Merge branch 'main' into full_annotation

Author: oscarandersson8218

.ci/docker/requirements-ci.txt

@@ -16,18 +16,21 @@ hypothesis==6.84.2
 parameterized==0.9.0
 
 # Doc build requirements, same as https://github.com/pytorch/pytorch/blob/main/.ci/docker/requirements-docs.txt
-sphinx==5.3.0
+sphinx==7.2.6
+sphinxcontrib.katex==0.9.10
+breathe==4.36.0  # only if generating C++
+exhale==0.3.7  # only if generating C++ docs
+docutils==0.18.1,<0.21
+sphinx-design==0.6.1
+sphinxcontrib-mermaid==1.0.0
+myst-parser==3.0.1  # if want to contribute in markdown
+sphinx-gallery==0.14.0  # only if hosting interactive tutorials
+sphinx-sitemap==2.7.1
 sphinx-reredirects==0.1.4
-sphinx-gallery==0.14.0
-breathe==4.34.0
-exhale==0.2.3
-docutils==0.16
 matplotlib>=3.9.4
+sphinx-copybutton==0.5.2
 # PyTorch Theme
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git#egg=pytorch_sphinx_theme
-myst-parser==0.18.1
-sphinx_design==0.4.1
-sphinx-copybutton==0.5.0
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
 
 # script unit test requirements
 yaspin==3.1.0

.ci/scripts/test_backend_linux.sh renamed to .ci/scripts/test_backend.sh

@@ -10,16 +10,26 @@ SUITE=$1
 FLOW=$2
 ARTIFACT_DIR=$3
 
-REPORT_FILE="$ARTIFACT_DIR/test-report-$FLOW-$SUITE.csv"
+REPORT_FILE="$ARTIFACT_DIR/test-report-$FLOW-$SUITE.json"
 
 echo "Running backend test job for suite $SUITE, flow $FLOW."
 echo "Saving job artifacts to $ARTIFACT_DIR."
 
-# The generic Linux job chooses to use base env, not the one setup by the image
 eval "$(conda shell.bash hook)"
 CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
 conda activate "${CONDA_ENV}"
 
+if [[ "$(uname)" == "Darwin" ]]; then
+    bash .ci/scripts/setup-conda.sh
+    eval "$(conda shell.bash hook)"
+    CONDA_RUN_CMD="${CONDA_RUN} --no-capture-output"
+    ${CONDA_RUN_CMD} pip install awscli==1.37.21
+    IS_MACOS=1
+else
+    CONDA_RUN_CMD=""
+    IS_MACOS=0
+fi
+
 export PYTHON_EXECUTABLE=python
 
 # CMake options to use, in addition to the defaults.
@@ -50,11 +60,14 @@ if [[ "$FLOW" == *arm* ]]; then
     .ci/scripts/setup-arm-baremetal-tools.sh
 fi
 
-# We need the runner to test the built library.
-PYTHON_EXECUTABLE=python CMAKE_ARGS="$EXTRA_BUILD_ARGS" .ci/scripts/setup-linux.sh --build-tool cmake --build-mode Release --editable true
+if [[ $IS_MACOS -eq 1 ]]; then
+    SETUP_SCRIPT=.ci/scripts/setup-macos.sh
+else
+    SETUP_SCRIPT=.ci/scripts/setup-linux.sh
+fi
+CMAKE_ARGS="$EXTRA_BUILD_ARGS" ${CONDA_RUN_CMD} $SETUP_SCRIPT --build-tool cmake --build-mode Release --editable true
 
