| Index | index by Group | index by Distribution | index by Vendor | index by creation date | index by Name | Mirrors | Help | Search |
| Name: libonnxruntime1 | Distribution: openSUSE Tumbleweed |
| Version: 1.27.0 | Vendor: openSUSE |
| Release: 1.2 | Build date: Wed Jul 1 09:55:07 2026 |
| Group: Development/Libraries/Other | Build host: reproducible |
| Size: 19874847 | Source RPM: onnxruntime-1.27.0-1.2.src.rpm |
| Packager: https://bugs.opensuse.org | |
| Url: https://github.com/microsoft/onnxruntime | |
| Summary: ONNX Runtime shared library | |
This package contains the ONNX Runtime shared library needed at runtime by applications and language bindings.
MIT AND Apache-2.0 AND MPL-2.0 AND BSL-1.0 AND BSD-2-Clause
* Wed Jul 01 2026 Paolo Stivanin <info@paolostivanin.com>
- Update to version 1.27.0:
* [CUDA] Optimize QMoE SoftmaxTopK router for small-batch decode (#29026)
* fix: NodeJS pkging stage needs to use CFS (#29007)
* Disable OrtEp::ort_version_supported sanity check to work around EPs that don't set it correctly. (#28960)
* 1.27.0 - cherry pick 2 (#28900)
* 1.27.0 - cherry pick 1 (#28817)
* Switch NPM publishing to consume from CUDA 13 pipeline (#28773) (#28792)
* Skip SetupDi device discovery if Win32k system calls are disabled (#28535)
* fix(quantization): validate bias scale in QDQ Conv → QLinearConv fusion (#28229)
* Use abseil for readable POSIX stack traces in debug builds (#28405)
* feat(quantization): add opset-21 block_size attribute to QDQ (#28522)
* fix: preserve Q/DQ around Resize when mode is linear or cubic (#28414)
* Fix PRelu returning NaN for infinite inputs on CPU EP (#28750)
* [MLAS][KleidiAI] Apply fp16 split independent fixes (#28720)
* Zero-copy I/O for plugin EPs with HOST_ACCESSIBLE memory (#28037)
* [CoreML EP] Add GatherND builder (#28598)
* Fix ARM CPUIDInfo bounds handling for unknown CPU vendors (#28344)
* Honor disable_synchronize_execution_providers for CUDA graph replay (#28686)
* [MLAS] Fix MatMulNBits CPU LUT GEMM path to apply optional bias (#28742)
* [WebGPU] Support models with opset 24 ops and KV-shared decoder layers (Gemma 4) (#28501)
* Add LpNormalization support for CUDA Execution Provider (#28724)
* Unify path validation in EPs, address security gaps (#28725)
* Add boundary checks for input validation (#28747)
* [CoreML EP] Support bool Cast in ML Program (#28595)
* [CUDA] Add QMoE tests for standard swiglu and update doc (#28741)
* Add Implicit inputs for conditional nodes branches part of node fusio… (#28690)
* Fix CUDA 13 nuget packaging pipeline failures and enable CUDA 13.3 builds (#28736)
* [CUDA] Chunked dequant+GEMM in MatMulNBits to reduce peak GPU memory for large N (#28712)
* Fix IsInfTest (T4 GPU does not support BF16) (#28727)
* [CUDA Plugin EP] Add provider options: user_compute_stream, do_copy_in_default_stream, use_ep_level_unified_stream, external allocator (#28603)
* [NvTensorRTRTX EP] Skip GPU JIT during compile-only sessions (#28503)
* Bump tmp from 0.2.5 to 0.2.7 in /js/web (#28694)
* Bump tmp from 0.2.4 to 0.2.6 in /onnxruntime/test/wasm (#28683)
* Bump qs and body-parser in /onnxruntime/test/wasm (#28644)
* Fix CPU QLinearConv: support per-channel weight zero points with distinct values (#28456)
* Build onnxruntime with training, training_apis, tests and gcc-15 (#27191)
* Update microsoft_gsl from v4.0.0 to v4.2.1 (fixes C4875 deprecation) (#28527)
* Introduce checks to prevent buffer overflow, add tests (#28713)
* CUDA Plugin EP: NHWC Cleanup & Hardening (#28612)
* Add explicit CUDA version suffix to GPU release artifacts (#28691)
* Update CUDA 12.8 to 13.0 in CI workflows (#28458)
* Add `tools/python/compile_contributors.py` `--paths` option (#28710)
* Flash Attention style tiled computation for CPU GQA quantized KV cache (#28695)
* Address external path references in loaded models (#28709)
* fix(partitioning_utils): include Loop/If/Scan implicit inputs in MetaDef (#28608)
* fix(quantization): emit axis on DequantizeLinear for per-channel dynamic quantization (#28228)
* Replace direct inclusion of <cub/cub.cuh> with "core/providers/cuda/cu_inc/cub.cuh" wrapper. (#28705)
* [WebGPU EP] Add environment variable to dump shader code to a file, move shader key validation to nightly build (#28674)
* Address CUDA ReduceL2 negative results and add test coverage (#28681)
* [WebGPU] QKV and MLP layer fusions for Qwen3-style models (#28280)
* Update protobuf version to 6.33.5 (#28684)
* [webgpu] Per-graph buffer manager for WebGPU multi-graph capture (#28260)
* Update GQA op schema (#28689)
* Use fp32 accumulation in SkipLayerNorm/EmbedLayerNorm CUDA kernels (#28682)
* [WebGPU plugin EP] Enforce minimum ORT version specified by `plugin-ep-webgpu/MIN_ONNXRUNTIME_VERSION` (#28687)
* Guard removed APIs in TRT 11 (#28611)
* [CoreML EP] Add Sin and Cos unary ops (#28596)
* Speed up CUDA CI build: split into per-arch OBJECT libraries, add --flash_nvcc_threads, and enable quick build mode (#28645)
* Propagate outer scope type info to subgraphs during verification (#27707)
* Fix CUTLASS FMHA BiasLoader alignment for unaligned kernel path (#28369)
* remove some webgpu ci tasks to speedup ci (#28642)
* Check weight shape dimensions in ConvTranspose shape inference msrc116345 (#28524)
* Add --test_parallel to reduce CTest concurrency for ASan builds (#28675)
* Fix plugin EP test failure when host ORT lacks newer data types (#28659)
* lstm support for webgpu (#27881)
* rev some npm packages (#28547)
* Update GatherBlockQuantized to support 2-bits (#28530)
* Raise protobuf minimum version in Python requirements (#28558)
* [Annotations] Address cuda orginal matching (#28623)
* Optimize MatMulNBits 2-bit + float zero_point CPU dequantization with multi-threaded kernel (#28589)
* Fix Erf NaN preservation on x86_64 FMA3 path (#28615)
* Additional diagnostics for DML failure path (#28495)
* Add example and documentation for kOrtEpDevice_EpMetadataKey_OSDriverVersion (#28282)
* [MLAS] RVV-Optimized LLM Operators for RISC-V (#28518)
* Fix: Accept 'CPU' as a valid provider name in SessionOptionsAppendExecutionProvider (#28625)
* Fix Reshape with allowzero=1 producing wrong shape for zero-size tensor in chained Reshape (#28455)
* [WebGPU] LinearAttention: increase tile_v when subgroups are available (#28519)
* fix(qdq): skip DQ forward propagation when DQ input is constant (#28521)
* Optimize MLAS quantized KV-cache GEMM kernels (follow-up to #28578) (#28606)
* Fix CUDA build with contrib ops disabled (#28554)
* QMoE CUDA: input validation, prepack cleanups, and packaging pipeline fix (#28607)
* [NVEP]: fix test for multi-gpu situation (#27837)
* Parallelize CPU ScatterElements kernel via ThreadPool (#28588)
* Validate seqlens_k against cos_cache bounds in GroupQueryAttention to… (#28277)
* Add component governance manifest for WebGPU EP (#28599)
* [CoreML EP] Support Gather with scalar 'indices' (#28278)
* Validate conv bias shape in WordConvEmbedding to prevent OOB read (#28279)
* Fix oob bias access for MatMulIntegerToFloat and DynamicQuantizeMatMul (#28499)
* [WebGPU plugin EP] Add Win ARM64 Python package (#28526)
* Address MaxUnpool shortcomings msrc116345 (#28550)
* Move `OrtEp` sanity checks to plugin EP creation and remove check for `OrtEp::ort_version_supported` upper bound (#28601)
* Fix Whisper genai_config context_length using wrong config attribute (#28600)
* QMoE CUDA: Rename build options, refactor PrePack, add GPU kernels (#28583)
* CPU GroupQueryAttention: Quantized KV Cache with SIMD-optimized MLAS kernels (#28578)
* Fail loudly when MatMulNBits receives unsupported block_size on CPU EP (#28590)
* feat(orttraining): add CPU fallback for FusedAdam optimizer (#28233)
* fix: propagate output_dtype attribute when inserting Q after DQ (#28144)
* fix: guard ORT_USE_CPUINFO on __linux__ to fix FreeBSD build (#28362)
* Fix EpGraphAssignmentInfo missing issue on nodes assigned to CPU-EP by InsertCastTransformer (#28268)
* [CoreML EP] Add Identity, Ceil, Tile builders + drop trivial-only partitions (#28293)
* Add linux-arm64 binaries to the foundry nuget package (#28579)
* Derive ORT_TEST_VERBOSE from System.Debug via runtime coalesce in plugin test pipelines (#28517)
* Add quantized KV cache support in CPU GroupQueryAttention (#28576)
* add more ORT_ENFORCE when building the tree structure of TreeEnsemble operators (#27677)
* QMoE CUDA EP — FP4/FP8/WFP4AFP8 Quantized Mixture-of-Experts + MoE GEMM Refactor (#28467)
* [MLAS] Correcting Compilation Errors: onnxruntime/onnxruntime/core/mlas/lib/riscv64/sconv_depthwise_kernel_rvv.cpp:138:18: error: unused variable 'pad_bottom' (#28506)
* Fix Cargo registry auth for network-isolated Linux Python wheel builds (#28556)
* Fix protoc error when cross compiling on a linux aarch64 host (#28567)
* Add EP and hardware device type to Windows ML telemetry (#28477)
* [MLAS] test: accept canonical NaN in activation NaN round-trip check (#28538)
