause permanent hangs and require manual intervention.
- Runtime Errors and Assertion Failures in Compilation and Functionalization: Bugs include assertion errors when combining nested compile regions with input mutation and float8 tensors, rejection of custom operators with wildcard alias annotations during functionalization, and NameErrors due to stride mismatches in torch.cond codegen. These cause compilation failures and runtime errors.
- Issues with Artifact Uploading and Downloading in CI: Bugs in artifact handling include uploading JSON test reports as logs zip files, and downloading artifacts from mismatched workflow runs due to substring matching of job IDs, causing incorrect artifact mirroring and downstream consumer errors.
- Failures and Bugs in Specific PyTorch Operations and Tests: Failures include
torch.jit.tracefailing on MaxUnpool2d due to a TypeError, FP8 decode codegen test failures on Orin platform, and vLLM multimodal pooling tests failing due to CUDA driver initialization errors after a Triton update.
- Requests for New CUDA Version Support: There is a request to add CUDA 13.3 as an available installation option to leverage new CUDA features and updates.
- High Overhead and Performance Investigation Requests: An issue discusses high overhead in autograd.Function implementations with many inputs and outputs, opening the topic for investigation and optimization.
- Proposal for Distributed Rendezvous Backend Pluggability: A proposal suggests refactoring rendezvous functions to support multiple backends via a pluggable registration model and URL hook preprocessing, improving scalability and customization while maintaining backward compatibility.
- Incorrect Device Filter Logic in Distributed Broadcast: A bug in the device filter for sm90+ GPUs causes an incorrect workaround to be applied on older devices, with observations that the workaround may be unnecessary for Tesla V100 GPUs.
- Silent Data Corruption on Apple MPS Devices: A silent and nondeterministic corruption bug occurs in
F.linearandnn.Linearon Apple MPS devices when using bias with large 3D inputs, caused by a 16-bit batch-index wraparound in Apple's kernel, resulting in incorrect outputs without errors or warnings.
- Proposal for Schedule-Aware Multi-Stream Orchestration in TorchInductor: A discussion explores implementing schedule-aware multi-stream orchestration in TorchInductor wrapper code to automatically assign CUDA streams and synchronize independent compute branches, aiming to improve parallelism beyond user-defined stream semantics.
- Bug in Threshold OpInfo Reference Test Due to NumPy Changes: The
nn.functional.thresholdOpInfo reference test fails with an OverflowError for unsigned integer types on NumPy 2.5+ due to changes in handling out-of-range Python integers innp.where, causing cross-platform test failures.
- Bug in FxGraphCachePickler Causing Cache Bypass: FxGraphCachePickler fails to generate stable cache keys and raises BypassFxGraphCache when encountering unpicklable reference-opaque inputs like ProcessGroup, disabling graph cache and causing recompilation on every invocation.
- Bug in SampleInput Metadata Causing vmap Test Failures: The
reference_inputs_elementwise_ternaryfunction generatesSampleInputs with incorrect metadata, causing vmap exhaustive tests to fail due to broadcast-shape errors.
- Proposal to Add Test for Syncing CPython 3.13 Diff Files: A proposal requests adding a test to verify that
.difffiles intest/cpython/v3_13/remain in sync with their adapted test files by regenerating diffs from upstream tags and asserting matches to prevent silent drift.