 EXIT_CODE=0
-python -m executorch.backends.test.suite.runner $SUITE --flow $FLOW --report "$REPORT_FILE" || EXIT_CODE=$?
-
+${CONDA_RUN_CMD} pytest -c /dev/nul -n auto backends/test/suite/$SUITE/ -m flow_$FLOW --json-report --json-report-file="$REPORT_FILE" || EXIT_CODE=$?
 # Generate markdown summary.
-python -m executorch.backends.test.suite.generate_markdown_summary "$REPORT_FILE" > ${GITHUB_STEP_SUMMARY:-"step_summary.md"} --exit-code $EXIT_CODE
+${CONDA_RUN_CMD} python -m executorch.backends.test.suite.generate_markdown_summary_json "$REPORT_FILE" > ${GITHUB_STEP_SUMMARY:-"step_summary.md"} --exit-code $EXIT_CODE

.ci/scripts/test_backend_macos.sh

@@ -237,7 +237,7 @@ if [[ "${CUSTOM}" == "ON" ]]; then
   EXPORT_ARGS="${EXPORT_ARGS} model.use_sdpa_with_kv_cache=true"
 fi
 if [[ "${QE}" == "ON" ]]; then
-  EXPORT_ARGS="${EXPORT_ARGS} quantization.embedding_quantize=\"8,1024\""
+  EXPORT_ARGS="${EXPORT_ARGS} quantization.embedding_quantize=\"8,768\""
 fi
 if [[ "${MPS}" == "ON" ]]; then
   EXPORT_ARGS="${EXPORT_ARGS} backend.mps.enabled=true model.enable_dynamic_shape=false debug.verbose=true"

.ci/scripts/test_llama.sh

@@ -5,6 +5,7 @@ set -euxo pipefail
 # Args / flags
 # -------------------------
 TEST_WITH_RUNNER=0
+USE_TORCHAO_KERNELS=0
 MODEL_NAME=""
 
 # Parse args
@@ -22,10 +23,14 @@ while [[ $# -gt 0 ]]; do
     --test_with_runner)
       TEST_WITH_RUNNER=1
       ;;
+    --use_torchao_kernels)
+      USE_TORCHAO_KERNELS=1
+      ;;
     -h|--help)
-      echo "Usage: $0 <model_name> [--test_with_runner]"
+      echo "Usage: $0 <model_name> [--test_with_runner] [--use_torchao_kernels]"
       echo "  model_name: qwen3_4b | phi_4_mini"
       echo "  --test_with_runner: build ET + run llama_main to sanity-check the export"
+      echo "  --use_torchao_kernels: use torchao kernels for linear and tied embedding"
       exit 0
       ;;
     *)
@@ -42,6 +47,13 @@ fi
 
 MODEL_OUT=model.pte
 
+
+# Default to XNNPACK
+BACKEND_ARGS="-X --xnnpack-extended-ops"
+if [[ "$USE_TORCHAO_KERNELS" -eq 1 ]]; then
+  BACKEND_ARGS="--use-torchao-kernels"
+fi
+
 case "$MODEL_NAME" in
   qwen3_4b)
     echo "Running Qwen3-4B export..."
@@ -58,12 +70,12 @@ case "$MODEL_NAME" in
       --output_name $MODEL_OUT \
       -kv \
       --use_sdpa_with_kv_cache \
-      -X \
-      --xnnpack-extended-ops \
       --max_context_length 1024 \
       --max_seq_length 1024 \
+      --metadata '{"get_bos_id":199999, "get_eos_ids":[200020,199999]}' \
+      --verbose \
       --dtype fp32 \
-      --metadata '{"get_bos_id":199999, "get_eos_ids":[200020,199999]}'
+      ${BACKEND_ARGS}
     ;;
 
   phi_4_mini)
@@ -81,12 +93,12 @@ case "$MODEL_NAME" in
       --output_name $MODEL_OUT \
       -kv \
       --use_sdpa_with_kv_cache \
-      -X \
-      --xnnpack-extended-ops \
       --max_context_length 1024 \
       --max_seq_length 1024 \
+      --metadata '{"get_bos_id":199999, "get_eos_ids":[200020,199999]}' \
+      --verbose \
       --dtype fp32 \
-      --metadata '{"get_bos_id":199999, "get_eos_ids":[200020,199999]}'
+      ${BACKEND_ARGS}
     ;;
 