* Validate sparse tensor external file paths (#28408)
* Fix XNNPACK Gemm SIGSEGV on missing and scalar 'C' bias (#28546)
* Fix int32 overflow in CUDA Cast and UnaryElementWise kernels for tensors with >2^31 elements (#28386)
* [Security] Validate CPU cache_indirection beam indices (#28486)
* Skip file path validation in CreateEpContextModel when the output is not a file (#28431)
* Add unit test for fix of plugin EP library handle leak in PR 28396 (#28430)
* Support kquant mixed for whisper export (#28515)
* contrib: add mask/input shape consistency checks in MaxpoolWithMask::Compute (#28223)
* Reject recursive local function definitions during model load (#28187)
* [WebGPU plugin EP] Package improvements (#28525)
* webgpu: Generalize FlashAttention prefill shared-memory path (#28520)
* fix: avoid vector copies in CheckIfSubtreesAreEqual (#27854)
* Fix Inverse kernel rank underflow before indexing trailing dims (#28400)
* [MLAS] Fix scratch buffer over-allocation in batched/grouped NCHW Conv (#28482)
* Fix mlasi_sve.h preprocessor guards to allow clang compilation (#28507)
* [QNN EP] ETW log level rule change (#27593)
* Fix input validation and null-pointer dereference in STFTDecomposition graph transformer (#28465)
* Fix WeaklyCanonicalPath ERROR_ACCESS_DENIED in Windows AppContainers (#28509)
* Build issues on AIX for POWER10 and POWER11 (#26704)
* Fill LSTM CUDA operator opset gap: extend coverage from opset 14 to opset 22 (#27737)
* Validate per-column weight_scale/weight_zero_point shape in CPU QAttention; harden integer arithmetic in QAttention and AttentionBase (#28480)
* enable dynamic max_k_step in FA for nvidia (#28511)
* [MLAS] Add an NHWC implementation of convolution to avoid transposes (#26834)
* Improve error reporting for pre-allocated outputs with a wrong shape (#28481)
* fix(security): add SafeInt overflow protection in Expand and constant folding output size limit (#28055)
* fix: support N-D weights with unit leading dims in MatMulNBitsQuantizer (#28178)
* Fix out-of-bounds read in CropBase: validate scale attribute length (#28399)
* CUDA Plugin Packaging Pipeline Fix TSAUpload and Timeout (#28497)
* fix: auto-upgrade model opset to 21 for int16/uint16 QDQ quantization (#28202)
* Bump next from 16.1.5 to 16.1.7 in /js/web/test/e2e/exports/testcases/nextjs-default (#27705)
* Address string attribute issue (#28494)
* Optimize FlashAttention for M4 Max (20x speedup) (#27780)
* Address latent bug in LabelEncoder (#28496)
* Fill opset gap for RandomNormal, RandomNormalLike, RandomUniform, RandomUniformLike CUDA ops with BFloat16 support (#27759)
* Fixes for unicode path handling (#28390)
* Add float zero point support for 2-bit LUT GEMM in MatMulNBits (#28354)
* Bound total output allocation size in Tile kernel (#28070)
* Fix OOB reads in SoftmaxCrossEntropyLoss via label bounds validation (#28004)
* Use weights_only=True for remaining torch.load() calls (#28421)
* Add LabelEncoder_4 fast path for session creation (#27526)
* Add RVV (RISC-V Vector Extension) optimized convolution and pooling kernels for the NCHWc blocked format in MLAS (#28411)
* Address Resize kernel shortcomings (#28402)
* [webgpu] Optimize LinearAttention Op with subgroup (#28412)
* Fill CUDA EP opset gap for Sin and Cos operators (7→22) (#27756)
* Use symbol file to export symbols in onnxruntime library on AIX platform as well (#28016)
* [WebGPU] Correct MatMul bias input indexing in `MatMul::ComputeInternal` (#28475)
* fix: skip DQ->MatMulNBits fusion when weight/scale initializer is shared (#28326)
* [WebGPU] Fix SkipSimplifiedLayerNormalization bias (#28427)
* Add CPU QMoE 2-bit support and LUT GEMM fast path (#28185)
* [ Cuda] ConvTranspose-22 (#27710)
* Fix sigmoid transformation in TreeEnsembleClassifier for all-positive weights with LOGISTIC post_transform (#27536)
* Enable QuickGeluFusion on WebGPU EP and fix fp16 shader (#28410)
* Fix runtime-unresolvable type annotations in Session and InferenceSession (#27802)
* Add DiT attention fusion for F5-TTS and diffusion transformer models (#27999)
* Add CPU Cast op support for Float8E8M0 (#28435)
* [CUDA] PagedAttention: use exact max_query_len on FA path (#28409)
* [GQA] Make present_key/present_value outputs optional and add Gemma4 support (#28242)
* Validate NodeArg creation (#28429)
* Fix CUDA plugin EP test pipeline (#28453)
* Address tokenizer shortcomings (#28428)
* Fix Clang + libc++ compilation and portability issues (#28049)
* Handle empty initializers gracefully in optimizer passes (#27976)
* [WebGPU Plugin EP] Packaging pipelines minor updates (#28460)
* Harden OneHot operator input validation and output size computation (#28014)
* Remove data hash from hashing code, to prevent unbounded growth issue (#28363)
* Don't pin SelectorActionTransformer replacement nodes to CPU (#28288)
* [WebGPU] Fix numerical stability issue in QMoE op (#28434)
* feat(quantization): add ActivationRestrictedAsymmetric option (#28237)
* Fix session use-after-free when UserLoggingFunction is used (#28314)
* Fix CrossEntropyLoss block to support multi-output models (#28232)
* Add PEP 561 `py.typed` marker to the `onnxruntime` package (#28438)
* Add FLOAT8E8M0 data type support in ONNX Runtime (#28381)
* [CoreML EP] Add FusedConv support (#28289)
* Fix loader-refcount leak in ProviderLibrary::Load on GetProvider miss (#28396)
* Fix CUDA plugin EP packaging test pipeline: host network and add cudnn to path (#28373)
* Fix dangling pointer of temporary return value (#28419)
* Remove Universal Package publishing from plugin EP packaging pipelines. (#28418)
* [Plugin EP] Session options getters (#28377)
* [C#] Add EP tests for Cuda Plugin (#28375)
* [CUDA Plugin EP] Add NuGet packaging pipeline (#28378)
* Fix CPU Attention softcap/attn_mask ordering (onnx#7867, #7913) + consolidate CUDA spec coverage tests from #28371 (#28379)
* Suppress test warnings in transformers tests and fix CUDA CI (#28391)
* Fix Subgraph_t issues with TRT RTX ver 1.5.x (#28361)
* [WebGPU plugin EP packaging] Remove explicit ORT package dependency (#28384)
* feat(quantization): add calibration cache to quantize_static (#28221)
* Refactor and modernize StringNormalizer. (#28320)
* [Plugin EP] Add OrtEp::OnSessionInitializationEnd() (#28319)
* [CoreML EP] Support pre-opset-13 Split via 'split' attribute (#28270)
* Add M-tile loop with dispatch capping for Intel Xe2/3-LPG (#28250)
* Prevent double-free in OrtModelEditorApi ownership transfer (#28123)
* webgpu: Support QKV bias in FlashAttention for MultiHeadAttention (#28380)
* Fill CUDA EP opset gaps for Round and Equal operators (#27754)
* Bump brace-expansion in /js/react_native/e2e (#27894)
* Fill Reshape CUDA operator opset gap from 23 to 25 (#27742)
* Fill RNN CUDA operator opset gap (14 → 22) (#27743)
* Fill CUDA opset gap for ReduceMax and ReduceMin (18 → 20) (#27755)
* Fill CUDA Cast operator opset gap: extend registration from opset 23 to 25 (#27744)
* [WebGPU plugin EP] NuGet packaging (#28313)
* Add CUDA plugin EP Python package pipeline (#28299)
* Fix round_prefer_ceil nearest mode for negative halfway values in Resize op (#28345)
* fix(ci): 'rc' qualifier ignored when packaging `onnxruntime-node` (#28350)
* Bump postcss from 8.5.3 to 8.5.13 in /js/web/test/e2e/exports/testcases/vite-default (#28304)
* Fix CApi tests on S390x (#28074)
* Fix CUDA Attention dispatch: skip MEA when head_size != v_head_size in GQA (#28358)
* chore: rename `ort_api_1_to_26` to `ort_api_1_to_27` (#28341)
* fix(ci): test pipeline didn't correctly specify `ReleaseVersionSuffix` (#28346)
* fix: make sympy an optional runtime dependency (#28141)
* Fix ReshapeFusion dropping allowzero on inferred 0-sized intermediate dims (#28349)
* Suppress -Wmaybe-uninitialized for onnxruntime_pybind11_state under pybind11 3.0 (#28251)
* adds foundry local packaging to webgpu plugin ep packaging pipeline (#28300)
* Fix BitShift UB when shift amount >= bit width (#28272)
* Eliminate Legacy MHA Unfused path from ONNX Attention; unify on 3-tier dispatch with causal alignment fix (#27992)
* Add plugin-ep-webgpu/RELEASE.md (#28321)
* refactor(ci): simplify build date/time metadata propagation (#28294)
* Bump plugin-ep-webgpu/VERSION_NUMBER to 0.2.0. (#28322)
* Bump version for 1.27.0 (#28324)
* Support ONNX overloaded functions (IR version 10+) (#28275)
* docs: add opset version notation legend to OperatorKernels.md (#28143)
* Add option to memory map .ORT model loads (#28164)
* add support for DFT with onesided=True and inverse=True (irfft) (#27028)
* Wed Jul 01 2026 Paolo Stivanin <info@paolostivanin.com>
- Move bundled() provides (date, onnx, SafeInt; eigen3 on Leap/SLE)
from the main preamble into the libonnxruntime1 and python3XX-
onnxruntime subpackages, so they ship on the built binaries (no
standalone onnxruntime package is produced) and are found by
bundled-library tracking.