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: 20
Summarized Issues:
- Build and CI Failures Due to External Dependencies: Several issues report build or CI failures caused by external factors such as broken links, server outages, or platform-specific incompatibilities. The nightly lint-urls job fails due to 28 broken links requiring fixes or allowlisting, while a temporary SSL certificate failure on a PyTorch download server caused build pipeline failures mitigated by fallback servers. Additionally, segmentation faults occur on Intel XPU devices due to SYCL ABI incompatibilities, and MPS CI tests fail due to inconsistent behavior of
torch.divwith nonfinite values on MPS devices. - [issues/189055, issues/189056, issues/189243, issues/189341]
- Compiler and Code Generation Bugs in Inductor and Dynamo: Multiple issues describe bugs in PyTorch's Inductor compiler and Dynamo-based exporters causing crashes, incorrect outputs, or compilation failures. These include a
KeyErrorfrom self-edges in memory dependencies duringtorch.compile, incorrect dropping of explicitdtypeincumsumoptimizations leading to wrong data types, and incorrect ONNX export ofnn.GRUattributes causing numerical discrepancies. - [issues/189225, issues/189271, issues/189518]
- Platform-Specific Backend Bugs and Deadlocks: There are platform-specific bugs causing deadlocks or incorrect results in PyTorch backends. The ROCm wheels deadlock indefinitely on hosts without AMD kfd drivers due to static constructors triggering thread pool finalization deadlocks, and the MPS backend's
torch.linalg.solvereturns incorrect results when broadcasting right-hand side tensors, unlike CPU implementations. - [issues/189110, issues/189134]
- Test Failures and Disabling on XPU Platforms: Several tests related to external constants and pinned memory are consistently failing on XPU platforms, leading to their disabling in multiple test suites. These failures affect the stability of the test suite on XPU and are reported across different test wrappers.
- [issues/189510, issues/189511, issues/189514, issues/189661]
- Documentation Improvement Request: There is a request to improve the Tanh function documentation by including example output directly in the docs, enabling users to see results without running code.
- [issues/189384]
- Unicode and Encoding Issues on Windows: A
UnicodeDecodeErroroccurs on Windows with Korean locale duringtorch.compiledue to an em-dash character in a Jinja template file being read without UTF-8 encoding, causing compilation failures. - [issues/189342]
- Random Number Generator API Proposal: A proposal suggests stabilizing and standardizing accelerator RNG APIs under a unified
torch.acceleratornamespace to provide consistent, backend-agnostic RNG state management across different accelerator devices. - [issues/189520]
- Wheel Build Configuration Regression: The PyTorch 2.13.0+cpu Linux wheel is built with XNNPACK disabled, causing
optimize_for_mobile()to fail with internal assertion errors for traced modules, whereas the previous 2.12.1 CPU wheel included XNNPACK and worked correctly. - [issues/189532]
- Autoscaling and Resource Queueing Issues: A fleet-wide autoscaling problem caused scale-up stalls starting July 9, resulting in synchronized ramp-ups and long queues of 5-7 hours across multiple pools.
- [issues/189558]
- Numerical Precision Bug in RMSNorm Module: PyTorch 2.6.0's
nn.RMSNormproduces significantly incorrect normalization constants onbfloat16tensors with small variances due to precision loss, causing output divergence from expected float32 results; this was fixed in later versions. - [issues/189581]
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: 345
Key Open Pull Requests
1. [DO NOT MERGE][ROCm] Use torch.version.hip to detect ROCm when /opt/rocm is absent: This pull request updates the ROCm detection mechanism in PyTorch's CI and runtime to use torch.version.hip for identifying ROCm environments when the traditional /opt/rocm path is absent, particularly addressing compatibility with TheRock ROCm wheels, improving GPU device count accuracy, and adding various fixes and version-gated test skips to maintain CI stability across ROCm 7.14 preview transitions.
- URL: pull/189229
- Associated Commits: c785b, 0ab9a, 385b3, 62c77, 6df26, c2e3a, 2050e, be3ec, 5527d, 8939d, 46766, dfe06, 3f4f5, a7c9d, ca482, 26ad7, 13c93, bc137, 7e519, c9cb6, 83c9c, b3daa, 41781, 3cec8
2. [precompile] Add dynamic shapes to torch.compiler.precompile: This pull request adds opt-in support for dynamic (unbacked) shapes to torch.compiler.precompile by capturing user-marked symbolic dimensions to enable a single compiled artifact to handle varying runtime sizes, improves the robustness of the _resolve_fake_mode function to correctly process symbolic graphs with None outputs, and restores deferred test coverage for device-generic numerics, subclass/DTensor, and standalone-process execution.
