105 Pull requests merged by 35 people

• [Executorch][llm] Enable local global attention in export_llama script

  #10836 merged May 14, 2025

• Change lowbit example to use 4-bit as default in example

  #10865 merged May 14, 2025

• [LlamaDemo] Replace some tokens

  #10863 merged May 14, 2025

• Update CI for HF Optimum models

  #10820 merged May 14, 2025

• Update documents for Express SDK update

  #10462 merged May 14, 2025

• Remove EXECUTORCH_SEPARATE_FLATCC_HOST_PROJECT

  #10860 merged May 13, 2025

• Use the install method for flatc

  #10859 merged May 13, 2025

• Allow graceful handling of cpuinfo init failure

  #10826 merged May 13, 2025

• Arm backend: Decompose sum in pass

  #10852 merged May 13, 2025

• Build flatcc for the host

  #10855 merged May 13, 2025

• Arm backend: Merge decompose/convert meandim pass

  #10844 merged May 13, 2025

• BUCK forward fix on NXP backend

  #10838 merged May 13, 2025

• Qualcomm AI Engine Direct - Flags for CI

  #9536 merged May 13, 2025

• [Executorch][llm] Enable leveraging ring kv cache via module swap

  #10835 merged May 13, 2025

• Add floatValue to ExecuTorch value

  #10823 merged May 13, 2025

• Update ownership for the build system

  #10837 merged May 13, 2025

• [Executorch][llm] Make custom update cache op operate on indices

  #10834 merged May 13, 2025

• Delete executorch_print_configuration_summary

  #10806 merged May 13, 2025

• Arm backend: Fix mypy linting in pre-push

  #10850 merged May 13, 2025

• [Executorch][llm] Add ring buffer based kv cache and mask calculation to MHA

  #10833 merged May 13, 2025

• Move dependent options to default preset

  #10805 merged May 13, 2025

• Arm backend: Refactor pass tests for TOSA V1.0

  #10843 merged May 13, 2025

• Arm backend: Update partitioner de-tagging iteration order

  #10813 merged May 13, 2025

• Use certifi certs for buck download

  #10095 merged May 13, 2025

• Correct model name in examples/arm/run.sh

  #10815 merged May 13, 2025

• [ET-VK] Removing un used push constants for conv2d pw.

  #10841 merged May 13, 2025

• Default to file load mode in module

  #10827 merged May 13, 2025

• [ET-VK] Removing un used push constants for conv2d pw.

  #10814 merged May 13, 2025

• Move simple options to default preset

  #10804 merged May 13, 2025

• Refactor _get_source_transforms to remove args

  #10519 merged May 13, 2025

• [Executorch][llm] Add support for ring kv cache and ring attention

  #10832 merged May 13, 2025

• [Executorch][llm] Enable local global attention in export_llama script

  #10612 merged May 13, 2025

• [Executorch][llm] Enable leveraging ring kv cache via module swap

  #10611 merged May 13, 2025

• [Executorch][llm] Make custom update cache op operate on indices

  #10610 merged May 13, 2025

• [Executorch][llm] Add ring buffer based kv cache and mask calculation to MHA

  #10609 merged May 13, 2025

• [Executorch][llm] Add support for ring kv cache and ring attention

  #10608 merged May 13, 2025

• Delete EXECUTORCH_BUILD_ANDROID_JNI

  #10803 merged May 13, 2025

• Move OPTIMIZE_SIZE to default preset

  #10802 merged May 13, 2025

• Refactor _to_edge_and_lower_llama to remove args

  #10520 merged May 13, 2025

• Move EXECUTORCH_ENABLE_EVENT_TRACER to default preset

  #10801 merged May 12, 2025

• [jit] Remove @torch.jit.export

  #10824 merged May 12, 2025

• Rename "topic: not user facing"

  #10828 merged May 12, 2025

• Rename "topic: not user facing"

  #10791 merged May 12, 2025

• mediatek llama runner use executorch_core

  #10754 merged May 12, 2025

• Android Qwen thinking mode prompt support

  #10668 merged May 12, 2025

• [jit] Remove TorchScript from doc

  #10825 merged May 12, 2025

• Move EXECUTORCH_ENABLE_PROGRAM_VERIFICATION to default preset

  #10800 merged May 12, 2025

• Forward-fixing G3 lt kernel

  #10812 merged May 12, 2025

• Move EXECUTORCH_LOG_LEVEL to default preset

  #10799 merged May 12, 2025

• Xnnpack test for program-data separation

  #10817 merged May 12, 2025

• Update backends-coreml.md

  #10816 merged May 12, 2025

• [ET-VK] Return fence after waiting is done.