* Tue Jun 30 2026 Paolo Stivanin <info@paolostivanin.com>
- Fix build on openSUSE Leap / SLE (suse_version < 1699): re-bundle eigen.
* The 1.26.0 refactor unbundled eigen3 assuming the system package is 5.0,
which holds on Tumbleweed / Factory but not on Leap/SLE 16.x, which ship
the eigen 3.4.0 release. ONNX Runtime's element_wise_ops.cc (Min/Max ops)
calls ArrayBase::min/max<Eigen::PropagateNaN>(...); those NaN-propagation
overloads were added on eigen's 3.4 maintenance branch *after* the 3.4.0
tag, so the build failed to compile against the system 3.4.0.
* Bundle the eigen 3.4-branch snapshot ORT pins in cmake/deps.txt (the same
commit upstream consumes) for suse_version < 1699 only: drop the
system-eigen3 find via sed and point FETCHCONTENT_SOURCE_DIR_EIGEN3 at the
bundled source, exactly as date/onnx/SafeInt are handled. Factory keeps
using the system eigen 5.0 unchanged. Preserves NaN-propagation semantics.
* Make BuildRequires: eigen3-devel Factory-only (suse_version >= 1699), since
Leap/SLE now build the bundled eigen, and add Provides: bundled(eigen3) =
3.4.90 there accordingly.
* Tue Jun 30 2026 Paolo Stivanin <info@paolostivanin.com>
- Fix build on openSUSE Leap / SLE (Leap 16.0, 16.1 and 15.7), guarded by
suse_version < 1699 so Tumbleweed / Factory keep the upstream behaviour:
* Protobuf: Leap/SLE pull protobuf21 from SUSE:SLFO:Main, which ships no
CMake config -- unlike Factory's protobuf21-devel, which provides
cmake(protobuf). ONNX Runtime declares Protobuf with "FIND_PACKAGE_ARGS
NAMES Protobuf protobuf", and the NAMES keyword forces config-only mode,
so detection failed and the disconnected FetchContent fallback aborted
configure. Rewrite that declare to "FIND_PACKAGE_ARGS MODULE" so CMake's
module-mode FindProtobuf locates the standard-path protobuf Leap does
ship (headers, libprotobuf.so, protoc) and creates the protobuf:: targets.
* Abseil: ONNX Runtime's hardcoded ABSEIL_LIBS list is generated for abseil
20250814 and names targets (e.g. absl::tracing_internal) absent from the
older system abseil on Leap (20240722), which failed the CMake generate
step. Prune ABSEIL_LIBS to the targets that actually exist; a no-op on
Factory, whose abseil (20260107) provides all of them.
* Fri Jun 26 2026 Paolo Stivanin <info@paolostivanin.com>
- Refactor to be less bundled-aware and more patched/upstream (Fedora-style):
* Unbundle flatbuffers, cpuinfo, boost/mp11 and eigen3; only date, onnx
(C++) and SafeInt remain bundled (openSUSE ships no package for them).
* Add system-mp11.patch and system-flatbuffers.patch; rename
eigen-system.patch to system-eigen3.patch (drop EXPERIMENTAL wording,
eigen 5.0 builds fine). cpuinfo needs no patch in 1.26 (system CMake
config provides cpuinfo::cpuinfo).
* BuildRequires: flatbuffers-devel (also provides flatc), cpuinfo-devel,
libboost_headers-devel.
* Regenerate the flatbuffers schemas with the system flatc in %build.
* Drop CMAKE_DISABLE_FIND_PACKAGE_flatbuffers/mp11 and the
FETCHCONTENT_SOURCE_DIR overrides for the now-system deps; remove the
obsolete bundled source zips.
* Fri Jun 26 2026 Paolo Stivanin <info@paolostivanin.com>
- Packaging changes accompanying the 1.26.0 update:
* Build now requires C++20 — BuildRequires bumped to gcc-c++ >= 11 and
cmake >= 3.28 (upstream enforces GCC >= 11.1 / CMake 3.28).
* Disable warnings-as-errors (--compile-no-warning-as-error): GCC 15 emits
false positives (e.g. -Wfree-nonheap-object in skip_layer_norm_fusion.cc)
that upstream's per-target -Werror would otherwise turn into build failures.
* Bundled onnx updated 1.20.1 -> 1.21.0 (upstream cmake/deps.txt); sympy is
now optional upstream but kept as a runtime Requires for shape inference.
* Unbundle dependencies openSUSE ships at a compatible version, per Factory
review feedback:
- protobuf: build against the protobuf21 compat package (21.12, matches
upstream cmake/deps.txt and ships protoc) instead of the bundled copy.
- abseil-cpp + re2: build against the system libraries; abseil's CMake
config is ExactVersion, so abseil-system-version.patch drops ORT's
"FIND_PACKAGE_ARGS 20250814" gate, letting find_package use Factory's
abseil (20260107). re2 follows as a matched pair. NixOS and Fedora also
build onnxruntime against abseil 20260107.
- nlohmann_json, ms-gsl: use the system header-only libraries.
- Switch FETCHCONTENT_TRY_FIND_PACKAGE_MODE from NEVER to OPT_IN.
* Still bundled: flatbuffers (Factory's CMake config exports only the shared
target, not flatbuffers::flatbuffers), mp11 (no Boost::mp11 target from
Factory's Boost), eigen (Factory ships 5.0, ORT needs 3.4), onnx (no C++
package), cpuinfo (devel ships no CMake config), date and SafeInt (not
packaged).
* Drop BSD-3-Clause from the License tag (its only carriers, protobuf and
re2, are now system libraries).
