init version

Pull Request resolved: https://github.com/pytorch/pytorch/pull/165116
Approved by: https://github.com/PaulZhang12

.ci/docker/build.sh

@@ -181,7 +181,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
+    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950;gfx1100"
     if [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
     fi
@@ -344,7 +344,7 @@ docker build \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
-       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a;gfx942}" \
+       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}" \
        --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
        --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-27664085f804afc83df26f740bb46c365854f2c4
+7416ffcb92cdbe98d9f97e4e6f95247e46dfc9fd

.ci/docker/libtorch/build.sh

@@ -46,9 +46,9 @@ case ${DOCKER_TAG_PREFIX} in
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-        # add gfx950 conditionally starting in ROCm 7.0
+        # add gfx950, gfx115x conditionally starting in ROCm 7.0
         if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
+            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950;gfx1150;gfx1151"
         fi
         DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg ROCM_VERSION=${GPU_ARCH_VERSION}"
         ;;

.ci/docker/manywheel/Dockerfile_s390x

@@ -115,6 +115,9 @@ RUN env GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=True pip3 install grpcio
 # cmake-3.28.0 from pip for onnxruntime
 RUN python3 -mpip install cmake==3.28.0
 
+ADD ./common/patch_libstdc.sh patch_libstdc.sh
+RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh
+
 # build onnxruntime 1.21.0 from sources.
 # it is not possible to build it from sources using pip,
 # so just build it from upstream repository.

.ci/docker/manywheel/build.sh

@@ -84,9 +84,9 @@ case ${image} in
         DEVTOOLSET_VERSION="11"
         GPU_IMAGE=rocm/dev-almalinux-8:${GPU_ARCH_VERSION}-complete
         PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-        # add gfx950 conditionally starting in ROCm 7.0
+        # add gfx950, gfx115x conditionally starting in ROCm 7.0
         if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
+            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950;gfx1150;gfx1151"
         fi
         DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}"
         ;;

.ci/docker/manywheel/build_scripts/ssl-check.py

@@ -10,11 +10,6 @@ BAD_SSL = "https://self-signed.badssl.com"
 
 print("Testing SSL certificate checking for Python:", sys.version)
 
-if sys.version_info[:2] < (2, 7) or sys.version_info[:2] < (3, 4):
-    print("This version never checks SSL certs; skipping tests")
-    sys.exit(0)
-
-
 EXC = OSError
 
 print(f"Connecting to {GOOD_SSL} should work")

.ci/docker/requirements-ci.txt

@@ -120,9 +120,8 @@ ninja==1.11.1.4
 numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
 numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #Description: Just-In-Time Compiler for Numerical Functions
-#Pinned versions: 0.54.1, 0.49.0, <=0.49.1
+#Pinned versions: 0.55.2, 0.60.0
 #test that import: test_numba_integration.py
-#For numba issue see https://github.com/pytorch/pytorch/issues/51511
 #Need release > 0.61.2 for s390x due to https://github.com/numba/numba/pull/10073
 
 #numpy
@@ -242,10 +241,9 @@ pygments==2.15.0
 #Pinned versions: 14.1.0
 #test that import:
 
-scikit-image==0.19.3 ; python_version < "3.10"
-scikit-image==0.22.0 ; python_version >= "3.10"
+scikit-image==0.22.0
 #Description: image processing routines
-#Pinned versions:
+#Pinned versions: 0.22.0
 #test that import: test_nn.py
 
 #scikit-learn

.ci/magma-rocm/Makefile

@@ -5,7 +5,7 @@ DESIRED_ROCM ?= 7.0
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
-#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201
+#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201
 
 DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	-v $(shell git rev-parse --show-toplevel)/.ci:/builder \
@@ -18,7 +18,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 .PHONY: all
 all: magma-rocm70
 all: magma-rocm64
-all: magma-rocm63
 
 .PHONY:
 clean:
@@ -34,8 +33,3 @@ magma-rocm70:
 magma-rocm64: DESIRED_ROCM := 6.4
 magma-rocm64:
 	$(DOCKER_RUN)
-
-.PHONY: magma-rocm63
-magma-rocm63: DESIRED_ROCM := 6.3
-magma-rocm63:
-	$(DOCKER_RUN)

.ci/pytorch/build.sh

@@ -233,7 +233,9 @@ if [[ "${BUILD_ENVIRONMENT}" != *cuda* ]]; then
   export BUILD_STATIC_RUNTIME_BENCHMARK=ON
 fi
 