- URL: pull/189165
- Associated Commits: cb985, deb06, f0089, 59c1e, 72acc, f1620, 884d9, 8f292, 48584, ea151, a5a02, ccbad, c992c, 1eb52, 7c49f, d36f0, 22269, 80d55, 3618a, 4021d, 76af6
3. Support multiple pools in cudagraph_trees: This pull request is a work-in-progress implementation that extends the cudagraph_trees component to support multiple CUDA memory pools (MemPools) for capture-time allocations, enabling the mode="reduce-overhead" feature by managing the full lifecycle of one or more user-registered pools across streams and improving memory management flexibility and efficiency in PyTorch's CUDA graph capture system.
- URL: pull/189230
- Associated Commits: d4e82, 52a43, 24944, 21158, 4bbf3, edc89, 1f58b, 78b94, 14bb0, d66ab, a1f92, 6b9cc, f93a6, 23b2b, d0e5a, 4bdaa, f9425, ae4aa, 96410, db3fe
Other Open Pull Requests
- Performance Optimizations and Benchmarking Features: Multiple pull requests introduce new benchmarking and optimization features, including gemv kernels with GPU-specific heuristics for the MPS backend and an opt-in polyhedral fusion optimization for Triton. These changes aim to improve performance and facilitate benchmarking workloads across different backends.
- pull/189628, pull/189460
- XPU Bridge and Backend Fixes: Several pull requests address fixes and improvements to the XPU bridge, including linking xpu_hal for generator bridge registration, updating build configurations for Windows and Linux, improving optional type portability, and fixing utility functions to handle backend changes. These ensure better integration and stability of XPU support in PyTorch.
- pull/189396, pull/189129
- CuptiMonitor Refactoring and Efficiency Improvements: Two pull requests refactor CuptiMonitor by moving flush calls to a native decode thread to reduce GIL contention and converting CuptiMonitor into a process-wide singleton with a global configuration function. These changes improve efficiency and enforce consistent configuration management.
- pull/189083, pull/189185
- Memory Allocation and CUDA Backend Enhancements: Pull requests modify symmetric memory allocation layouts to prevent signal pad pollution and enable expandable segments for CUDA memory allocation on Windows by porting driver APIs and adapting system calls. These updates optimize memory handling and improve cross-platform support.
- pull/189088, pull/189163
- Continuous Delivery Pipeline Consolidation: One pull request consolidates Linux, Windows, and macOS wheel-build pipelines into a unified directory structure, centralizes helper scripts, unifies numpy version pinning, refactors environment setup, and adds documentation. This streamlines the CD process and fixes latent bugs across platforms.
- pull/189275
- torch.compiler.precompile API Development: Multiple pull requests implement the torch.compiler.precompile API with ahead-of-time compilation, add the inductor backend as default, and provide comprehensive user-facing documentation. These changes establish a robust precompile feature with caching, metadata, and support for eager and inductor backends.
- pull/189620, pull/189621, pull/189622
- Process Group and Communication Object Fixes: A pull request fixes process group destruction by deduplicating communication objects to ensure each backend's comm is finalized once, preventing crashes during model parallel teardown scenarios.
- pull/189074
- Bug Fixes in Attribute Handling and GIL Acquisition: Pull requests fix bugs in constant-fold fallback for attribute calls, reduce redundant call_obj_hasattr overrides by improving attribute error handling, and enhance GIL acquisition by fixing exception-escaping issues and adding interpreter finalization checks. These improve stability and correctness in attribute and threading behavior.
- pull/189091, pull/189092, pull/189128
- Triton and CUTLASS Backend Positional Argument Fixes: One pull request extends a fix for positional argument overflow in TritonBenchmarkRequest to the CUTLASS backend, trimming input tensors to match kernel signatures and preventing runtime errors.
- pull/189186
- VariableTracker and Iterator Method Migration: Two pull requests migrate VariableTracker classes and iterator length_hint methods from manual call_method implementations to declarative tp_methods tables, improving method dispatch and maintainability.