* Fri Jun 26 2026 Paolo Stivanin <info@paolostivanin.com>
- Update to version 1.26.0:
* 1.26.0 - cherry-pick for RC2 (#28347)
* GridSample: harden float->int64 casts against NaN/Inf/out-of-range coords (#28302)
* Relax GQA seqlens_k shape validation for backward compat with older models (#28259)
* Fix DoubleQDQPairsRemover adding spurious dimension to scalar scale/zero-point (#28057)
* remove weights_are_all_positive_ from TreeEnsemble (#27552)
* fix(ci): incorrect relative template includes for setup-feeds (#28312)
* Add QNN Plugin EP repo link to README (#28225)
* Include license file in built distributions (#27783)
* Fix CUDA 13 build error in gqa_unfused_attention.cu (#28309)
* [OVEP] OpenVINO EP 1.26.0 Development Release Updates (#28297)
* Fix NVTX profiling: restore missing node_compute_range_.Begin() call (#28273)
* Fix session logger use-after-free during EP teardown under VERBOSE logging (#28274)
* Implement CUDA EP Plugin profiling API (#28216)
* Language bindings for Hardware Device and EP Compatibility APIs (#28128)
* Propagate python authentication to docker contexts (#28295)
* Use CUDART_VERSION reduction compatibility in GQA attention (#28296)
* WebGPU plugin EP Python packaging (#28226)
* [NvTensorRTRTX EP] Fix build breaks with TRT-RTX 1.5+ and 1.6+ (#28263)
* Fix NPM packaging (#28238)
* Add RISC-V Vector (RVV) support for CPU Execution Provider (#28261)
* ICM fixes (6/n) (#28255)
* [React Native] Add react-native.config.js and Expo plugin MainApplication patch to fix autolinking (#28266)
* [Cuda] Upgrade cutlass to 4.4.2 (#28276)
* [WebGPU] Add GridSample operator (#28264)
* Add update_inplace overload accepting OrtValue for device-to-device copy (#28256)
* Fix cpuinfo init on Linux without CPU sysfs lists (#28230)
* [OVEP] Updating OV version to 2026.1.0 (#28170)
* [WebNN] Rename roundingType to outputShapeRounding for pool2d ops (#28172)
* webgpu: merge batchA into M dimension when batchB==1 (#28197)
* Add aarch64 wheel build to CUDA 13 Python packaging pipelines (#27760)
* WebGPU: Support Split-K with batch size > 1 (#28151)
* Replace unsafe `reinterpret_cast` with C API calls in `include/onnxruntime/ep/adapter/op_kernel_info.h` (#28081)
* Add position_ids bounds validation to WebGPU/JS RotaryEmbedding kernels (#28214)
* Fix universal package version validation comment and add SHA prefix (#28248)
* fix out of boundary vector per class in SVM (#27952)
* Improve SparseTensors public API input validation as well as sparse utilities (#28227)
* [CoreML EP] Add QuickGelu support (#28184)
* chore(ci): temporarily remove react-native from NPM required publish pkg set (#28254)
* [CUDA] PagedAttention: add SM<80 fp16 fallback via memory-efficient attention (#28200)
* security: replace unrestricted setattr with allowlist in Python backend (#28083)
* [VitisAI] pass base timestamp for vitisai profiling (#27808)
* Address CI build failure (#28243)
* ICM fixes (5/n) (#27971)
* chore(ci): `NIGHTLY_BUILD` env-var is now set via pipeline parameter (#28241)
* Fix heap OOB read in RNN operator via sequence_lens=0 (#28052)
* [WebGPU] Fix MHA to ignore past key/value when no present outputs requested (#28027)
* Add regression test for quantize_static with in-memory ModelProto (#23268) (#28201)
* GQA unfused attention with FP32 QK accumulation (fixes #28195) (#28198)
* webgpu: Refactor SubgroupMatrixMatMulNBits to vendor-agnostic config … (#28109)
* Add LabelEncoder CUDA execution provider for numeric types (#28045)
* Add packaging pipeline for CUDA plugin EP (#28152)
* [CI] Remove win gpu doc gen and trt cuda mini CI pipelines (#28218)
* Fix missing include for SetRawDataInTensorProto in NV TensorRT RTX tests (#28065)
* [CoreML EP] Add HardSigmoid support (#28182)
* [Plugin WebGPU EP] Fix API initialization error handling (#28211)
* Fix heap OOB write in EmbedLayerNormalizationShapeInference (#28176)
* Add workflow to replace Windows GPU Doc Gen CI Pipeline (#28192)
* fix: add missing BackendKernelSelectorConfig parameter to QNBitGemmPackQuantBDataSize_Lasx (#28179)
* Add exponential-backoff option for thread pool spin loop (#28096)
* [Core] Add correctness tests for SpaceToDepth and MobileClip Attention fusion (#28168)
* Make sure in memory initialziers are not propagated externally (#28189)
* [MLAS] Enable FP16 for Gelu (#26815)
* Update TRT CI to replace Linux_TRT_Minimal_CUDA_Test_CI pipeline (#28194)
* Address parameter validation across all EPs for Convo kernels (#28142)
* Removes duplicate macos dylib and fixes packaging for the Microsoft.ML.OnnxRuntime.Foundry package (#28191)
* Fix overflow in CopyCpuTensor for sub-byte types (#28171)
* Allow comments in `build.py` argument files (#27624)
* Refine specification of CMake C/C++ standard version requirement (#28126)
* Fix string tensor deserialization in ORT format models (#28133)
* Add MatMul double implementation, inspect optimizers add test coverage (#28145)
* Test Driver update (#28190)
* Apply model size limit before converting size_t input to int32_t (#28112)
* Enable API doc generation builds in PR checks (#28124)
* Implement GetAvaiableResource() callback (#28103)
* [webgpu] support arbitrary input_channel size for im2col (#27038)
* Fix lack of auth on python packaging (#28118)
* Fix RemoveDuplicateCastTransformer incorrectly eliminating lossy Cast chains targeting bool (#28102)
* Add seqlens_k bounds validation in GroupQueryAttention to prevent GEMM OOB (#28031)
* CoreML: Add support for Pad with 'reflect' for ML Program (#28073)
* Fix doxygen issues (#28122)
* ICM fixes (1/n) (#27906)
* [NvTensorRTRTX EP] Implement GetHardwareDeviceIncompatibilityDetails with driver and compute capability checks (#27577)
* Fix packaging pipeline TSA Upload warnings (#28127)
* Add INT8, INT16, and UINT8 type support for CUDA TopK operator (#27862)
* Use cuBLAS status APIs for error strings (#27946)
* WebGPU plugin EP pipeline updates (#28121)
* Address build failure on ARM (#28120)
* Disable QDQ related transformers for WebNN EP (#28020)
* Use weights_only for torch.load checkpoints (#28097)
* Use SafeInt for size arithmetic in CPU tensor operators to prevent overflow (#28060)
* Update worker thread pool to use time based wait. (#27916)
* Fix int32 overflow in CUDA Gather kernel for large tensors (#28108)
* [CORE]: Improve filesystem error messages during Linux device discovery (#27289)
* fix(ml): use SafeInt checked arithmetic in ML operator coefficient size validation (#28001)
* Set SOVERSION 1 for Apple builds to match Linux library versioning (#28101)
* Fix CoreML EP issue with external weight path handling. (#28062)
* [CI] fix: use internal feed for NPM on ADO (#28079)
* Remove old custom EP load infrastructure (#28095)
* Do not create numpy on top of Tensor non-owning buffer (#28088)
* Bump vite from 6.4.1 to 6.4.2 in /js/web/test/e2e/exports/testcases/vite-default (#27991)
* Validate token_id bounds in NGramRepeatBlock to prevent OOB write (#28039)
* Bump picomatch from 2.3.1 to 2.3.2 in /js/web (#27850)
* Bump picomatch from 2.3.1 to 2.3.2 in /js/react_native/e2e (#27849)
* Bump picomatch from 2.3.1 to 2.3.2 in /js/common (#27847)
* Bump picomatch from 2.3.1 to 2.3.2 in /js (#27846)
* removes webgpu from onnxruntime-foundry-nuget package (#27929)
* Added a dimension check in `MatMulComputeHelper` to ensure the inner dimension `K_` (#28053)
* Update global-agent comments to use version-agnostic GitHub URL (#28061)
* [WebGPU EP] Reduce forward declaration boilerplate in kernel registration (#27977)
* Fix heap OOB write in MaxPoolGrad via indices bounds validation (#27903)
* Fix the python pipeline (#28084)
* Add more checks and add functional coverage for bifurcation_detector (#28068)
* Bump follow-redirects from 1.15.6 to 1.16.0 in /js/web (#28056)
* Fix FlashAttentionDecodeSplitVx indirect dispatch input ordering (#27926)
* Bump follow-redirects from 1.15.6 to 1.16.0 in /onnxruntime/test/wasm (#28076)
* Bump electron from 38.1.2 to 39.8.5 in /js/web (#27974)
* Add tensor size validation for MatMulBnb4 to prevent OOB read via K/N attribute mismatch (#27995)
* Bump picomatch from 4.0.2 to 4.0.4 in /js/web/test/e2e/exports/testcases/vite-default (#27844)
* [CI] chore: remove QNN and python DML from release meta-pipeline (#28077)
* [C] fix: ADO CI spuriously failing format check due to picking up NuGet.config (#28064)
