77 Pull requests merged by 34 people

• Properly set mutable buffer lifespans

  #12182 merged Jul 8, 2025

• Quant doc updates

  #12260 merged Jul 8, 2025

• [release/0.7 only] Fix check_c10_sync being out of sync

  #12252 merged Jul 8, 2025

• executorch quantier numeric debugging update for recent torchao changes

  #12185 merged Jul 8, 2025

• (release branch only) Revert "Respect selective build for dtype_specialized_elementwise_fn_impl in elementwise_util (#11975)"

  #12184 merged Jul 8, 2025

• Add last_token_pos in llama_transformer (#11793)

  #12239 merged Jul 8, 2025

• [Yolo12] Revert of the revert of the Yolo12 Sample

  #12163 merged Jul 8, 2025

• Quant doc updates

  #12240 merged Jul 8, 2025

• Arm backend: Add decomposition pass and test for asin

  #12241 merged Jul 7, 2025

• Resubmission of Arm backend: Split executor runner into init + run

  #12197 merged Jul 7, 2025

• Remove the legacy export

  #12218 merged Jul 7, 2025

• fix type promotion for div in RemoveMixedTypeOperators

  #12157 merged Jul 7, 2025

• annotate the rms_norm

  #12238 merged Jul 7, 2025

• Add coverage for minimum in TestPasses.test_remove_mixed_type_operators

  #12156 merged Jul 7, 2025

• Don't consider system buck2

  #12245 merged Jul 7, 2025

• Manual cherry-pick: Parallelize optimized op_log_softmax

  #12246 merged Jul 7, 2025

• Clean up TestPasses.test_remove_mixed_type_operators

  #12155 merged Jul 7, 2025

• Added mse range setting

  #11857 merged Jul 7, 2025

• Added Pybindings for Method.h/cpp

  #12158 merged Jul 4, 2025

• Arm backend: Remove upsample ops from ops_to_not_decompose on U55

  #12236 merged Jul 4, 2025

• Arm backend: Add pass and test for adaptive_avg_pool2d

  #12190 merged Jul 4, 2025

• Arm backend: Introduce TOSA backend dialect

  #12195 merged Jul 4, 2025

• Qualcomm AI Engine Direct - GA Whisper

  #12102 merged Jul 4, 2025

• Fixed bug for 16a4w ptq

  #12167 merged Jul 3, 2025

• Implemented Runtime Intermediate Output Extraction Based on Corresponding AOT Operators

  #12212 merged Jul 3, 2025

• Parallelize optimized op_log_softmax

  #12099 merged Jul 3, 2025

• update building from source doc to make the backend support guidance clearer

  #12169 merged Jul 3, 2025

• Update SwiftPM pins in Xcode projects

  #12115 merged Jul 3, 2025

• Fix Qwen export command backslashes

  #12181 merged Jul 3, 2025

• Set Weight Loading to be strict

  #12151 merged Jul 3, 2025

• Fix typing_stubs deps.

  #12179 merged Jul 3, 2025

• Arm backend: Add unit tests for per-channel quantization

  #12192 merged Jul 3, 2025

• Plumbing for Arm Example Runnner to use Zephyr Toolchain

  #12078 merged Jul 3, 2025

• Arm backend: Add sign decomposition pass and test

  #12159 merged Jul 3, 2025

• Arm Zephyr cmake Preset

  #11923 merged Jul 3, 2025

• Adjust tolerance for fp16 exp & gelu ops test to handle reasonable calculation discrepancies

  #12150 merged Jul 3, 2025

• [Executorch][llm] Fix ring kv cache when used with quantized kv cache and sdpa

  #12143 merged Jul 3, 2025

• [Release 0.7] Update all docs to reflect the new branch 'release/0.7'