- pull/189434, pull/189435
- Test Refactoring for Device Parametrization: A pull request refactors distribution tests from pytest style to PyTorch's TestCase framework with device-parametrized testing, enabling tests to run across multiple devices including custom backends.
- pull/189508
- Inductor Backend and Scheduler Fixes: Pull requests fix non-deterministic gradient computation in Inductor's slice_scatter backward fusion and enable CUDA Graph by default in the AOTInductor backend with runtime and compile-time improvements.
- pull/189129, pull/189191
- MPS Backend Linear Algebra Enhancements: A pull request adds the geqrf function and refactors the linalg_qr kernel for the MPS backend to use a device-agnostic implementation shared with CPU and CUDA.
- pull/189192
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: 166
Key Closed Pull Requests
1. Add CUDA Support for Fused _add_relu With Half and BFloat16: This pull request adds CUDA support for the fused _add_relu operation with half-precision (float16) and bfloat16 data types, enabling a more efficient combined add and ReLU computation on CUDA by avoiding separate kernel launches and ensuring feature parity between CPU and CUDA backends.
- URL: pull/189155
- Associated Commits: 6c049, dd652, 9609e, 9962a, 4b1d8, dbb77, 5bc2f, 7d73a, 383b9, a7d92, ab294, ce383, 7e818, 8947a, caac4, 53f7f
- Associated Commits: 6c049, dd652, 9609e, 9962a, 4b1d8, dbb77, 5bc2f, 7d73a, 383b9, a7d92, ab294, ce383, 7e818, 8947a, caac4, 53f7f
2. [MPS] Migrate lu solve to metal kernels: This pull request migrates the LU solve operation to Metal kernels for improved performance on MPS devices, demonstrating significant speedups across various matrix sizes and right-hand side counts, and includes fixes related to torch.linalg.solve.
- URL: pull/189200
- Associated Commits: 2c243, 72643, 3db11, d42da, 258f5, 9c54d, 4ff95, 10fe0, 37669, d7d1a, 0076f, 8831b, 442b6
- Associated Commits: 2c243, 72643, 3db11, d42da, 258f5, 9c54d, 4ff95, 10fe0, 37669, d7d1a, 0076f, 8831b, 442b6
3. Add multigpu pytest marker to partition distributed tests by GPU count: This pull request introduces a multigpu pytest marker that automatically classifies distributed tests based on their GPU requirements by inspecting their base test classes, enabling the partitioning of the test suite to run multi-GPU tests on a multi-GPU CI runner and single-GPU or single-process tests on a cheaper single-GPU runner, thereby optimizing GPU resource usage and test execution efficiency without requiring manual test annotations.
- URL: pull/189232
Other Closed Pull Requests
- CUPTI Integration and Compatibility: Multiple pull requests enhance CUPTI support by adding a version-gated resolver for the CUPTI header cupti_activity.h, enabling the CUPTI profiler's approximate-clock timestamp callback, and providing a helper to install newer CUPTI headers from the standalone CUDA CUPTI redistributable tarball. These changes ensure compatibility with CUPTI API versions 13.3.0 and above, allow accurate re-timing of activity records, and address outdated headers in CUDA-13 containers.
- ARC/OSDC Runner and Workflow Updates: Several pull requests remove the arc experiment and token from the runner determinator and workflow configurations, migrate the hourly test-stats uploader to an OSDC/ARC runner running in a container without host Docker, and consolidate build jobs by removing legacy EC2 jobs. These changes finalize the OSDC/ARC migration, simplify workflow routing, and update S3 access methods to assume role/arc for gha-artifacts.
- Distributed Process Group and Communication Improvements: Pull requests improve distributed process group handling by preventing crashes when splitting groups with unsupported backends, enabling lazy NCCL group initialization, ensuring custom backends fall back properly, and modifying dist.split_group to preserve user-provided rank order. These updates enhance robustness and provide explicit control over group rank assignments.