* Move WASM builds to large VMs (#28069)
* [CUDA] QuantizeLinear and DequantizeLinear opset 25 (#28046)
* Bump lodash from 4.17.23 to 4.18.1 in /js/react_native/e2e (#27966)
* fix: idempotent feed setup (#28034)
* [CUDA EP Plugin] ResourceAcountant integration (#28028)
* Handle int overflow in rnn (#28003)
* [CI] fix: bad relative ref to setup-feeds template (#28050)
* ICM fixes (2/n) (#27922)
* [Optimizer] Fix ConstantFolding crash with missing optional outputs (#27620)
* Bump yaml from 2.7.0 to 2.8.3 in /js/react_native/e2e (#27863)
* Bump next from 16.1.5 to 16.2.3 in /js/web/test/e2e/exports/testcases/nextjs-default (#28036)
* Bump fast-xml-parser from 4.5.3 to 4.5.6 in /js/react_native/e2e (#28017)
* [QNN-EP] Add additional guards for file mapping (#27871)
* fix a security issue in SVM* (#27950)
* Bump lodash from 4.17.23 to 4.18.1 in /js/web (#27990)
* [WebGPU EP] Fuse QMoE 1-token decode path to reduce GPU dispatches (#27998)
* centralise feed authentication for ADO pipelines (#27997)
* [VitisAI] fix dangling pointer (#27949)
* Fix plugin EP profiling timestamp skew on macOS (#27994)
* Add CUDA Graph support for the CUDA plugin EP (#28002)
* Modify scale & offset of WhereDummyDq (#27109)
* [Plugin EP] Port graph capture/replay APIs (#27958)
* ICM fixes (3/n) (#27925)
* mlas/arm64: add BF16 fast-math conv kernels for NCHW/NCHWc paths (#27878)
* webgpu support for qwen3.5 (#27996)
* Add pre-commit git hook to run lintrunner on staged files (#28013)
* Add bounds validation for LinearClassifier coefficients (#27989)
* Validate g_idx values in MatMulNBits to prevent OOB read (#27582)
* [CPU/CUDA ep] Improve DeformConv op performance (#27824)
* [webgpu] Set `is_channels_last` to true by default in `ComputeMatMul` (#27674)
* fix target_ids out of boundary in TreeEnsemble* (#27951)
* Remove unnecessary model package test (#28015)
* Model Package Support (#27786)
* [web] 2 small fixes to JavaScript code (#27985)
* Add INT8, INT16, and UINT8 type support for CPU TopK operator (#27860)
* webgpu: Fix buffer overflow in BufferManager::Upload causing data corruption (#27948)
* Refactor Intel SubgroupMatrix MatMulNBits (#27911)
* Implement/migrate Arena allocators to cuda plugin ep. (#27931)
* Fix GatherND zero-dimension index validation bug (#28006)
* Add AI agent instructions (#27856)
* Add LinearAttention and CausalConvState ops for Qwen3.5 (#27907)
* Bump lodash from 4.17.23 to 4.18.1 in /onnxruntime/test/wasm (#27969)
* CUDA Plugin Cleanup for Shared Kernel Helpers (#27915)
* Add CUDA plugin EP Sync support for IOBinding (#27919)
* Fix build failure when both CUDA and MiGraphX execution providers are enabled (#27798)
* Add CUDA Plugin EP CI and fix Windows plugin build support (#27959)
* [CORE][NVEP]: add support for Vulkan interop (#27456)
* Add webgpu plugin EP pipeline (#27841)
* Add __array__ and DLPack protocols to OrtValue (#27980)
* Fill CUDA Transpose operator opset gap (23 → 25) (#27740)
* +rotemb, +rmsnorm, reshape->opset-25, transpose->opset-24 (#27752)
* Fix heap OOB write in Attention PrepareMask via negative mask_index values (#27789)
* Fix ADO pools for webgpu & other (#27988)
* Fix heap OOB read in SVMRegressor via attribute size validation (#27901)
* linear attention signature (#27842)
* Bump version to 1.26.0 (#27955)
* [DOCS ]Add annotated partitioning documentation (#27972)
* ICM fixes (4/n) (#27957)
* Fix narrow-to-wide string conversion bugs in dll_load_error.cc (#27777)
* [webgpu] house keeping after ORT switch to c++20 (#27904)
* Fix CUDA ONNX Attention: min_bias_align crash on SM<80 and MEA NaN for fully-masked batches (#27831)
* ONNX 1.21.0 integration (#27601)
* [Core] MobileClip Attention Fusion (#27883)
* Cleanup: Consolidate `OpKernel::UseSharePrePackedBuffers_V2` and `OpKernel::UseSharePrePackedBuffers` (#27924)
* [WebGPU EP] Support Conv3D (#27917)
* Cleanup for op schema API tests for plugin EPs (#27921)
* Add ORT_UNIT_TEST_MAIN_DYNAMIC_PLUGIN_EP_CONFIG_JSON_FILE env var (#27945)
* Fix CPU Attention overflow issue (#27822)
* [DML EP] Disable 4-D MVN tests for DirectML (#27930)
* Fix typo in CApiTest.VersionConsistencyWithApiVersion. (#27902)
* webgpu lpnorm support (#27876)
* Fix CPU LRN int overflow issues (#27886)
* [MLAS] Add depthwise with multiplier conv special kernel for NCHW data layout on Avx512 (#27874)
* Plugin EP event profiling APIs (#27649)
* [VitisAI] external_ep_library typo fix (#27647)
* S390x test fixes (#27404)
* [QNN-EP] Fix use-after-free of logger object (#27804)
* webgpu: Optimize DP4A SmallM MatMulNBits tiling (#27910)
* [Plugin EP] Add plugin EP APIs to retrieve ONNX operator schemas (#27713)
* [CI] fix: missing `branch` specifier in `schedule` directive (#27914)
* CUDA Plugin EP: Core Implementation (#27816)
* [Plugin EP] Add plugin EP Sync API (#27538)
* [CI] chore: bump `actions/cache@v5` (#27866)
* Fix run-level profiling for subgraph operators (#27870)
* webgpu: Increase MatMulNBits K-parallelism with tile_size_k_vec=32 (#27834)
* Add per-session thread pool work callbacks API (#27253)
* Fix integer modulo by zero crash in CPU EP Mod operator (#27833)
* fixes to pass webnn DequantizeLinear compliance tests over webgpu ep (#27778)
* Add cron job to release pipeline (#27864)
* fix potential out of boundary issue when initializer a SVMClassifier (#27699)
* Annotation based partitioning along with resource accounting (#27595)
* Fix overflow in DmlGraphFusionHelper::ProcessInputData (#27815)
* Fix new-delete mismatch in DML EP's QuantizeLinear operator (#27823)
* Add MHA fusion for Nemotron speech conformer encoder (#27764)
* Fix WebGPU buffer segment offset alignment (#27853)
* update jsvascript dependencies (#27838)
* Shell injection constant strings (#27840)
* [CI] feat: use ccache & vcpkg cache for linux workflows (#27623)
* Disable KleidiAI for older versions of MSVC without Aarch64 SME support. (#27825)
* Route fp16 HQNBIT_CompInt8 (4-bit and 8-bit) through fp32 MLAS path in MatMulNBits (#27820)
* Fix --build_wasm_static_lib implicitly enable --build_wasm (#27342)
* [wasm] fix non-SIMD sgemm implementation (#27819)
* mlas/arm64: Add AArch64 assembly path for NCHWc float kernel and wire into build (#27788)
* Only validate REACT_NATIVE if required (#27827)
* webgpu: Fix multi-session GPU profiling event attribution (#27689)
* Fill CUDA EP opset gap for Shape operator (23→25) (#27734)
* Update CUDA TopK kernel registration to opset 24 with BFloat16 support (#27735)
* [CPU] Handle ONNX domain Gelu and HardSigmoid activations in the NCHWc transformer suite (#27821)
* Fix WebGPU device destroyed on session release, breaking session recreation (#27634)
* [WebGPU] Einsum fixes for 5D tensors (#27779)
* Make WebGPU EP compatible with EP API (#26907)
* [MLAS] Add fused Silu and Gelu kernels for AVX512 (#27690)
* Bump flatted from 3.3.3 to 3.4.2 in /js (#27799)
* Bump flatted from 3.3.3 to 3.4.2 in /js/react_native/e2e (#27785)
* Bump rollup from 4.35.0 to 4.59.0 in /js/web/test/e2e/exports/testcases/vite-default (#27463)
* Extend DQ→MatMulNBits fusion to support Gemm + per-tensor/per-channel quantization (#27769)
* fix webnn test case for webgpu ep: 'transpose float32 1D constant tensor default op… (#27773)
* fix webnn/where complicance tests for webgpu (#27776)
* Fix NeonFp16DequantB8Bit reference to match kernel fp16 precision (#27812)
* Fix QNN SDK version propagation in Linux ort-qnn wheel build (#27800)
* [CUDA] Extend Pad support through opset 25 with wrap mode (#27774)
* [CUDA] RoiAlign for opset versions 16 and 22 (#27646)
* Fix non-ASCII Unicode model path crash across session and provider code (#27724)
* [CPU/CUDA EP] Add DeformConv op support (#27393)
* Fix tf.function retracing in TensorFlow benchmark (#27665)
* [CPU] Update the NCHWc transformer to handle more patterns (#27691)
* [Core] Add SpaceToDepth fusion pattern (#27747)
* Add opset 21/23 CUDA kernel registrations for Flatten, Identity, If, Loop, Scan, ConstantOfShape, Size (#27728)
* Bump flatted from 3.2.7 to 3.4.2 in /onnxruntime/test/wasm (#27787)
* Fix out of bounds read issue in cross.entropy.cc (#27568)
* Fill CUDA EP opset gap for GRU operator (14 → 22) (#27738)
* [React Native] Fix float16 tensor crash by mapping to Uint16Array (#27549)
* Dev/erscor/2026/3/19 make react native optional (#27772)
* [WebGPU] DequantizeLinear op fixes (#27706)