  #12187 merged Jul 3, 2025

• [Executorch][llm] Make mask tensor float only for sdpa

  #12142 merged Jul 3, 2025

• executorch quantier numeric debugging update for recent torchao changes

  #12173 merged Jul 2, 2025

• Add code to Bundleio to generate error stats

  #12051 merged Jul 2, 2025

• Supporting Zephyr with Executorch as a Module

  #12174 merged Jul 2, 2025

• Use shared log_softmax kernels from PyTorch

  #12172 merged Jul 2, 2025

• Reapply changes reverted by #11506

  #12176 merged Jul 2, 2025

• Install headers from runtime/executor and extension/module in CMake build

  #12175 merged Jul 2, 2025

• Deprecate tag qdq pass in xnnbackend

  #12170 merged Jul 2, 2025

• Reapply changes reverted by #11506

  #12121 merged Jul 2, 2025

• [Executorch][llm] Make runner return error if execution was not successful

  #12141 merged Jul 2, 2025

• Remove _skip_type_promotion config

  #12149 merged Jul 2, 2025

• fix is_inplace_node check

  #12071 merged Jul 2, 2025

• Revert "Arm backend: Split executor runner into init + run"

  #12171 merged Jul 2, 2025

• use graph.output_node

  #12139 merged Jul 2, 2025

• Use shared log_softmax kernels from PyTorch

  #12098 merged Jul 2, 2025

• update building from source doc to make the backend support guidance clearer

  #12140 merged Jul 2, 2025

• Revert "Qualcomm AI Engine Direct - CI for Non-LLM GA model"

  #12166 merged Jul 2, 2025

• Introducing NXP Neutron runtime

  #10563 merged Jul 2, 2025

• [BE] Add selected custom ops to CI

  #11744 merged Jul 2, 2025

• Arm backend: Fix bug in decompose_linear_pass

  #12160 merged Jul 2, 2025

• Arm backend: Split executor runner into init + run

  #12162 merged Jul 2, 2025

• Arm backend: Add mean.dim to CheckProperQuantization

  #12127 merged Jul 2, 2025

• add custom annoatation for new model export

  #12126 merged Jul 2, 2025

• Add Pybindings for Program.h/cpp

  #12016 merged Jul 2, 2025

• Fixes in to_executorch for while

  #12062 merged Jul 1, 2025

• Improve memory planning for submodule hierarchies.

  #11860 merged Jul 1, 2025

• [Release 0.7] Update CI pins

  #12137 merged Jul 1, 2025

• remove unnecessary print

  #11925 merged Jul 1, 2025

• Enable training for pybindings

  #12144 merged Jul 1, 2025

• Specialize BroadcastIndexesRange for the case where there is only 1 contiguous input

  #12023 merged Jul 1, 2025

• [Executorch][llm] Fix ring kv cache when used with quantized kv cache and sdpa

  #12132 merged Jul 1, 2025

• [Executorch][llm] Make mask tensor float only for sdpa

  #12131 merged Jul 1, 2025

• [Executorch][llm] Make runner return error if execution was not successful

  #12129 merged Jul 1, 2025

• Updating documentation for cmake dtype selective build

  #12123 merged Jul 1, 2025

• Bump version in apple.yml for nightly builds

  #12118 merged Jul 1, 2025

• Revert "[Example] Yolo12 Detection sample with OpenVINO/XNNPACK backend"