-if [[ "$BUILD_ENVIRONMENT" == *-debug* ]]; then
+if [[ "$BUILD_ENVIRONMENT" == *-full-debug* ]]; then
+  export CMAKE_BUILD_TYPE=Debug
+elif [[ "$BUILD_ENVIRONMENT" == *-debug* ]]; then
   export CMAKE_BUILD_TYPE=RelWithAssert
 fi
 
@@ -299,6 +301,11 @@ else
       python -m build --wheel --no-isolation
     fi
     pip_install_whl "$(echo dist/*.whl)"
+    if [[ "$BUILD_ENVIRONMENT" == *full-debug* ]]; then
+      # Regression test for https://github.com/pytorch/pytorch/issues/164297
+      # Torch should be importable and that's about it
+      pushd /; python -c "import torch;print(torch.__config__.show(), torch.randn(5) + 1.7)"; popd
+    fi
 
     if [[ "${BUILD_ADDITIONAL_PACKAGES:-}" == *vision* ]]; then
       install_torchvision

.ci/pytorch/check_binary.sh

@@ -67,7 +67,7 @@ fi
 #       wheels with cxx11-abi
 
 echo "Checking that the gcc ABI is what we expect"
-if [[ "$(uname)" != 'Darwin' &&  "$(uname -m)" != "s390x" ]]; then
+if [[ "$(uname)" != 'Darwin' ]]; then
   # We also check that there are cxx11 symbols in libtorch
   #
   echo "Checking that symbols in libtorch.so have the right gcc abi"

.ci/pytorch/macos-test.sh

@@ -256,7 +256,7 @@ test_torchbench_smoketest() {
   local device=mps
   local dtypes=(undefined float16 bfloat16 notset)
   local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(llama BERT_pytorch dcgan yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor vgg16)
 
   for backend in eager inductor; do
 
@@ -319,7 +319,7 @@ test_aoti_torchbench_smoketest() {
   local device=mps
   local dtypes=(undefined float16 bfloat16 notset)
   local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(llama BERT_pytorch dcgan yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor vgg16)
 
   echo "Launching torchbench inference performance run for AOT Inductor and dtype ${dtype}"
   local dtype_arg="--${dtype}"

.ci/pytorch/test.sh

@@ -838,7 +838,7 @@ test_dynamo_benchmark() {
       elif [[ "${suite}" == "timm_models" ]]; then
         export TORCHBENCH_ONLY_MODELS="inception_v3"
       elif [[ "${suite}" == "torchbench" ]]; then
-        export TORCHBENCH_ONLY_MODELS="hf_Bert"
+        export TORCHBENCH_ONLY_MODELS="BERT_pytorch"
       fi
     fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
@@ -869,13 +869,13 @@ test_inductor_torchbench_smoketest_perf() {
   mkdir -p "$TEST_REPORTS_DIR"
 
   python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --float16 --training \
-    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only hf_Bert \
+    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only BERT_pytorch \
     --output "$TEST_REPORTS_DIR/inductor_training_smoketest.csv"
   # The threshold value needs to be actively maintained to make this check useful
   python benchmarks/dynamo/check_perf_csv.py -f "$TEST_REPORTS_DIR/inductor_training_smoketest.csv" -t 1.4
 
   # Check memory compression ratio for a few models
-  for test in hf_Albert timm_vision_transformer; do
+  for test in BERT_pytorch yolov3; do
     python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --amp --training \
       --disable-cudagraphs --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" \
       --only $test --output "$TEST_REPORTS_DIR/inductor_training_smoketest_$test.csv"
@@ -886,7 +886,7 @@ test_inductor_torchbench_smoketest_perf() {
   done
 
   # Perform some "warm-start" runs for a few huggingface models.
-  for test in AlbertForQuestionAnswering AllenaiLongformerBase DistilBertForMaskedLM DistillGPT2 GoogleFnet YituTechConvBert; do
+  for test in AllenaiLongformerBase DistilBertForMaskedLM DistillGPT2 GoogleFnet YituTechConvBert; do
     python benchmarks/dynamo/huggingface.py --accuracy --training --amp --inductor --device cuda --warm-start-latency \
       --only $test --output "$TEST_REPORTS_DIR/inductor_warm_start_smoketest_$test.csv"
     python benchmarks/dynamo/check_accuracy.py \