* Fill GlobalAveragePool and GlobalMaxPool opset gap in CUDA provider (1→22) (#27733)
* Add SkipLayerNorm fusion with bias Add (#27765)
* [CUDA] Refactor CUDA helpers or ops for plugin EP (#27719)
* Enable KleidiAI for asymmetric 4-bit MatMulNBits on ARM64 (#27751)
* Support softcap and softmax_precision in Attention(CUDA) (#27714)
* Java: avoid provider resource extraction when library already exists in onnxruntime.native.path (#27668)
* Fix integer division by zero crash in CPU EP Div operator (#27693)
* Fill Squeeze and Unsqueeze CUDA opset gaps to opset 25 (#27739)
* Improve checks in concat_slice_elimination against optional attrs and unsqueeze_elimination against invalid model (#27638)
* Add optional router_weights input to QMoE for separate selection/aggregation routing (#27687)
* Fix onnxruntime-qnn python wheel build on WSL (#27730)
* Add ORT_ENFORCE for swiglu_fusion==1 in QMoECPU to prevent out-of-bounds read (#27748)
* Fix large model check in Intel's Neural Compressor (#27746)
* [MLAS][KleidiAI]Catlaw01/sgemm epilogue neon opt (#27609)
* Fix WebGPU ConvTranspose shader bugs for non-vectorizable input channels (#27749)
* Offload H100 issues to A10 pools (#27711)
* [MLAS] Enable BF16 KAI SBGemm on NCHWc ARM builds (#27703)
* [QNN-EP] Disable file mapping for embedded cache (#27627)
* Release Packaging Pipeline Orchestration (#27611)
* fix: invalid ccache args when running with old ccache version (#27698)
* [CANN] Fix test errors during onnxruntime-cann build and compilation (#27091)
* Add fp16 support for 8-bit MatMulNBits on ARM64 and fix pre-existing bugs (#27692)
* CUDA Resize: add optimized 3D nearest resize kernel for 5D up/down sa… (#27578)
* Address Pad Reflect vulnerability (#27652)
* [Build] Fix clang build issues for CPU and CUDA builds (#27669)
* Fix MLAS qgemm dispatch and kernel regressions in quantized conv tests (#27671)
* [CUDA] DecoderMaskedMultiHeadAttention files consolidation (#27688)
* [webgpu] fix condition of DAWN_ENABLE_VULKAN and DAWN_ENABLE_D3D12 (#27694)
* Upgrade minimatch 3.1.2 to 3.1.4 (CVE-2026-27904) (#27667)
* [MLAS] Integrate KleidiAI BF16 SME2 Kernel Through Mlas SBGEMM Path (#26773)
* [Build] Fix debug build (#27659)
* Fix NvTensorRTRTXExecutionProvider capability claims (#27631)
* [web] Fix 2 bugs in emdawnwebgpu (patches dawn) (#27641)
* Plugin EP: Fix bug that incorrectly assigned duplicate MetDef IDs to fused nodes in different GraphViews (#27666)
* [Plugin EP] Check for nullptr before dereferencing (#27644)
* Add RotaryEmbedding fusion for Qwen3 on-the-fly RoPE patterns (#27590)
* [VitisAI] Fix unused warnings (#27630)
* Add C# and Python language bindings for GetCompatibilityInfoFromModel / GetCompatibilityInfoFromModelBytes (#27565)
* Cleanup external data path validation (#27539)
* Revert "[QNN EP] Fix error messages being logged as VERBOSE instead o… (#27650)
* Fix DQ→MatMulNBits fusion for FP16 models on CPU EP (#27640)
* Skip building custom_op_library if CUDA_MINIMAL is enabled (#27520)
* Add PE version info to onnxruntime_providers_vitisai.dll (#27626)
* [Cuda Plugin] Refactor CUDA ops — Move more shared CPU/CUDA helper code from .cc to headers (Part 2) (#27628)
* Fix build errors of GridSample and test failures in test_attention_fusion.py (#27642)
* [TensorRT RTX] Fix compiler warnings on Linux and Windows (#27602)
* [js/rn] Add 16KB page size alignment for Android (#27523)
* Accept legacy WebGPU/WebNN memory info names for backward compatibility (#27637)
* Optimize ONNX Attention KV cache with ConcatNewToPast and add release-build kernel safety (#27613)
* [CUDA] Support volumetric (3-D) grid sampling in the CUDA GridSample operator (#27201)
* Add Qwen3 model type support to Python transformer optimizer (#27556)
* Use SafeInt in implementation of Col2Im (#27625)
* Update to C++20 (#27178)
* MLAS/POWER10: Optimize Sgemm PackA kernel using VSX intrinsics and assembly. (#27575)
* [Cuda Plugin] Refactor CUDA ops — Move shared CPU/CUDA helper code from .cc to headers (#27617)
* Fix WebGPU EP crash on exit (#27569)
* Add PCI bus fallback for Linux GPU device discovery in containerized environments (#27591)
* Fix Potential Integer Truncation Leading to Heap Out-of-Bounds Read/Write (#27544)
* Use `_tpause` instead of `__builtin_ia32_tpause` (#27607)
* Expose Python `OrtDeviceVendorId` enum and use it for vendor-aware `OrtDevice` aliases (#27594)
* [CUDA] Add Validation of batch_indices in RoiAlign (#27603)
* chore: tooling shouldn't ignore forge hidden folders (#27580)
* Add validation of position_ids in RotaryEmbedding operators (#27597)
* Update .NET action version (#27618)
* Extend DQ→MatMulNBits fusion to support 2/8-bit weights and Cast(fp16→fp32) patterns (#27614)
* Add KernelInfo string-array attribute APIs to the ORT C and C++ APIs (#27599)
* Remove some dead YML code (#27530)
* [EP Plugin Adapter] support LoggingManager::HasDefaultLogger() (#27587)
* Bump tar from 7.5.9 to 7.5.11 in /js/node (#27616)
* [CUDA Plugin] Refactoring Einsum (#27606)
* Add validation and unit tests AffineGrid (#27572)
* Address Tile kernel dim overflow and generate tests (#27566)
* Add cherry_pick.py script for ORT release (#27333)
* Add validation for negative position_ids in EmbedLayerNorm (#27573)
* Add WebGPU support for TopK (#27560)
* Support Attention(24)-CUDA and disjoint from contrib op (#27542)
* Add fix for double allocation MiGraphX EP (#27551)
* Session config to create weightless EPContext models during compilation (#27197)
* [QNN EP] Enable offline x64 compilation with memhandle IO type (#27479)
* Adding CIG context creation in OrtFactory (#27267)
* Update CUDA minimum required version to 12.0 (#27570)
* Add missing attribute array validation for TreeEnsembleClassifier (#27571)
* webgpu: support bool for Expand, Flatten, Gather and Unsqueeze (#27561)
* webgpu: Split `MatMulWriteFnSource()` into smaller functions (#27448)
* [ARMNN EP] Remove ArmNN Execution Provider (#27447)
* security fix for transpose optimizer (#27555)
* Add validation of onnx::TensorProto data size before allocation (#27547)
* Fix GPT-2 no-past attention fusion for transformers >= 4.27 (#27449)
* DQ→MatMulNBits fusion transformer for NvTensorRtRtx ep (#27466)
* Remove tools/ci_build/github/azure-pipelines/build-perf-test-binaries-pipeline.yml (#27517)
* Fix RoiAlign heap out-of-bounds read via unchecked batch_indices (#27543)
* [WebNN] Support DepthToSpace op (#27508)
* Add/Update telemetry events (#27356)
* [EP Plugin] use ORT_EP_API_ADAPTER instead of BUILD_WEBGPU_EP_STATIC_LIB (#27540)
* Merge ConstantOfShape and Cast in clean_graph optimization (#27482)
* Add OrtEnv.DisableDllImportResolver to prevent fatal error on resolver conflict (#27535)
* [WebNN] Improve MultiHeadAttention op implementation (#27494)
* [EP API] support OpKernelInfo::GetConfigOptions for adapter (#27541)
* increase kMaxValueLength to 8192 (#27521)
* Account for ORT_NO_EXCEPTIONS builds in Lora test (#27537)
* Remove some dead code. (#27320)
* Fix incorrect warning suppression in projects using precompiled headers on Windows (#27039)
* Add "library_path" metadata entry to OrtEpDevice instances for plugin and provider bridge EPs (#27522)
* [NvTensorRTRTX EP]: Add missing override specifiers to suppress warnings (#27288)
* [CXX Lora] Prevent heap OOB from maliciously crafted Lora Adapters. (#27518)
* Support broadcasting skip shapes in SkipLayerNorm fusion (#27489)
* Fix for Cuda Graph in pre-compiled path (#27477)
* [Build] Fix pybind11 vcpkg configuration (#27514)
* Fix -Warray-bounds build error in MLAS on clang 17+ (#27499)
* Fix double-free in TRT EP custom op domain Release functions (#27471)
* Suppress spurious Array Out of Bounds warnings produced by GCC 14.2 compiler on Linux builds (#27454)
* Split out Linux CUDA Python package builds into separate stages (#27490)
* Add pre-layer normalization support to attention fusion (#27418)
* Fix O(n²) model load time for TreeEnsemble with categorical feature chains (#27391)
* Remove s_kernel_registry_vitisaiep.reset() in deinitialize_vitisai_ep() (#27295)
* Reduce allocation size in test Gather_oveflow_check from 4GiB to >2GiB (#27483)
* [CPU] GridSample operator performance improvement on bilinear interpolation… (#27359)
* [EP API] header-only adapter for EP API (#26919)
* [MLAS/CPU EP]: Introduce a backend kernel selector config in MLAS (#27136)