  #12136 merged Jul 1, 2025

• Qualcomm AI Engine Direct - CI for Non-LLM GA model

  #11357 merged Jul 1, 2025

• Arm backend: Measure and show time per model during testing

  #12128 merged Jul 1, 2025

• Arm backend: Add pytests time stats for unit test

  #12110 merged Jul 1, 2025

44 Pull requests opened by 23 people

• Delete opt_mul_scalar_out

  #12145 opened Jul 1, 2025

• [not for merge] Add armv7 support in buck shim

  #12146 opened Jul 1, 2025

• refactor XNNPACK's ukernel config srcs

  #12152 opened Jul 1, 2025

• refactor XNNPACK's ukernel config srcs

  #12153 opened Jul 1, 2025

• Move xnnpack to 5220835694

  #12154 opened Jul 1, 2025

• Qualcomm AI Engine Direct - LE support

  #12164 opened Jul 2, 2025

• Qualcomm AI Engine Direct - gpu support part1

  #12165 opened Jul 2, 2025

• [cadence] add logging init to cadence tests

  #12177 opened Jul 2, 2025

• Fix dtype selective build check in elementwise_util

  #12183 opened Jul 2, 2025

• 0.7 release ready

  #12189 opened Jul 3, 2025

• Qualcomm AI Engine Direct - CI for Non-LLM GA model

  #12191 opened Jul 3, 2025

• Added support for --lib-name in manual kernel registration

  #12193 opened Jul 3, 2025

• [ET-VK][ez] Fix partitioner logic

  #12196 opened Jul 3, 2025

• Binary Comparison Ops

  #12198 opened Jul 3, 2025

• [ET-VK][Ops] aligning Q/DQ/CQP op inputs with ATen impl

  #12199 opened Jul 3, 2025

• [ET-VK][ez][Ops] registering Q/DQ/CQP ops and specifying optimal storage

  #12200 opened Jul 3, 2025

• [ET-VK][ez] enabling fp64->fp32 converison for vulkan compatibility

  #12201 opened Jul 3, 2025

• [ET-VK] lowering ExecuTorch tensor dtype for Vulkan tensor dtype to enable 64bit

  #12202 opened Jul 3, 2025

• [ET] correcting cpu ref quantize_per_channel logic to align with ATen

  #12203 opened Jul 3, 2025

• [ET-VK][Ops] quantize_per_channel reference impl and testing

  #12204 opened Jul 3, 2025

• [ET-VK][Ops] quantize_per_channel shaders and impl

  #12205 opened Jul 3, 2025

• [ET-VK][Ops] dequantize_per_channel reference impl and testing

  #12206 opened Jul 3, 2025

• [ET-VK][Ops] dequantize_per_channel shaders and impl

  #12207 opened Jul 3, 2025

• [ET-VK][Ops] quantize_per_tensor.tensor variant

  #12208 opened Jul 3, 2025

• [ET-VK][Ops] dequantize_per_tensor.tensor variant

  #12209 opened Jul 3, 2025

• [ET-VK][testing] Q/DQ/CQP op comprehensive delegate dynamic quantization testing

  #12210 opened Jul 3, 2025

• Add type error suppressions for upcoming upgrade

  #12213 opened Jul 3, 2025

• Remove unused compute_fun argument to validate_elementwise_fn_inputs

  #12217 opened Jul 3, 2025

• Milestone2.1: Partition to_dim_order_copy op in XNN delegate

  #12220 opened Jul 3, 2025

• Qualcomm AI Engine Direct - GA model enablement (T5)

  #12234 opened Jul 4, 2025

• Arm backend: Add initial module tests for Stable Diffusion 3.5 Medium

  #12242 opened Jul 7, 2025

• Arm backend: Match fp32->int32 cast between pytorch and TOSA's CAST

  #12243 opened Jul 7, 2025

• Buckify runtime

  #12244 opened Jul 7, 2025

• [RFC] Add TrainingModule and SGD JNI + PTE-only Training Workflow

  #12247 opened Jul 7, 2025

• Add XOR example to Android JNI setup

  #12250 opened Jul 7, 2025

• Updated the comparison logic to handle sequences separately

  #12251 opened Jul 7, 2025

• Bump PyTorch pin to 20250706

  #12253 opened Jul 7, 2025

• Remove mkldnn flag

  #12254 opened Jul 7, 2025

• [BE] Clean pte_data_map

  #12255 opened Jul 8, 2025

• [Android] Format all Java files

  #12256 opened Jul 8, 2025

• Don't force extension_module to build as a shared library by default

  #12257 opened Jul 8, 2025

• Add support for absolute mem_id/offset placement constraints.

