pypi / executorch
executorch
What changed in executorch 1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
2 breaking57 new APIs and options2 performance4 behavior and defaults3 tooling18 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initexecutorch 186 changes
Breaking changes 2
Softmax, log-softmax, mean, and sum now accumulate in FP32 for improved numerical accuracy.
Removed the legacy toggles from the runtime and export path.
New APIs and options 57
Added support for the Ethos-U65 target.
Added a public API for retrieving the thread count.
The export pipeline supports multi-method models and per-backend load options.
Added Neutron support for
rsqrt, Hardswish,softmax, reflected padding, scalar elementary arguments, and AdaptiveAvgPool1d.The AOTI backend gained opt-in ROCm support, packed W4 execution, and Voxtral Realtime coverage.
Added Qualcomm support for the
as_stridedoperator.LLM Export adds BF16 activation support.
Added BF16-by-FP32 GEMV decode support.
XNNPACK gained BF16 fully connected delegation.
Python wheels now include a C++ SDK with linkable delegate libraries and packaged TorchAO kernels.
Memory planning can account for target memory maps and shared allocations with offsets.
Added opt-in explicit-layout lowering for Cortex-M, with convolution and pooling kernels.
Arm support now includes the Ethos-U65 target.
Arm export now supports Ethos-U85 KV caches.
MLX lowering now supports flip.
Qualcomm enables Gemma 4 text models.
Shared pre-decomposition and graph-transform hooks include common view/permute propagation and reinplace passes.
Qualcomm adds support for Hadamard-transform flows.
Added INT4 short-query kernels for CUDA.
Added INT5 short-query kernels for CUDA.
Added INT6 short-query kernels for CUDA.
Qualcomm support now includes LLM quantization-aware distillation.
Metal adds addmm support.
Metal adds split_copy support.
The export pipeline now supports models with multiple methods.
MLX lowering now supports native_group_norm.
MLX adds flat, ring, and cell off-graph KV-cache layouts with shared pools.
Added Qualcomm support for the
pdistoperator.Loading now supports options configured per backend.
Vulkan adds quantized convolution paths using unsigned dot products.
Added a quantized MoE feed-forward operator with INT4 packing and fused projections.
Added a quantized stacked-halves RoPE operator.
Added Qualcomm support for the
reflection_pad3doperator.Added Qualcomm support for the
scatter_addoperator.Added Qualcomm support for the
scatter_reduceoperator.CUDA inference can target shared memory across architectures.
Added Qualcomm support for the
sortoperator.MLX lowering now supports trunc.
Vulkan adds two-dimensional dispatch helpers and device-safe dispatch geometry.
MLX lowering now supports upsample_nearest2d.
Vulkan can wrap external images and buffers using caller-provided dimensions.
Added 2-bit weight quantization support.
Added an MLX handler for bitwise_or.
Added BFloat16 and FP16 activation support.
Ethos-U and VGF dependencies can be installed with the optional package extras.
Added a grammar field to GenerationConfig.
Added a grammar_type field to GenerationConfig.
Added an MLX handler for hardtanh.
Added an MLX handler for leaky_relu.
Added partitioning, tensor dumping, and runtime heap profiling.
Added support for quantizing LLMs with multiple batches.
Added addmm, select_scatter, variance, diagonal, fill, cdist, LayerNorm variants, and hyperbolic functions.
Added BFloat16-output quantized embeddings on CPU.
Added quantized same-padding convolution support in XNNPACK.
Expanded lowering for activation, binary, conversion, data layout, FFT, reduction, shape, and unary operators.
Added or improved profiling, shader segments, BFloat16, MXFP, grid sampling, and persistent IO mappings.
Added desktop support for Linux and Windows GPUs and opt-in wheel builds.
Performance 2
BF16 SDPA was optimized.
Decoupled calibration, quantization, compilation, and evaluation stages for faster workflows.
Behavior and default changes 4
Arm support was expanded for symbolic shapes, dynamic convolution, and pooling.
Migrated the Android API from Java to Kotlin.
Expanded operator coverage and removed several broadcasting restrictions.
Migrated the Android API from Java to Kotlin.
Tooling 3
Added scratch-buffer planning for AoT.
Added a tool reporting ANE, GPU, and CPU dispatch.
Added backend profiling.
Fixes and other changes 18
The supported CMake floor was lowered so builds with CMake 3.26 through 3.28 work again.
Fixed NaNs in CUDA split-K attention for Voxtral.
Failures for invalid device-planned copies are clearer.
Failures caused by missing delegate data are clearer.
PTE and PTD files are checked for schema versions, with clearer failures for invalid device-planned copies and missing delegate data.
Improved direct serialization for embedded flows.
Added explicit dim-order legalization for channels-last portable fallbacks.
Expanded support to additional CPU variants, including more Corstone-300 variants and M0+.
Migrated the Android APIs from Java to Kotlin.
Fixed partition, format, and conversion issues.
Fixed correctness.
Fixed a null dereference when op-package options are absent.
Fixed same padding for even kernels in XNNPACK.
Improved preservation of symbolic input metadata during exported-program pass replay.
Hardened validation and segment loading against null metadata.
Migrated the Android API from Java to Kotlin.
Fixed incorrect baked-in paths within wheels.
Fixed XNNPack PreLU in quantized graphs.