   *)
@@ -104,6 +116,10 @@ if [[ $MODEL_SIZE -gt $EXPECTED_MODEL_SIZE_UPPER_BOUND ]]; then
 fi
 
 # Install ET with CMake
+EXECUTORCH_BUILD_KERNELS_TORCHAO="OFF"
+if [[ "$USE_TORCHAO_KERNELS" -eq 1 ]]; then
+  EXECUTORCH_BUILD_KERNELS_TORCHAO="ON"
+fi
 if [[ "$TEST_WITH_RUNNER" -eq 1 ]]; then
   echo "[runner] Building and testing llama_main ..."
     cmake -DPYTHON_EXECUTABLE=python \
@@ -120,6 +136,7 @@ if [[ "$TEST_WITH_RUNNER" -eq 1 ]]; then
         -DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_LLM=ON \
         -DEXECUTORCH_BUILD_KERNELS_LLM=ON \
+        -DEXECUTORCH_BUILD_KERNELS_TORCHAO=${EXECUTORCH_BUILD_KERNELS_TORCHAO} \
         -Bcmake-out .
     cmake --build cmake-out -j16 --config Release --target install
 

.ci/scripts/test_torchao_huggingface_checkpoints.sh

@@ -57,7 +57,7 @@ jobs:
       script: |
         set -eux
 
-        source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
+        source .ci/scripts/test_backend.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
 
   test-backend-macos:
     if: ${{ inputs.run-macos }}
@@ -81,4 +81,4 @@ jobs:
         # This is needed to get the prebuilt PyTorch wheel from S3
         ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
 
-        source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
+        source .ci/scripts/test_backend.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

.github/workflows/_test_backend.yml

@@ -12,27 +12,39 @@ jobs:
       - name: Add open issues and open, non-draft PRs to org project (excluding certain authors)
         uses: actions/github-script@v7
         with:
+          github-token: ${{ secrets.ET_EXT_CONTRIB }}
           script: |
             const projectId = "PVT_kwDOAUB9vs4A_PUL"; // PyTorch org project 136
             const owner = 'pytorch';
             const repo = 'executorch';
 