  #12266 opened Jul 8, 2025

• Link backend to prtable libs

  #12268 opened Jul 8, 2025

• Delete unused import statement for Executorch TOSA dialect

  #12269 opened Jul 8, 2025

16 Issues closed by 7 people

• Update quantization overview doc page

  #10603 closed Jul 8, 2025

• Support for WebGL/WebGPU

  #12249 closed Jul 7, 2025

• CoreML fails to lower pooling with single-element padding list

  #11696 closed Jul 7, 2025

• CoreML fails to lower ConvTranspose with output padding

  #11705 closed Jul 7, 2025

• CoreML conv fails to lower with circular padding mode

  #11703 closed Jul 7, 2025

• CoreML missing partition constraint for tensors with rank greater than 5

  #11694 closed Jul 7, 2025

• CoreML missing partitioner constraint for PixelUnshuffle on older targets

  #11711 closed Jul 7, 2025

• CoreML doesn't handle negative dim values correctly in cumsum

  #11716 closed Jul 7, 2025

• fail to build QNN backend on macOS

  #8082 closed Jul 3, 2025

• Deprecate tag implicit qdq pass

  #11588 closed Jul 2, 2025

• freqs_cos data type of Llama3.2-1B mismatch with activation

  #8614 closed Jul 1, 2025

• QNN model running on SM8750

  #8475 closed Jul 1, 2025

• SM8750 Htp not added in Android apk building

  #8454 closed Jul 1, 2025

• XNNPACK Pin Update

  #11933 closed Jul 1, 2025

• New SDOT Kernels in KleidiAI

  #11931 closed Jul 1, 2025

• Update Android Executorch example (Image Segmentation) documentation accounting for the newly added feature

  #11915 closed Jul 1, 2025

32 Issues opened by 13 people

• assert isinstance(quantization_spec, QuantizationSpec)

  #12267 opened Jul 8, 2025

• Add quantization documentation to the Qualcomm docs

  #12259 opened Jul 8, 2025

• Add quantization documentation to vulkan docs

  #12258 opened Jul 8, 2025

• [RFC] Unified Recipe Management System for ET backends and users

  #12248 opened Jul 7, 2025

• [Arm backend] can't convert yolo models with aot_arm_compiler

  #12237 opened Jul 4, 2025

• Vulkan divide with truncate rounding mode doesn't match eager/portable

  #12235 opened Jul 4, 2025

• Vulkan squeeze errors out for negative (relative) dims

  #12233 opened Jul 4, 2025

• Vulkan mean errors out during lowering

  #12232 opened Jul 4, 2025

• Vulkan embeddings give incorrect outputs

  #12231 opened Jul 4, 2025

• Vulkan partially delegated, decomposed cross product gives incorrect outputs

  #12230 opened Jul 4, 2025

• Vulkan floor_divide gives incorrect outputs with integer divisor

  #12229 opened Jul 4, 2025

• Vulkan partially delegated replication_pad2d fails to run

  #12228 opened Jul 4, 2025

• Vulkan amax/amin fail to lower with index out of range

  #12227 opened Jul 4, 2025

• Vulkan replication_pad3d pte fails to load with missing shader error

  #12226 opened Jul 4, 2025

• Vulkan linalg norm fails with arg out of range

  #12225 opened Jul 4, 2025

• Vulkan model with (adaptive) avgpool1d (or maxpool1d) fail to load

  #12224 opened Jul 4, 2025

• Vulkan transposed conv 1d with padding outputs don't match eager

  #12223 opened Jul 4, 2025

• Vulkan index select outputs don't match eager/portable

  #12222 opened Jul 4, 2025

• How to build executorch with ANDROID_ABI=armeabi-v7a

  #12221 opened Jul 4, 2025

• Update devtools documentations with numerical accuracy debugging

  #12216 opened Jul 3, 2025

• Update runtime documentations with new features around program/data separation

  #12215 opened Jul 3, 2025

• Update LLM documents and websites

  #12214 opened Jul 3, 2025

• Temporarily disable benchmark run on Apple private devices

  #12211 opened Jul 3, 2025

• Arm Bare Metal cmake Preset

  #12186 opened Jul 2, 2025

• Fix ios demo app CI

  #12178 opened Jul 2, 2025

• iOS demo app does not build

  #12168 opened Jul 2, 2025

• IndexError in Conv1dToConv2d pass due to incorrect argument count check

  #12161 opened Jul 2, 2025

• Set KleidiAI as default

  #12148 opened Jul 1, 2025

• Update ET to Kleidi Commit

  #12147 opened Jul 1, 2025

• RFC: Create high level CMake targets

  #12138 opened Jul 1, 2025

• XNNPack execution fails on the listed models

  #12134 opened Jul 1, 2025

• Refactor cmake Selective Build Option ROOT_OPS

  #12133 opened Jul 1, 2025

57 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [et] generate debug handle before opeartor decomposition

  #11997 commented on Jul 8, 2025 • 10 new comments

• [Backend Tester] Add FACTO operator test skeleton

  #11953 commented on Jul 8, 2025 • 9 new comments

• Pass Dependencies When Proposing Partitions VIA New Group-Based Partitioner

  #12072 commented on Jul 7, 2025 • 8 new comments

• [Backend Tester] Add compliance suite skeleton and operator tests

  #11960 commented on Jul 2, 2025 • 7 new comments

• Refactor XNN workspace sharing to allow runtime gating

  #11748 commented on Jul 2, 2025 • 3 new comments

• Milestone2.2: Optimize transposes in XNNPACK partition by removing redundant to_copy ops

  #11316 commented on Jul 7, 2025 • 2 new comments

• Arm backend: Make per-channel quantization default

  #11873 commented on Jul 7, 2025 • 1 new comment

• Introduce sym_max and sym_min ops to executorch

  #12037 commented on Jul 2, 2025 • 1 new comment

• Install libcpp from release package other than apt-get

  #11832 commented on Jul 8, 2025 • 1 new comment

• Fix for build issue in greater_lesser_equal ISA proto use

  #12116 commented on Jul 1, 2025 • 1 new comment

• Arm backend: Create ethosu directory

  #11849 commented on Jul 3, 2025 • 0 new comments

• Arm backend: Move Ethos-U backend to generate TOSA-1.0

  #11852 commented on Jul 3, 2025 • 0 new comments

• [XNNPACK] Add support for Linear fused BatchNorm

  #11805 commented on Jul 3, 2025 • 0 new comments

• NXP backend: Add quantization of aten.view

  #11784 commented on Jul 1, 2025 • 0 new comments

• Introduce apply_torch_ops_aten_passes to test mul fusion e2e

  #11741 commented on Jul 3, 2025 • 0 new comments

• NXP backend: Improve quantization annotation process

  #11908 commented on Jul 3, 2025 • 0 new comments

• Expose a method getter api in Module

  #11929 commented on Jul 1, 2025 • 0 new comments

• Split neutron backend test based on executor dependency

  #11934 commented on Jul 7, 2025 • 0 new comments

• Reapply "Implement unary_ufunc functions using elementwise_util (#9386)"

  #11943 commented on Jul 3, 2025 • 0 new comments

• [Backend Tester] Add CoreML tester implementation

  #11959 commented on Jul 2, 2025 • 0 new comments

• [ET-VK][Ops] linear_qta8a_qga4w_qta8o test framework

  #12005 commented on Jul 7, 2025 • 0 new comments

• [ET-VK][Ops] linear_qta8a_qga4w_qta8o impl and shaders

  #12006 commented on Jul 7, 2025 • 0 new comments

• Enable Gemma3 1B on ExecuTorch

  #12048 commented on Jul 1, 2025 • 0 new comments

• Add export recipes for xnnpack (#12069)