  #10808 merged May 12, 2025

• Xnnpack test for program-data separation

  #10532 merged May 12, 2025

• Make a separate target for kernel utils

  #10788 merged May 12, 2025

• Move EXECUTORCH_PAL_DEFAULT to default preset

  #10798 merged May 12, 2025

• [llava] Remove torch.jit.save in llava example

  #10794 merged May 12, 2025

• NXP Backend: Add eIQ Neutron Backend

  #10196 merged May 12, 2025

• Arm backend: Rescale fixes for TOSA 1.0

  #10809 merged May 12, 2025

• Arm backend: Fix ensures check in UnsqueezeScalarPlaceholdersPass

  #10811 merged May 12, 2025

• [ET-VK] Return fence after waiting is done.

  #10787 merged May 12, 2025

• Introduce assertj test lib to make the throw exception test more accu…

  #10779 merged May 10, 2025

• Remove FLATC_EXECUTABLE and the ability to bring your own flatc

  #10781 merged May 9, 2025

• fix bug with sequential backends

  #10708 merged May 9, 2025

• bugfix

  #10793 merged May 9, 2025

• Use torchtune 0.6.1

  #10792 merged May 9, 2025

• Reapply #9841: Migrate elementwise_util callers to the variants with out_dtypes in template arguments

  #10491 merged May 9, 2025

• Reapply #9842: Save some size in dtype_util when dtype selective build is not in use

  #10490 merged May 9, 2025

• Save some size in pattern/{bitwise,comparison}_op.h

  #10489 merged May 9, 2025

• Cortex-M: Use q/dq ops in Arm Ethos Runner

  #10782 merged May 9, 2025

• Arm backend: Suppress colors in pre-push if non-interactive

  #10783 merged May 9, 2025

• fix transpose / permutations fusion pass

  #10780 merged May 9, 2025

• Use std::align_alloc in file_data_loader

  #10660 merged May 9, 2025

• Arm Backend: Use tosa_ref_model only if it is avaiable

  #10778 merged May 9, 2025

• to make TIE quantized conv operator to fall back to hifi quantized conv op instead of cpu op for shapes not supported by the TIE kernel.

  #10770 merged May 9, 2025

• : constant fold None

  #10762 merged May 9, 2025

• Make constant_folding's _DEFAULT_SKIP_TARGETS public

  #10760 merged May 8, 2025

• Extract trace from prepare_and_convert and remove export_program

  #10493 merged May 8, 2025

• Create a macos-arm64 preset

  #10768 merged May 8, 2025

• Convert the unit test from java to kotlin

  #10702 merged May 8, 2025

• Allow options to be set by presets

  #10767 merged May 8, 2025

• Minor vector sizing change.