* Remove cudaStreamSynchronize from CUDA LLM ops for CUDA graph capture compatibility (#27484)
* Improve error message when EPContext node is not assigned to an EP (#27474)
* webgpu: support int64 for Unsqueeze and Expand operators (#27478)
* Move JAR testing pipelines to canonical pipeline template (#27480)
* Fix WebGPU static destruction crash by heap-allocating contexts map (#27470)
* Support boolean attention mask in Attention(23) CUDA - MHA case (#27428)
* Fix BART attention fusion for SDPA pattern from transformers >= 4.49 (#27458)
* Fix Cast node naming collisions and opset 10 Resize in float16 conversion (#27469)
* Build Windows ARM64X binaries as part of packaging pipeline (#27316)
* softplus support for webgpu (#27457)
* Remove unused cusparse library dependency (#27467)
* upgrade dawn to ec7b457e5bb1fcec6f59733c4f3dd84d2f885a38 (#27427)
* [web] remove the unhelpful "Unknown CPU vendor" warning. (#27399)
* [web] fix usage of wasmBinary together with a blob URL for .mjs (#27411)
* [MLAS] Add kleidiai pad ptr invalidation test case (#27465)
* Avoid repetitive creation of fp4/fp8 native-custom-op domains for NvTensorRtRtx EP (#27192)
* Support TensorScatter (24) - CUDA (#27446)
* Add optional dependency for quantization utilities (ml_dtypes) (#27435)
* Fix 3d attention mask broadcasting in MHA (#27464)
* Fix GatherCopyData Integer Truncation Leading to Heap Out-of-Bounds Read/Write (#27444)
* Update python_requires (#27354)
* Fix SkipLayerNorm fusion incorrectly applied when gamma/beta are not 1D (#27459)
* optimized qmoe code path for 1 token (#27383)
* [wasm] add flags for WebGPU build to reduce binary size (#27371)
* WebGPU: Optimize GraphCacheManager by removing pending_buffers (#27453)
* [MLAS] ukernel interface logging changes (#27439)
* Fix VersionNumber usage in React Native Android build with major version validation (#27385)
* Add OrtModel input support for Compile API (#27332)
* fix: OrtValue.from_dlpack rejects zero-size tensors as non-contiguous (#27451)
* webgpu: support head_sink in flash attention (#27410)
* Improve Pre Packing for 2 bit LUT kernels (#27131)
* fix: out of bounds access for resize operation (#27419)
* Add validity check on MaxUnpool indicies. (#27432)
* webgpu: Add configurable multi rotary cache concat offset parameter (#27434)
* Enable Python 3.14 CI and Upgrade Dependencies (#27401)
* Fix validation for external data paths for models loaded from bytes (#27430)
* [CUDA EP] Add resize op version 19 support for CUDA (#27415)
* [Build] Fix nuget pipelines to pass IsReleaseBuild to validate-package (#27387)
* MatmulNBits prepacking scales fix (#27412)
* Normalize Release/Non Release DML packaging pipelines (#27395)
* Move QNN publishing to release pipeline solely (#27420)
* Fix build issues in winml_adapter (#27423)
* Fix DllImportResolver (#27397)
* TensorScatter-24 (CPU) (#27389)
* Fix error where bytes is not assigned for dynamic qgemm pack b size (#27421)
* Fix refcount bug in map input conversion that caused shutdown segfault (#27413)
* Support nonpad kv seqlen within opset 24 Attention (CPU) (#27384)
* chore: increase parallelism for win CI (#27376)
* QMoE CPU Performance Update (Up to 4x on 4-bit) (#27364)
* [js/web] Use embedded WASM module in Blob URL workers when wasmBinary is provided (#27318)
* add option to disable generation contrib ops (#27369)
* [MLAS] Adding DynamicQGemm function pointers and ukernel interface (#27403)
* Prevent cross-EP Cast fusion in RemoveDuplicateCastTransformer (#27363)
* [build] Fix building with torch nightly (#27398)
* [EP API] extract common code for EP API adapter (#26879)
* Log filesystem path instead of string for C++20 compliance (#26745)
* allow a build that disable string type (#27348)
* Fix incorrect pad indices in AveragePool count_include_pad computation (#27375)
* [MLAS] Avoid default parameters in MLAS function signatures (#27128)
* [Node.js] Fix float16 tensor input support (#27327)
* build: Make sympy an optional dependency (#27200)
* Exclude ONNX docstrings in WebAssembly build (#27370)
* Fix QMoE CPU Operator (#27360)
* Fix GatherND division by zero when batch dimensions mismatch (#27090)
* Enable Robust Symlink Support for External Data (HF Cache Support) (#27374)
* Add zero point support to dp4a 2-bit dequantization in the WebGPU MatMulNbits (#27325)
* [CUDA] Support FP8 (E4M3) KV Cache for Group Query Attention (#27321)
* Bump qs from 6.14.1 to 6.14.2 in /onnxruntime/test/wasm (#27345)
* Bump tar from 7.5.7 to 7.5.9 in /js/node (#27372)
* Bump qs from 6.14.1 to 6.14.2 in /js/web (#27344)
* [MLAS] Removed memcpy step by storing result in C if possible (#27367)
* [DNNL] Fix DNNL build (#27334)
* [Build] Fix DML Nuget Pipeline for Release (#27349)
* [Build] Fix python packaging pipeline (#27350)
* Fix misaligned addresses while reading tensor attributes from raw data buffers (#27312)
* Doc fixes (#27228)
* Remove AzurePowerShell pwsh parameter in tools/ci_build/github/azure-pipelines/templates/publish-symbolrequestprod-api.yml (#27339)
* Add boundary checks and add tests for SparseTensorProtoToDenseTensorProto (#27323)
* mlas/arm64: add NEON conv asm kernels and tune NCHWC kernel selection (#27099)
* Update ORT_API_VERSION (#27280)
* BaseTester: support plugin EPs with compiled nodes and registered kernels (#27176)
* Fix NuGet DLL Loading on Linux and macOS (#27266)
* Fix CUDA minimal build requiring CUDNN_HOME (#27308)
* [CUDA] Fix build errors (#27319)
* Fix ConfigEps execution provider ordering in test files (#27014)
* Update Microsoft.ML.OnnxRuntime.Foundry Package (#27294)
* Sync optimizer opset versions with CPU kernel registrations (#27270)
* Check NodeArg has name in ORT formal model load (#27297)
* Fix WebGPU Conv auto_pad=SAME_UPPER padding calculation (#27249)
* win arm64 python packages (#27299)
* Speed up iOS/macOS CocoaPods packaging pipeline (#27279)
* Improve encapsulation of weight_data_type_freq_ (#27292)
* [CUDA] GroupQueryAttention with XQA and Quantized KV Cache Support (#27246)
* [onnxruntime_perf_test] Fix Mac peak working set size value (#27302)
* Support group query attention in Attention(23) CUDA (#27082)
* [Build] Fix java macos (#27271)
* [Build] Adjust nvcc_threads for CI (#27296)
* [MLAS] Fix Lut GEMM Flakiness and Accuracy (#27216)
* [Shape Inference] Fix GQA shape inference for present outputs (#27250)
* [CI] Build NCHWc code on Windows/Linux ARM64 CIs (#27273)
* Improve WebGPU MatMulNBits to support zero pointer for 2bits (#27285)
* Add Contributor Compilation Tool for Release Notes (#27284)
* [BUILD] Fix Build Errors and Warnings in CUDA Providers (#27276)
* Fix out-of-bounds read vulnerability in ArrayFeatureExtractor (#27275)
* Add support for CUDA architecture family codes (#27278)
* Log Framework name to more Windows ML relevant events (#27256)
* [WebNN] Support more features for GQA (#27234)
* [TRT RTX EP] Add support for D3D12 external resource import (#26948)
* Implement compiled model compatibility APIs in example plugin EP and add tests (#27088)
* [Kernel] Fix bug in Einsum implementation when a lone operand had a reduction operation (#27225)
* Fix WebGPU ConvTranspose bias validation in TypeScript and C++ implementations (#27213)
* Record service in telemetry events (#27252)
* [Kernel] Fix einsum operator for empty inputs (#27226)
* Add notes for bumping version as part of a release (#27080)
* Remove test folder from dawn source tree (#27247)
* [webgpu] Enable profiling for graph capture (#27058)
* Optimize for VSCode C++ extension file scan (#27245)
* [webgpu] support auto pad for im2col-matmul (#26771)
* Bump protobuf from 6.33.0 to 6.33.5 in /tools/ci_build/github/linux/python (#27240)
* [WebGPU EP] Disable `Split-K` when `use_deterministic_compute` is true (#27086)
* Update xnnpack.cmake for WASM build (#24860)
* [OVEP] ORT 1.24 Release Patch (#27238)
* Bump next from 15.4.10 to 16.1.5 in /js/web/test/e2e/exports/testcases/nextjs-default (#27194)
* Fix regression in agents for Windows ARM64 (#27235)
* [webgpu] Optimize string stream used in WebGPU EP (#27223)
* [CPU/CUDA] Add bias input validations for ConvTranspose (#27209)
* [CUDA] Run FlashAttention regression test only when FlashAttention is available (#27206)
* Specify attention-23 kernel and relax assertion in prepare qkv (#27217)
* Fix for #25145 (#26994)
* Fix WebGPU MoE swiglu_limit (default to infinity) (#27221)
* Fix Conv LHS packing padding/uninitialized ptrs V2 (#27215)
* [QNN EP] Enablement of 64bit Udma mode (#26677)
* [web] use shorter memory info name for WebGPU buffer and WebNN tensor (#27207)
* [MLAS] Fix rotary interleaved NEON kernel (#26390)
* POWER : Fix build failure due to unsupported cpuinfo on ppc64le (#27120)
* [MLAS] Fix Flaky LuT GEMM Tests by Replacing Gather with Shuffle (#27174)
* Fix API doc comment for OrtApi::RunOptionsEnableProfiling (#27195)
* Bump lodash from 4.17.21 to 4.17.23 in /onnxruntime/test/wasm (#27106)
* Bump lodash from 4.17.21 to 4.17.23 in /js/web (#27105)
* Fix: Replace pkg_resources with importlib.metadata in machine_info.py (#27157)
* Add API GetTensorElementTypeAndShapeDataReference (#27175)
* Bump tar and cmake-js in /js/node (#27193)
* remove coloredlogs (#27135)
* [MLAS] Fix Data Race in MlasLutGemm by Serializing LUT Generation (#27179)
* [WebGPU EP] Reduce duplicated code in `MatMulReadFnSource()` (#27151)
* Add LpNormalization-22 and update the implementation to respect ONNX spec (#27164)
* [QNN-EP] Implement file mapped weights feature (#26952)
* webgpu: optimize Gemm and MatMul using subgroup feature (#26433)
* Bump lodash from 4.17.21 to 4.17.23 in /js/react_native/e2e (#27134)
* Deprecate transformers model examples (#27156)
* Apply absl cuda warning patch to othe OS (#27126)
* Engine compatibility validity API implementation (#26774)
* [CPU] Fix arithmetic overflow and legacy TODO in Det operator (#27070)
* perftest: support plugin eps for compile_ep_context (#27121)
* Remove x86 from nuget (#27124)
* [webgpu] Support int64 for range (#26673)
* [webgpu] fix broadcast for SkipLayerNorm (#27107)
* [webgpu] Use LazyRelease for prepack allocator (#27077)
* [QNN EP] Fix error messages being logged as VERBOSE instead of ERROR (#24931)
* Fix GQA Parity (#27108)
* Add enable_profiling in runoptions (#26846)
* [webgpu] Support Identity (#27067)
* Add absl cuda warnings patch (#27096)
* Linux device discovery for TRT-RTX Ep (#26210)
* Fix OrtApi static_assert violation, add instructions for updating additional API structs. (#27100)
* [test] refactor common test target settings (#27013)
* Move model compatibility checks ahead of session initialization (#27037)
* Add dedicated API to support extracting compatibility string from model metadata (#27015)
* Fix failing mainline build on Arm64 linux (#27101)
* [webgpu] Optimize generic 4D Transpose using OIHW2OHWI Program (#26942)
* Bump version to 1.25.0 (#27048)
* Fix GPU JAR testing on Linux (#27011)
* [test] move IO binding test to onnxruntime_provider_tests (#27055)
* Print tensor for new packed type of 2 bits (#27064)
* Fix Doxygen documentation build error in onnxruntime_c_api.h (#27083)
* Disable Float32_2Bits_Asymmetric_256x256 test (#27046)
* [QNN EP] Add FusedMatMul operator support (#27044)
* [NV TRT RTX EP] Add dimension override for TopK model in TestSessionOutputs unittest (#26975)
* [webgpu] Fix a bug for im2col (#27069)
* [EP ABI] Update Graph_GetGraphView() implementation (#26711)
* Add a new pipeline for CUDA 13 nuget builds (#27023)
* [Reland] Attention(23) CUDA (#27030)
* [EP ABI] Add CreateCustomOpDomains() API for plugin EP to register custom ops (#27050)
* upgrade emsdk to 4.0.23 (#27029)
* Fix some issues when building with the latest CUDA and cuDNN versions (#27049)
* Introducing BF16 Pointwise NCHWc Convolution for Arm64 (#26838)
* Add INT2 and UINT2 support for QDQ, transpose and cast ops (#27022)
* [QNN EP] Add QuickGELU operator support for QNN provider (#27034)
* Tweak external resource importer handle structs (#27040)
* Added support for QMX kernels in MLAS (#26849)
* [CXX] Enable users to specify custom OrtSyncStream via RunOptions (#26988)
* Raise the timeout for the ios simulator job (#27045)
* Implement multithreading in qgemm_kleidi (#26301)
* [QNN-EP] Support alternate Layernorm fusion pattern in QNN preprocess (#26060)
* [MLAS/NEON] Add dedicated kernel for depthwise convolution for ARM64 using NEON intrinsics (#26688)
* [QNN-EP] Add MatMulNBits translation for GPU (#26340)
* Fix warning around ununsed code in QNN Android Emulator builds by clang (#27026)
* [OVEP] OpenVINO EP Features and bug-fixes for ORT-1.24 - Follow up (#27007)
* Add API to get ep graph partitioning info (#26781)
* [QNN EP] Support quantized BatchNorm with per-channel DQ params on QNN HTP (#26959)
* Disable matmul 1d tests on DML (#27021)
* Fix ClipQuantFusion crash when Clip has multiple input edges (#27016)
* [QNN EP] Add support for handling 0-dimension for Concat Op (#27000)
* Revert "Attention(23) CUDA" (#27020)
* [MLAS/CPU EP] Improve performance of Silu activation path within the QuickGelu CPU kernel (#26753)
* [NV TensorRT RTX EP] Fix external tensorrt_plugins load path (#26814)
* Implement new experimental lookup-based matrix multiplication method(TMAC) (#26695)
* [QNN EP] Add BFloat16 dtype support in QNN EP (#26987)
* Add type definitions, registration, utilities for INT2/UINT2 support (#26824)
* Add check that tensor sizes match in DataTransferManager::CopyTensors (#27008)
* Attention(23) CUDA (#26466)
* Thu Jun 25 2026 Paolo Stivanin <info@paolostivanin.com>
- Packaging fixes for Factory submission:
* Fix the python314 build flavor: the spec only handled 3.11/3.12/3.13
while _multibuild requests 3.11/3.13/3.14, so the python314 build aborted
with "Unknown flavor". The spec now targets 3.11/3.13/3.14.
* Add BSD-3-Clause to the License tag to cover the bundled, statically
linked protobuf and re2.
* Declare the bundled() provides on the python subpackage as well, via a
shared macro so the shared library and python packages stay in sync.
* Document in the spec why each dependency is bundled (the set matches
upstream cmake/deps.txt exactly and system versions are incompatible).
* Drop unused, unreferenced source archives (cxxopts, dlpack, FP16).
* Mon Mar 23 2026 Jürgen Löhel <juergen@loehel.de>
- Update to version 1.24.4:
* Added PCI bus fallback for Linux device discovery in container environments where standard GPU drivers are not loaded.
* Fixed null pointer dereference in Plugin Execution Provider output handling.
* Fixed duplicate MetaDef ID assignment causing session creation failures in graphs containing conditional branches.
* Reverted logging changes that could trigger segmentation faults during backend destruction.
* Replaced compiler-specific _tpause builtin with portable intrinsic to improve GCC/LLVM build compatibility.
* Updated minimum supported Python version to 3.11.
* Mon Mar 23 2026 Jürgen Löhel <juergen@loehel.de>
- Update to version 1.24.3:
* Fixed multiple security issues causing heap out-of-bounds access in core operators (Gather, RoiAlign, Resize) and malicious LoRA adapter handling.
* Fixed GatherND division-by-zero when batch dimensions mismatch.
* Improved validation of external model data paths when loading models from memory.
* Corrected SkipLayerNorm fusion when gamma/beta tensors are not 1-dimensional.
* Fixed QMoE CPU operator correctness issues.
* Fixed MatmulNBits prepacking scale handling.
* Fixed Python reference counting issue causing interpreter shutdown segmentation faults.
* Improved QMoE CPU performance (up to 4× for 4-bit workloads).
* Fixed quadratic model load time for TreeEnsemble models with categorical feature chains.
* MLAS: added DynamicQGemm interface and fixed dynamic qgemm packing size handling.
* Suppressed incorrect array out-of-bounds compiler warnings on recent GCC versions.
* Fixed -Warray-bounds build failure with newer clang versions.
* Increased internal configuration maximum value length limit.
* Mon Mar 02 2026 Jürgen Löhel <juergen@loehel.de>
- Update to version 1.24.2:
* Enable Robust Symlink Support for External Data for Huggingface Hub Cache.
* Added boundary checks for SparseTensorProtoToDenseTensorProto to improve robustness.
* Fixed an out-of-bounds read vulnerability in ArrayFeatureExtractor.
* Accuracy-Fix in Lut GEMM.
* Mon Feb 16 2026 Jürgen Löhel <juergen@loehel.de>
- initial commit with python bindings
/usr/lib64/libonnxruntime.so.1 /usr/lib64/libonnxruntime.so.1.27.0 /usr/share/doc/packages/libonnxruntime1 /usr/share/doc/packages/libonnxruntime1/ThirdPartyNotices.txt /usr/share/licenses/libonnxruntime1 /usr/share/licenses/libonnxruntime1/LICENSE
Generated by rpm2html 1.8.1
Fabrice Bellet, Mon Aug 3 00:14:01 2026