  #12070 commented on Jul 3, 2025 • 0 new comments

• [wip] Introduce MergedDataMap

  #12087 commented on Jul 3, 2025 • 0 new comments

• [wip] Add MergedDataMap to method

  #12088 commented on Jul 3, 2025 • 0 new comments

• Add `required_for_abi_source_only = True` to extension/android:executorch_llama

  #12093 commented on Jul 4, 2025 • 0 new comments

• Arm backend: Add Inception v3 test

  #12111 commented on Jul 2, 2025 • 0 new comments

• Use safe_numerics util from PyTorch

  #12125 commented on Jul 3, 2025 • 0 new comments

• Implement common backend operator-level tests

  #11910 commented on Jul 1, 2025 • 0 new comments

• Cross Compile Executorch for Arm Cortex M h/w (Raspberry PI Pico 2) and infer a simple model

  #11913 commented on Jul 1, 2025 • 0 new comments

• QNN GPU or DSP backend issue

  #5914 commented on Jul 2, 2025 • 0 new comments

• executorch model Inference time is higher than the torch model

  #10297 commented on Jul 2, 2025 • 0 new comments

• Python apis for loading and saving .ptd from dictionaries

  #8542 commented on Jul 2, 2025 • 0 new comments

• Shared memory for multiple entry points with XNNPACK delegate

  #11738 commented on Jul 3, 2025 • 0 new comments

• Error on import protable_lib via pybindings

  #9745 commented on Jul 3, 2025 • 0 new comments

• Add merge to NamedDataMap API

  #10551 commented on Jul 3, 2025 • 0 new comments

• Windows build failure: Failed to query buck for sources.

  #9616 commented on Jul 3, 2025 • 0 new comments

• [QCOM] [Llama] the size of w4a16 quantized Llama 3.2 1B Pte is too large

  #10226 commented on Jul 3, 2025 • 0 new comments

• Add GenerateFromPoS in Android LLAMA API

  #8290 commented on Jul 3, 2025 • 0 new comments

• Failed build of apple-ios example

  #11753 commented on Jul 4, 2025 • 0 new comments

• Flatten layer followed by linear layer causes HardFault on Cortex M4F

  #7651 commented on Jul 4, 2025 • 0 new comments

• Support for squeeze and select operators in ExecuTorch

  #12103 commented on Jul 7, 2025 • 0 new comments

• Remove patch for QNN to use the legacy export flow.

  #7373 commented on Jul 7, 2025 • 0 new comments

• Gradle Error while opening the project

  #11749 commented on Jul 7, 2025 • 0 new comments

• MPS delegate crashes on iOS 26

  #11655 commented on Jul 7, 2025 • 0 new comments

• Remove the outdated warning in building from source documentation

  #11229 commented on Jul 7, 2025 • 0 new comments

• [v0.7.0] Release Tracker

  #11075 commented on Jul 8, 2025 • 0 new comments

• Facing an issue with weird model sizes for *.pte model when compared to corresponding torch script model (*.pt)

  #11637 commented on Jul 8, 2025 • 0 new comments

• Query regarding support of Executorch for ARM Ethos-U65 backend

  #9356 commented on Jul 8, 2025 • 0 new comments

• Added tensor's dim order ambiguity check

  #10272 commented on Jul 5, 2025 • 0 new comments

• Remove strictness in export calls

  #10552 commented on Jul 8, 2025 • 0 new comments

• Buckify quantizer

  #10881 commented on Jul 7, 2025 • 0 new comments

• Update remove clone to drop no-op q/dq

  #10920 commented on Jul 7, 2025 • 0 new comments

• [pytorch hash update] update the pinned pytorch hash

  #10955 commented on Jul 8, 2025 • 0 new comments

• Add torchao kernels to xcframework

  #10963 commented on Jul 7, 2025 • 0 new comments

• NXP backend: Add support for depthwise and separable convolution.

  #11215 commented on Jul 7, 2025 • 0 new comments