             // List of authors to exclude
             const excludedAuthors = new Set([
-              "nil-is-all", "cbilgin", "KimishPatel", "psiddh", "digantdesai", "SS-JIA", "ahmtox", "mcr229", "shoumikhin",
+              "nil-is-all", "cbilgin", "kimishpatel", "psiddh", "digantdesai", "SS-JIA", "ahmtox", "mcr229", "shoumikhin",
               "manuelcandales", "metascroy", "cccclai", "rohansjoshi", "kirklandsign", "abhinaykukkadapu", "JacobSzwejbka",
               "Conarnar", "lucylq", "larryliu0820", "BujSet", "Gasoonjia", "Juntian777", "guangy10", "jackzhxng",
               "GregoryComer", "leafs1", "swolchok", "mergennachin", "tarun292", "byjlw", "jathu", "Jack-Khuu", "georgehong",
-              "zhenyan-zhang-meta", "silverguo", "dbort", "jorgep31415", "huydhn", "mcremon-meta", "trivedivivek", "angelayi",
-              "helunwencser", "hsharma35", "zhxchen17", "iseeyuan", "svekars", "nathanaelsee", "dulinriley", "jerryzh168",
+              "zhenyan-zhang-meta", "silverguo", "harishs88ss", "AlannaBurke", "dbort", "huydhn", "mcremon-meta", "trivedivivek", 
+              "angelayi", "helunwencser", "hsharma35", "zhxchen17", "iseeyuan", "svekars", "nathanaelsee", "dulinriley", "jerryzh168",
               "cmodi-meta", "bigfootjon", "sxu", "ydwu4", "Riandy", "tugsbayasgalan", "bsoyluoglu", "yangw-dev", "YIWENX14",
               "namanahuja", "yushangdi", "limintang", "pianpwk", "viveknayakatmeta", "andreanicastro", "JakeStevens",
-              "gmagogsfm", "zonglinpeng", "eigen-k", "derekxu", "salilsdesai", "skrtskrtfb", "pssrawat", "r-barnes", "pytorchbot",
-              "pytorchmergebot", "pytorchupdatebot", "facebook-github-bot", "Erik-Lundell", "zingo", "AdrianLundell",
-              "oscarandersson8218", "per", "Sebastian-Larsson", "SaoirseARM", "robell", "mansnils", "martinlsm", "freddan80",
-              "YufengShi-dudu", "tom-arm", "perheld", "Jerry-Ge", "gggekov", "fumchin", "wwwind", "haowhsu-quic", "shewu-quic",
-              "winskuo-quic", "chunit-quic", "DannyYuyang-quic", "chuntl", "cymbalrush", "DenisVieriu97", "billmguo",
-              "StrycekSimon", "jirioc", "robert-kalmar", "skywall", "neuropilot-captain"
+              "gmagogsfm", "zonglinpeng", "eigen-k", "derekxu", "salilsdesai", "skrtskrtfb", "pssrawat", "r-barnes",
+              "kalpit-meta-1", "Will-MingLun-Li", "KapJI", "piyengar", "j-bahr", "BoyuanFeng", "fgasperij", "DariusHolmgren",
+              "sammarden-meta", "kushrast", "meta-emilian", "Rittzz", "jeanschmidt", "copyrightly", "mikekgfb", "vmpuri",
+              "zonglinpengmeta", "maggiemoss", "aorenste", "hoangminhle98", "Solumin", "meyering", "rchen152",
+              "AishwaryaSivaraman", "migeed-z", "ebgraham", "Esteb37", "nausicaasnow", "Camyll", "ezyang", "huiyujie",
+              "dltn", "cjhopman", "blackm00n", "agunapal", "SamGondelman", "Ninja91", "ivayloen", "DrJessop", "rodrigos01meta",
+              "akrieger", "cmt0", "yiming0416", "ethansfng", "ThomasJannaud", "nirvanagth", "marcinkwiatkowski", "3l1",
+              "omerjerk", "nitish2112", "yipjustin", "ejnguyen", "andrewor14", "phaiting", "mgiordy", "LeeOHzzZ", "adicatana",
+              "Polyomino", "ezrilow", "navsud", "YifanShenSZ", "RdoubleA", "Olivia-liu", "Abhi-hpp", "Vysarat", "azad-meta",
+              "pytorchbot", "pytorchmergebot", "pytorchupdatebot", "facebook-github-bot", "app/dependabot", "Erik-Lundell",
+              "zingo", "AdrianLundell", "oscarandersson8218", "per", "Sebastian-Larsson", "SaoirseARM", "robell", "mansnils",
+              "martinlsm", "freddan80", "YufengShi-dudu", "tom-arm", "perheld", "Jerry-Ge", "gggekov", "fumchin", "wwwind",
+              "benkli01", "Tessil", "maddun01", "Michiel-Olieslagers", "armwaheed", "agrima1304", "emmakujala", "annietllnd", 
+              "haowhsu-quic", "shewu-quic", "winskuo-quic", "chunit-quic", "DannyYuyang-quic", "chuntl", "thchenqti", 
+              "jethroqti", "cymbalrush", "DenisVieriu97", "billmguo", "StrycekSimon", "jirioc", "robert-kalmar", "skywall", 
+              "MartinPavella", "roman-janik-nxp", "novak-vaclav ", "neuropilot-captain", "dijopaul", "cad-rlc", "cad-audio", 
+              "ynimmaga", "daniil-lyakhov", "emmanuel-ferdman", "cavusmustafa", "Jiseong-oh", "alexdean08"
             ]);
 
             async function addItem(contentId, type, number) {
@@ -79,11 +91,10 @@ jobs:
                   owner,
                   repo,
                   state: 'open',
-                  draft: false,
                 }
               );
               for (const pr of prs) {
-                if (!excludedAuthors.has(pr.user.login)) {
+                if (!pr.draft && !excludedAuthors.has(pr.user.login)) {
                   await addItem(pr.node_id, 'pr', pr.number);
                 }
               }

.github/workflows/add-unanswered-to-project.yml

@@ -31,7 +31,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        runner: [linux.2xlarge]
+        runner: [linux.4xlarge]
         docker-image-name: [
           executorch-ubuntu-22.04-gcc9,
           executorch-ubuntu-22.04-clang12,

.github/workflows/docker-builds.yml

@@ -746,8 +746,8 @@ jobs:
         # Install llama requirements
         bash examples/models/llama/install_requirements.sh
 
-        # install a recent version of torchtune.
-        PYTHON_EXECUTABLE=python python -m pip install torchtune==0.7.0.dev20250730  --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+        # install a recent version of torchtune (>= 20250730)
+        PYTHON_EXECUTABLE=python python -m pip install torchtune==0.7.0.dev20250929  --extra-index-url https://download.pytorch.org/whl/nightly/cpu
 
         # run llama runner in eager mode
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama_lora.sh
@@ -1010,6 +1010,7 @@ jobs:
         ./cmake-out/backends/vulkan/test/custom_ops/q4gsw_linear
         ./cmake-out/backends/vulkan/test/custom_ops/choose_qparams_per_row
         ./cmake-out/backends/vulkan/test/custom_ops/qdq8ta_conv2d_activations
+        ./cmake-out/backends/vulkan/test/custom_ops/q8ta_q8ta_q8to_add
 
         # "Classic" Operator tests
         PYTHON_EXECUTABLE=python bash backends/vulkan/test/scripts/test_op.sh --build

.github/workflows/pull.yml

@@ -594,15 +594,22 @@ jobs:
     strategy:
       matrix:
         model: [qwen3_4b, phi_4_mini]
+        runner: [linux.2xlarge]
+        docker-image: [executorch-ubuntu-22.04-clang12]
+        backend: [xnnpack]
         include:
           - model: qwen3_4b
-            test_with_runner: true
+            runner: linux.arm64.2xlarge
+            docker-image: executorch-ubuntu-22.04-gcc11-aarch64
+            backend: torchao
           - model: phi_4_mini
-            test_with_runner: false
+            runner: linux.arm64.2xlarge
+            docker-image: executorch-ubuntu-22.04-gcc11-aarch64
+            backend: torchao
       fail-fast: false
     with:
-      runner: linux.2xlarge
-      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      runner: ${{ matrix.runner }}
+      docker-image: ci-image:${{ matrix.docker-image }}
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 900
@@ -612,9 +619,14 @@ jobs:
         conda activate "${CONDA_ENV}"
 
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake
+
+        if [[ "${{ matrix.backend }}" == "torchao" ]]; then
+          BUILD_TORCHAO_EXPERIMENTAL=1 TORCHAO_BUILD_CPU_AARCH64=1 TORCHAO_BUILD_KLEIDIAI=1 TORCHAO_ENABLE_ARM_NEON_DOT=1 TORCHAO_PARALLEL_BACKEND=OPENMP pip install third-party/ao
+        fi
+
         pip install -U "huggingface_hub[cli]"
 
-        bash .ci/scripts/test_torchao_huggingface_checkpoints.sh ${{ matrix.model }} ${{ matrix.test_with_runner && '--test_with_runner' || '' }}
+        bash .ci/scripts/test_torchao_huggingface_checkpoints.sh ${{ matrix.model }} ${{ matrix.model != 'phi_4_mini' && '--test_with_runner' || '' }}  ${{ matrix.backend == 'torchao' && '--use_torchao_kernels' || '' }}
 
   test-multimodal-macos:
     if: ${{ !github.event.pull_request.head.repo.fork }}