  #10753 merged May 8, 2025

• Arm backend: Replace asserts with exceptions in permutation code

  #10774 merged May 8, 2025

• Arm backend: Remove redundant validation check for op_where

  #10773 merged May 8, 2025

• Automatically announce declared options

  #10766 merged May 8, 2025

• Arm Backend: Update unit tests for TOSA 1.0

  #10776 merged May 8, 2025

• Arm backend: Add model name to -llama_inputs

  #10775 merged May 8, 2025

• [ET-VK] Implement linear_qcs4w

  #10772 merged May 8, 2025

• [ET-VK] Introduce generic export pass for fusing Q/DQ nodes

  #10771 merged May 8, 2025

• [ET-VK] Implement linear_qcs4w

  #10588 merged May 8, 2025

• [ET-VK] Introduce generic export pass for fusing Q/DQ nodes

  #10525 merged May 8, 2025

• Tests use executorch_core

  #10764 merged May 8, 2025

• Update buck2 to 2025-05-06

  #10742 merged May 8, 2025

• Handle avg_pool2d with padding == 0 as no padding

  #10697 merged May 8, 2025

• Vulkan tests use executorch_core

  #10765 merged May 8, 2025

• Run apple.yml on ciflow/trunk

  #10759 merged May 7, 2025

• Add libextension_flat_tensor.a to build_apple_frameworks.sh

  #10758 merged May 7, 2025

• Add tools/cmake to unittests

  #10752 merged May 7, 2025

• Arm backend: add support for operator @

  #10749 merged May 7, 2025

• Arm backend: Adjust MaxPool2d padding when window is not divisible by stride

  #10751 merged May 7, 2025

• Test.cmake use executorch_core

  #10747 merged May 7, 2025

• Create a helper to define overridable configs

  #10731 merged May 7, 2025

• bump torchao pin

  #10743 merged May 7, 2025

• [cadence] add-delinearize-index-dep

  #10739 merged May 7, 2025

• Arm backend: Replace asserts with ValueError for slicing constraints

  #10748 merged May 7, 2025

• [CMake] llm_runner use executorch_core

  #10698 merged May 7, 2025

28 Pull requests opened by 21 people

• : constant fold None

  #10755 opened May 7, 2025

• move pattern

  #10756 opened May 7, 2025

• Prim ops move 2

  #10763 opened May 7, 2025

• Qualcomm AI Engine Direct - fix for pytorch uplevel

  #10769 opened May 8, 2025

• [ET-VK] Removing descriptor pool intialization from DescriptorPool ctor.

  #10777 opened May 8, 2025

• [ET-VK] Reducing memory wastage by tightening DescriptorPoolConfig values.

  #10784 opened May 9, 2025

• [ET-VK] Moving device capabilities check to DispatchNode and PrepackNode ctor.

  #10785 opened May 9, 2025

• Update export and build from source docs

  #10807 opened May 11, 2025

• Example external model to be used by the ahead of time arm compiler

  #10810 opened May 12, 2025

• Pass to replace Adaptive Avg. Pool with Aten Avg. Pool

  #10818 opened May 12, 2025

• Make qwen3 compatible with ao-quantized checkpoints

  #10822 opened May 12, 2025

• Forward fix on NXP backend

  #10829 opened May 12, 2025

• [ET-VK] custom memory pools

  #10831 opened May 12, 2025

• Support prequant qwen3

  #10839 opened May 13, 2025

• Adding operators: bitwiseand, bitwiseor, bitwisexor, embedding, eq, fmod, ge, gt, hardtanh, le, lt, maskedfill, ne, select_copy and quantisedmatmul

  #10840 opened May 13, 2025

• Hook up PreprocessAll flow to EdgeManager

  #10842 opened May 13, 2025

• Arm backend: Update NEGATE with TOSA 1.0 support

  #10845 opened May 13, 2025

• Arm backend: Add test for DeiT Tiny for TOSA BI

  #10846 opened May 13, 2025

• Arm backend: Reenable test_fuse_const_ops_tosa_BI

  #10847 opened May 13, 2025

• Arm backend: Add DecomposeLinalgVectorNorm pass + tests

  #10848 opened May 13, 2025

• Arm backend: Update operator support for TOSA-1.0+INT+u55

  #10849 opened May 13, 2025

• Arm backend: Refactor misc tests for TOSA V1.0

  #10851 opened May 13, 2025

• Add a pass to fuse mul.Scalar into dequant

  #10853 opened May 13, 2025

• [jit] Remove more reference to TorchScript

  #10856 opened May 13, 2025

• Add pass to convert kwargs to args + populate optional args.

  #10857 opened May 13, 2025

• Allow setting thread count from Java

  #10858 opened May 13, 2025

• Fix broken tests

  #10866 opened May 14, 2025

• Mostly sync BlasKernel.cpp with ATen ReducedPrecisionGemvFastPathKernel

  #10868 opened May 14, 2025

10 Issues closed by 6 people

• [CMake] optimized_kernels and quantized_kernel will depend on portable_kernels.

  #10677 closed May 14, 2025

• Beefing up CONTRIBUTING.md to lower the barrier of entry for external contributors

  #9582 closed May 13, 2025

• [ET-LLM] Use pytorch-labs/tokenizers in ET

  #8376 closed May 13, 2025

• Editable mode is error-ing out with flatc message

  #8784 closed May 13, 2025

• Add complex dtype support to op_sum_dim

  #10431 closed May 13, 2025

• Add complex dtype support to op_mul

  #10430 closed May 13, 2025

• Update buck version when there's an April 4 or newer release & clean up after #9890

  #9919 closed May 12, 2025

• [Build Presets] Create a GitHub workflow foundation

  #10715 closed May 10, 2025

• where is pytorch_tokenizers.tools.llama2c.convert?

  #10571 closed May 8, 2025

• How to use tokenizer.json in ExecuTorch Android demo (without tokenizer.model)?

  #10745 closed May 7, 2025

13 Issues opened by 9 people

• Running Qwen3 XNNPACK on Android fails while setting up pretokenizer

  #10867 opened May 14, 2025

• Make PR label checker non-mandatory

  #10864 opened May 13, 2025

• Automatically specify params and metadata for export LLM

  #10862 opened May 13, 2025

• Allow loading downstream HF repos

  #10861 opened May 13, 2025

• Add first class citizen support for bundling along strings, ints etc. as metadata

  #10854 opened May 13, 2025

• Create a minimal_executor_runner target

  #10830 opened May 12, 2025

• Out-variant kernels are lacking BC protection for new default args being added.

  #10821 opened May 12, 2025

• Consolidate executor_runners

  #10819 opened May 12, 2025

• Unable to run example script to generate Coreml Models

  #10797 opened May 9, 2025

• Create a landing page for executorch mobile

  #10796 opened May 9, 2025

• Remove all instances and usages of torchscript

  #10795 opened May 9, 2025

• Add timestamps for pte generation in CI

  #10761 opened May 7, 2025

• Deploying VITA-1.5 Multimodal Model with ExecuTorch

  #10757 opened May 7, 2025

65 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.

• Recipe and Input class definitions with e2e export

  #10034 commented on May 14, 2025 • 23 new comments

• Migrate ExecuTorch's use of pt2e from torch.ao to torchao

  #10294 commented on May 14, 2025 • 9 new comments

• NXP backend: Enable initial unit tests workflow

  #10258 commented on May 14, 2025 • 6 new comments

• Introducing NXP Neutron runtime

  #10563 commented on May 14, 2025 • 5 new comments

• NXP backend: Create NeutronAtenPassManager with initial BatchNorm fusing passes

  #10579 commented on May 14, 2025 • 5 new comments

• Update install script and building from source docs

  #10652 commented on May 12, 2025 • 4 new comments

• [executorch][android] Add Runtime.java to centralize native library l…

  #10672 commented on May 7, 2025 • 2 new comments

• Qualcomm AI Engine Direct - Enable custom operator

  #8726 commented on May 14, 2025 • 2 new comments

• Use std::string_view and std::optional

  #10541 commented on May 8, 2025 • 2 new comments

• Arm backend: Allocate the scratch buffer runtime rather than in the pte

  #10714 commented on May 14, 2025 • 1 new comment

• [docs][ez] Fix doc build workflow

  #8079 commented on May 13, 2025 • 0 new comments

• Adjust tolerance for quantized XNN conv1d tests

  #8093 commented on May 13, 2025 • 0 new comments

• Revert to use mean_out than mean_dim_out

  #8021 commented on May 13, 2025 • 0 new comments

• remove the exec_aten namespace

  #8018 commented on May 13, 2025 • 0 new comments

• [devtool] create stream_data_sink

  #8604 commented on May 13, 2025 • 0 new comments

• Add small check when input type is a list

  #9186 commented on May 8, 2025 • 0 new comments

• Add some basic xnnpack recipes

  #10035 commented on May 14, 2025 • 0 new comments

• Arm backend: Add TOSA support for GroupNorm

  #10198 commented on May 14, 2025 • 0 new comments

• Fix comment in memory_planning.py

  #8010 commented on May 13, 2025 • 0 new comments

• fix spec_prop_pass

  #7974 commented on May 13, 2025 • 0 new comments

• Added tensor's dim order ambiguity check

  #10272 commented on May 8, 2025 • 0 new comments

• Qualcomm AI Engine Direct - alias_copy op

  #10319 commented on May 13, 2025 • 0 new comments

• Bump PyTorch nightly pin past April 22 2025

  #10362 commented on May 13, 2025 • 0 new comments

• [ExecuTorch][#10447] Extend `PyBundledModule` with `extension.BundledModule`

  #10450 commented on May 13, 2025 • 0 new comments

• Introduce `platform-config` in CompileSpec for MediaTek backend

  #10464 commented on May 12, 2025 • 0 new comments

• Qualcomm AI Engine Direct - xr model enablement (mld_f)

  #10546 commented on May 12, 2025 • 0 new comments

• Arm backend: Fix sigmoid int16 and int32 flakyness

  #10548 commented on May 13, 2025 • 0 new comments

• Update MemoryPlanning Verifier to only assume model has user input if it has at least one tensor input

  #10617 commented on May 14, 2025 • 0 new comments

• Clean up eager quant in llm_export

  #10684 commented on May 9, 2025 • 0 new comments

• Fix preq embedding dtype check

  #10699 commented on May 9, 2025 • 0 new comments

• Add input size validation to Module.execute

  #10701 commented on May 12, 2025 • 0 new comments

• openvino_backend doesn't need to be static only

  #10732 commented on May 7, 2025 • 0 new comments

• [TEST] Split prim ops into its own

  #10741 commented on May 8, 2025 • 0 new comments

• [Build Presets] Create a windows-x86_64 preset

  #10723 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create a linux-x86_64 preset

  #10722 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create an android-x86_64 preset

  #10721 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create an android-arm64-v8a preset

  #10720 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create a ios-simulator-arm64 preset

  #10719 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create an ios-arm64 preset

  #10718 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create a macos-arm64 preset

  #10717 commented on May 9, 2025 • 0 new comments

• [Build Presets] Create foundation and default configurations

  #10716 commented on May 9, 2025 • 0 new comments

• [CMake] Decouple prim_ops from executorch

  #10704 commented on May 9, 2025 • 0 new comments

• [CMake] Duplicated entries in executorch_srcs.cmake

  #10687 commented on May 9, 2025 • 0 new comments

• executorch model Inference time is higher than the torch model

  #10297 commented on May 9, 2025 • 0 new comments

• Request for support of ExecuTorch pip package on linux aarch64

  #10651 commented on May 8, 2025 • 0 new comments

• Question about programmatically running inference on Android-based custom OS with vulkan delegate

  #10602 commented on May 8, 2025 • 0 new comments

• Return "platform not supported" when using PyTorch on intel-based Macbooks

  #9772 commented on May 8, 2025 • 0 new comments

• [Android] Add a Runtime.java

  #10439 commented on May 7, 2025 • 0 new comments

• [CMake] Enable BUILD_SHARED_LIBS=ON flag

  #10676 commented on May 7, 2025 • 0 new comments

• [pytorch hash update] update the pinned pytorch hash

  #4589 commented on May 14, 2025 • 0 new comments

• Building ExecuTorch on RPi5 with Clang 14.0.6 fails due to bfloat incompatibility

  #8924 commented on May 14, 2025 • 0 new comments

• Advice on how to run the training example in Android

  #10593 commented on May 13, 2025 • 0 new comments

• Undefined symbol while compiling runtime for MediaTek

  #10389 commented on May 13, 2025 • 0 new comments

• _load_for_executorch_from_buffer doesn't keep buffer alive

  #10725 commented on May 13, 2025 • 0 new comments

• Need a feature to get etdump while running LLAMA model on qnn with qnn_llama_runner

  #10580 commented on May 13, 2025 • 0 new comments

• [Neutron Backend] Move Neutron Backend to Dim Order Representation usage

  #10711 commented on May 12, 2025 • 0 new comments

• [CMake] Potentially duplicated srcs in llama_runner build

  #10686 commented on May 12, 2025 • 0 new comments

• Format CMakeLists.txt

  #10736 commented on May 12, 2025 • 0 new comments

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

  #10226 commented on May 12, 2025 • 0 new comments

• llama3.2 1B model run on QNN backend produce wrong result

  #5929 commented on May 12, 2025 • 0 new comments

• [QCOM] Support stable diffusion 2.1 on SM8750

  #10209 commented on May 12, 2025 • 0 new comments

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

  #9356 commented on May 12, 2025 • 0 new comments

• Add torchao kernels to xcframework

  #10694 commented on May 12, 2025 • 0 new comments

• Consolidate debug handle infra from Export Graph, torch.ao to ExecuTorch

  #10727 commented on May 9, 2025 • 0 new comments

• Python runtime API for operators

  #10726 commented on May 9, 2025 • 0 new comments