- Compile Worker and Sidecar Watchdog Enhancements: Multiple pull requests harden the Inductor SubprocPool by fixing file descriptor inheritance issues, add a compile-worker sidecar watchdog to report stuck jobs, and improve the watchdog with detailed phase reporting for stuck or slow workers. These changes improve shutdown robustness, enable visibility of wedged compile jobs, and enhance diagnostic capabilities without adding IPC overhead.
- pull/189290, [pull/189485](https://github.com/pytorch/pytorch/pull/189485], pull/189486
- Performance Baseline and Code Efficiency Updates: Pull requests regenerate x86_64 and aarch64 performance baselines for operator_benchmark to address baseline drift and apply missing std::move calls throughout the codebase as suggested by clang-tidy. These efforts ensure CI reflects current benchmark results accurately and improve move semantics for better performance.
- Flex_gemm and TensorSSA Matcher Improvements: A pull request extends the feed-main matcher in flex_gemm to recursively handle trailing shape-preserving pointwise nodes and axis-1 groups within TensorSSA fragments. This improves quant cast composition with physical axis-0 feeds and explicitly rejects axis-1 groups beyond fragment limits to avoid misleading errors.
- AArch64 SVE512 CPU Vectorization Support: One pull request introduces support for AArch64 SVE512 by adding compile-time detection, extending runtime CPU dispatch, reusing existing SVE kernels, updating test helpers, and enabling cross-build scripts. This facilitates development and testing on SVE512-capable hardware.
- CUDA Autotuning and cublasLt Candidate Selection: A pull request refactors the CUDA-specific test for cublasLt candidate tunable operations to ensure autotuning runs all candidates and selects the fastest one, avoiding defaulting to the Default TunableOp due to timing noise.
- Operator Schema and Python Signature Accuracy: A pull request fixes incorrect return type annotations for tuple types in operator schemas by dynamically constructing tuple annotations, improving compatibility with strict static type checkers and accuracy of generated Python signatures for multi-tensor returns.
- Functorch Compile_to_Python Testing: One pull request adds comprehensive unit and end-to-end tests for the compile_to_python composition layer in functorch, verifying emitted modules behave equivalently to eager execution and that source-emission helpers and structural guards function correctly.
- Git Timeout Handling in Shim Linters: A pull request addresses an uncaught TimeoutExpired exception in merge_base_with_main() by catching timeout errors during partial clone scenarios and falling back to a slower fetch method to prevent linter crashes in CI.
- Ownership and Access Control Updates: A pull request proposes adding user bohnstingl to merge_rules and CODEOWNER files for control flow and higher-order operators, although it was not merged.
- Reapplying Index_put_ Dispatch Change: A pull request attempts to reapply a previously reverted change dispatching
index_put_tofill_for scalar values, addressing issues caused by Meta internal CI failures.
- Test Shutdown Flakiness Fix: A pull request fixes flakiness in test_shutdown_terminates_sidecar_worker_pool by separating worker startup and shutdown phases, adding a readiness barrier, and extending timeouts to ensure reliable termination without false failures.
- Unmerged Softshrink Type Fix: A pull request addresses a type-related issue in the softshrink function fixing output mismatches for bfloat16 and float scalar inputs, but it was not merged.
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 |
|---|---|---|---|---|
| guilhermeleobas | 74 | 33 | 1 | 6 |
| ezyang | 39 | 22 | 0 | 44 |
| bobrenjc93 | 61 | 13 | 0 | 11 |
| huydhn | 37 | 19 | 0 | 3 |
| Skylion007 | 8 | 3 | 0 | 45 |
| dolpm | 34 | 5 | 0 | 4 |
| Isalia20 | 33 | 9 | 0 | 0 |
| williamwen42 | 26 | 16 | 0 | 0 |
| Copilot | 10 | 5 | 0 | 27 |
| frgossen | 14 | 11 | 11 | 4 |
Access Last Week's Newsletter: