pypi / onnxruntime
onnxruntime
What changed in onnxruntime 1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
5 breaking3 security2 deprecations40 new APIs and options5 performance12 behavior and defaults18 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initonnxruntime 185 changes
Breaking changes 5
The corresponding experimental lookup entries were removed as EPContext callback registration moved to stable APIs.
Out-of-range indices now produce zero output slices on CPU, CUDA, and WebGPU for integer and floating-point quantized formats, differing deliberately from ONNX Gather error semantics.
CPU FP16 Gemm and MatMul execution is gated on hardware acceleration; CPU-assigned FP16 nodes without a matching kernel now fall back to FP32.
CUDA fpA-intB builds now default to a compact kernel set for FP16 activations, INT4/INT8 weights, scale-only quantization, and block_size=32.
ImageScaler now requires bias to match the channel count.
Security 3
Canonicalized external-data locations to harden model loading.
Added an allowlist of safe LoRA adapter parameter data types.
Limited nested model-graph depth to harden model loading.
Deprecations 2
ACL EP is deprecated and will be removed in a future release; source builds using --use_acl now receive a deprecation warning.
Source builds using this option now receive a deprecation warning.
New APIs and options 40
Added as a weightless hyper-connection operator with CPU, CUDA, and WebGPU implementations.
Compilation can write external initializers and EP context data to buffers, enabling in-memory workflows for models larger than 2 GB.
CoreML is exposed through OrtEpFactory, device enumeration, SessionOptionsAppendExecutionProvider_V2, and EP selection policies; NPU and GPU selection have stated OS minimums.
Extended for Qwen4-Exp.
Added in Node.js/React Native, Objective-C/Swift, and Rust; JavaScript support is read-only and unsupported on Web/WASM.
Added callback support in C#, Python, and Java.
The activation attribute supports SiLU, its Swish alias, and Sigmoid gating across CPU, CUDA, and WebGPU; SiLU remains the default.
WebNN EP added support for GroupNorm.
WebNN EP added support for GroupNormalization.
Added as a weightless hyper-connection operator with CPU, CUDA, and WebGPU implementations.
Added as a weightless hyper-connection operator with CPU, CUDA, and WebGPU implementations.
Added to the standard, non-proxy ONNX Runtime Web run path.
Opt-in CPU MatMulNBits FP32 throughput mode for eligible x86/x64 nodes; it is session-wide, disabled by default, and trades single-request latency for batched throughput.
Disables telemetry.
This opt-in source-build option skips platform accelerator probing and reports only the CPU device.
Set to ON to select Windows TraceLogging telemetry.
Added as an override for the accelerated block-wise INT4/INT8 QMoE CPU expert path.
Plugin execution providers can be statically linked; a static WebGPU plugin build configuration was added.
CoreML can participate in device-based EP selection, including Apple Neural Engine selection through PREFER_NPU.
Added an accuracy_level attribute; INT4 uses FP32 activations by default, while accuracy_level=4 opts into INT8 activation computation.
Added as a weightless hyper-connection operator with CPU, CUDA, and WebGPU implementations.
Set to 1 to fail initialization on workspace declaration overruns or nonzero orphaned reservations; strict verification is unsupported for ORT-format model loads.
WebNN EP added support for SkipGroupNorm.
Added for packed DeepSeek Engram batching with request-local history.
Added the DeepSeek Engram contrib operator.
Added a compact operator and BFloat16 support for CUDA GatedDeltaNet.
Added Go bindings for the ONNX Runtime C API.
Extended memory importing with host-pointer support.
Added access to preallocated outputs through this API.
Added the DeepSeek Engram contrib operator.
Source builds can disable FP4 QMoE kernels with -Donnxruntime_USE_FP4_QMOE=OFF.
Set this source-build option to retain the full CUDA fpA-intB kernel set instead of the compact default set.
Setting this environment variable to 1 opts into an FP8 DeepGEMM MoE decode path for supported fixed-shape QMoE workloads on Hopper GPUs; the path is disabled on Windows.
Extended paged XQA with speculative decoding, additional head-group and head-size support, FP16 caches for head size 256, and native block tables for 128-token pages.
Added INT4 paged KV caches with per-channel scales.
Added SkipLayerNormalization support to the WebNN execution provider.
Added an opt-in split-K GEMV path for small-N FP16 MatMul shapes.
Validated node references and bounded subtree comparison.
Added the operator for continuous batching and compact variable-length causal-convolution state updates.
Exposed safe graph-capture I/O in Python.
Performance 5
Eligible CPU weight prepacking during session initialization is parallelized using the intra-op thread pool; cross-session shared caching retains its serialized path.
Added an opt-in CPU FP32 throughput path for eligible x86/x64 FP32-input, 4-bit, block-size-32, accuracy-level-4 nodes.
Added AVX2 LayerNorm/RMSNorm kernels.
Added fused CPU LinearAttention kernels for AVX-512, Arm64 NEON, and SVE.
Improved PagedAttention and added metadata, GPT-OSS support, and INT8 KV-cache block quantization.
Behavior and default changes 12
Supported native builds now default to 1DS telemetry; Windows TraceLogging remains selectable and telemetry can be disabled.
Named read/write callbacks for application-managed storage are stable; EP support negotiation prevents registered callbacks from silently falling back to filesystem I/O.
Extended to FP8/FP4 data, full-axis blocks, and scale broadcasting; CPU UINT8 dequantization also improved.
Extended for Qwen4-Exp; packed hashing also supports head offsets, EOS resets, and segment boundaries.
Model-package consumers should use this stable API lookup and the stable OrtModelPackageApi table instead of experimental lookup entries.
The model-package C and C++ APIs are stable for loading packages, selecting compatible model variants, and creating sessions.
WebNN supports NHWC Resize by inferring resample2d axes.
Registered BFloat16 LayerNorm/RMSNorm kernels on the CPU execution provider.
FP4 QMoE kernels are enabled by default in CUDA builds; source builds can opt out with the CMake option.
Added INT64 Gather support in the WebGPU execution provider.
WebGPU now supports forwarding onnxruntime_perf_test -i options.
Extended convolution fusion with eight additional activations, fused activation handling in the im2col path, and uniform-based activation parameters for Conv/MatMul.
Fixes and other changes 18
Remains selectable for Windows TraceLogging telemetry.
Misuse now reports errors instead of terminating the process.
Rejected non-finite CPU RoiAlign coordinates.
Enabled WebGPU LayerNorm fusion to fix FP16 inference correctness.
Python asynchronous-run resources are retained and C# arguments are pinned until completion.
Made Rust string-tensor output extraction safe.
Validated Rust tensor element types.
Plugin EP packaging now supports Linux AArch64.
Shape inference was hardened for this applicable contrib operator.
Avoids out-of-bounds reads in ConvTransposeWithDynamicPads.
Shape inference no longer performs an out-of-bounds read when beta is a scalar.
Shape inference was hardened for this applicable contrib operator.
Avoids an out-of-bounds read when a Loop body input is also an initializer.
Reusing a packed sub-byte buffer for a full-byte tensor no longer causes a heap buffer overflow.
Single-tensor data transfers now preserve streams during plugin EP execution.
Avoids an out-of-bounds read when a Scan body input is also an initializer.
Shape inference was hardened for this applicable contrib operator.
Graph validation now rejects inputs to operator schemas that declare zero inputs.