April 2026 monthly summary focusing on Dev work for microsoft/onnxruntime. Key feature work delivered MobileClip performance and optimization work, plus robustness and CI improvements across the repo.

March 2026 monthly summary for microsoft/onnxruntime. Delivered high-impact features and optimization work across CUDA, CPU, and ARM backends, with a focus on expanding hardware support, reducing memory traffic, and accelerating inference for real-time workloads. Business value was realized through broader support for volumetric data, improved operator fusion, and faster activation paths across key models.

February 2026 monthly summary: Delivered cross-OS ARM64 CI coverage, expanded MLAS runtime kernel selection and API cleanup, and strengthened core ops stability through targeted bug fixes and validations. Key accomplishments include enabling ARM64 NCHWc builds on Windows/Linux CI, introducing a backend kernel selector config in MLAS (with explicit parameter passing), adding ConvTranspose bias validation, hardening Einsum for empty inputs and lone operands, and improving CI reliability with conditional FlashAttention test skip on Windows. These efforts reduced CI flakiness, prevented runtime errors, and broadened ARM64 and CUDA-enabled platform support, enabling faster, more reliable releases and better developer experience.

January 2026 performance highlights for intel/onnxruntime focused on delivering measurable business value through kernel-level optimizations and CI reliability improvements. Key outcomes include throughput boosts for common activation paths, a dedicated ARM64 NEON kernel for depthwise convolution, and cleaner, more reliable CI pipelines that reduce flaky runs and accelerate validation cycles.

December 2025: Maintained build stability and improved issue traceability for intel/onnxruntime by focusing on maintenance and patch-tracking efforts. Key action was reverting a CMake configuration change that destabilized builds, coupled with a patch-tracking workflow to ensure underlying issues are resolved in a future release. This work safeguarded CI/release pipelines, reduced risk for downstream users, and improved cross-team visibility into ongoing fixes.

October 2025 monthly summary for intel/onnxruntime focused on CPU-side performance optimizations for convolution kernels. Implemented thread-aware execution paths and memory-efficiency improvements to improve throughput for NCHW Conv workloads across batched and grouped configurations, with attention to ARM64 and other CPU architectures. No major regressions observed; groundwork laid for broader performance gains in upcoming sprints.

Month: 2025-09 Summary of cross-repo development activity focusing on business value, performance, and stability. Highlights include low-precision support, ARM-centric optimizations, distributed AI kernel improvements, and cross-platform build/test robustness. Deliverables span ROCm/onnxruntime, microsoft/onnxruntime, microsoft/onnxruntime-genai, and intel/onnxruntime. Key accomplishments by repo: - ROCm/onnxruntime: • Added 4-bit FP4 data type support in ONNX Runtime (ORT) with FP4 casting integration and 4-bit tensor printing enhancements to improve statistics, debugging, and workflow efficiency. Commits: 16a842a41ac294c0f7c71e7e118a91b1ce5d326c; 4783e0ade83e134101b02e87ffef3f3e21a2b8d6. • Implemented 8-bit GEMM weights on ARM64 for quantized operations, including two kernel flavors (vdotq and vusdotq) and exposure through MatmulNBits; added comprehensive tests to validate performance gains on ARM64. Commit: 31dcc6062e919ce9a6ef53cc64d375f36946126b. • Fixed memory alignment for pre-packed weights buffer in x86 GEMM; restored stability and perf with added tests. Commit: 96f459500ec34d8d2b9fb44385e4efe67ce7fbd9. • Added ARM NCHWc build option to enable ARM kernels for higher-thread-count scenarios; option remains off by default pending stabilization. Commit: 04386c9250edba25f700ab756bef6e1e712fdf92. • MLAS_USE_SVE macro guard added to prevent pipeline crashes in tests/benchmarks, ensuring consistent CI results. Commit: 189e673d13c2268b090aa5da5ce8c28bf4912b34. - microsoft/onnxruntime: • Windows CUDA Profiler Test Stabilization: temporarily disables the profiler test for Windows CUDA builds to improve stability and CI reliability. Commit: bac0bff72b1b4e6fd68ae759a32644defac61944. • CUDA FP4 compatibility/workarounds for Windows builds: fixes to FP4 header usage and suppression of related warnings to ensure clean Windows builds. Commits: 99ee627d3ab1dc3b737ecc6aa0fe56bd616d8eb6; bdffd76c02b84b5aa0e130ef97196b4cdbfb6c6f. • Windows non-CUDA environment robustness: skip BFloat16 tests when CUDA is unavailable to reduce false negatives and improve compatibility. Commit: 6b81b5f602c173044ca2486df5ddc09f5b61110e. - microsoft/onnxruntime-genai: • Distributed TopK kernel with distributed selection for large vocabularies to improve device utilization and performance; includes new metadata/buffers and updated GetTopK usage. Commits: d5dc8cb02fd02b0dce99c6938449566371da0d28; ded6e97789ca718d76ce58bba4a2b483b10045ee. - intel/onnxruntime: • MLAS_USE_SVE macro defined to prevent pipeline crashes in tests/benchmarks; commit: 189e673d13c2268b090aa5da5ce8c28bf4912b34. • CUDA FP4 compatibility and Windows build warnings suppression to ensure Windows builds cleanly with CUDA fp4 usage. Commits: 99ee627d3ab1dc3b737ecc6aa0fe56bd616d8eb6; bdffd76c02b84b5aa0e130ef97196b4cdbfb6c6f. • ARM NCHWc build option addition via PR #25580 enabling ARM kernels for higher thread counts. Commit: 04386c9250edba25f700ab756bef6e1e712fdf92.

In June 2025, ROCm/onnxruntime delivered targeted improvements for quantized inference and test coverage across the XNNPACK Matmul path and CPU builds. Key outcomes include: activation broadcasting fix in XNNPACK Matmul for 1-D activations and correct batch size handling; enabling 8-bit weights in the MatmulNBits kernel via unpacked compute mode to support flexible quantization; and enabling 8-bit Matmul tests on CPU builds by adjusting MLAS header guards. These changes improve performance, flexibility, and reliability, expanding hardware support and accelerating deployment of quantized models. Commits linked to these changes provide traceability: 3426f646a1c1fb57ddf870acea8619579e8c1048, 3b855e1dd6de7d9059864921efca150bf06d5d62, 242cb4398a042221895b982c59f5069a491ffb49.