.ci/pytorch/win-test-helpers/test_libtorch.bat

@@ -15,37 +15,35 @@ if errorlevel 1 exit /b 1
 if not errorlevel 0 exit /b 1
 
 cd %TMP_DIR_WIN%\build\torch\test
+
+:: Enable delayed variable expansion to make the list
+setlocal enabledelayedexpansion
+set EXE_LIST=
 for /r "." %%a in (*.exe) do (
-    call :libtorch_check "%%~na" "%%~fa"
+  if "%%~na" == "c10_intrusive_ptr_benchmark" (
+    @REM NB: This is not a gtest executable file, thus couldn't be handled by
+    @REM pytest-cpp and is excluded from test discovery by run_test
+    call "%%~fa"
     if errorlevel 1 goto fail
+    if not errorlevel 0 goto fail
+  ) else (
+    if "%%~na" == "verify_api_visibility" (
+      @REM Skip verify_api_visibility as it is a compile-level test
+    ) else (
+      set EXE_LIST=!EXE_LIST! cpp/%%~na
+    )
+  )
 )
 
-goto :eof
-
-:libtorch_check
-
 cd %CWD%
 set CPP_TESTS_DIR=%TMP_DIR_WIN%\build\torch\test
 
-:: Skip verify_api_visibility as it a compile level test
-if "%~1" == "verify_api_visibility" goto :eof
+:: Run python test\run_test.py on the list
+set NO_TD=True && python test\run_test.py --cpp --verbose -i !EXE_LIST!
+if errorlevel 1 goto fail
+if not errorlevel 0 goto fail
 
-echo Running "%~2"
-if "%~1" == "c10_intrusive_ptr_benchmark" (
-  :: NB: This is not a gtest executable file, thus couldn't be handled by pytest-cpp
-  call "%~2"
-  goto :eof
-)
-
-python test\run_test.py --cpp --verbose -i "cpp/%~1"
-if errorlevel 1 (
-  echo %1 failed with exit code %errorlevel%
-  goto fail
-)
-if not errorlevel 0 (
-  echo %1 failed with exit code %errorlevel%
-  goto fail
-)
+goto :eof
 
 :eof
 exit /b 0

.ci/pytorch/win-test.sh

@@ -38,7 +38,7 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move this to .ci/docker/requirements-ci.txt
-python -m pip install "psutil==5.9.1" "pynvml==11.4.1" "pytest-shard==0.1.2"
+python -m pip install "psutil==5.9.1" nvidia-ml-py "pytest-shard==0.1.2"
 
 run_tests() {
     # Run nvidia-smi if available

.circleci/scripts/binary_populate_env.sh

@@ -71,14 +71,7 @@ export PYTORCH_BUILD_NUMBER=1
 
 # Set triton version as part of PYTORCH_EXTRA_INSTALL_REQUIREMENTS
 TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
-
-# Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
-TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"
-
-# CUDA 12.9/13.0 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]] || [[ "$DESIRED_CUDA" == "cu130" ]]; then
-  TRITON_CONSTRAINT="platform_system == 'Linux'"
-fi
+TRITON_CONSTRAINT="platform_system == 'Linux'"
 
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* &&  -n "${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:-}" && ! "$PYTORCH_BUILD_VERSION" =~ .*xpu.* ]]; then
   TRITON_REQUIREMENT="triton==${TRITON_VERSION}; ${TRITON_CONSTRAINT}"

.flake8

@@ -12,7 +12,7 @@ ignore =
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907,B908,B910
+    B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
     # these ignores are from flake8-comprehensions; please fix!
     C407,
     # these ignores are from flake8-logging-format; please fix!

.github/actions/linux-test/action.yml

@@ -274,8 +274,6 @@ runs:
           -w /var/lib/jenkins/workspace \
           "${DOCKER_IMAGE}"
         )
-        # Propagate download.pytorch.org IP to container
-        grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
         echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
         docker exec -t "${container_name}" sh -c "pip install $(echo dist/*.whl)[opt-einsum] && ${TEST_COMMAND}"
 

.github/actions/setup-linux/action.yml

@@ -28,6 +28,10 @@ runs:
         echo "instance-type: $(get_ec2_metadata instance-type)"
         echo "system info $(uname -a)"
 
+    - name: Print GPU info (if present)
+      shell: bash
+      run: if [ -f /usr/bin/nvidia-smi ]; then nvidia-smi; fi
+
     - name: Check if in a container runner
       shell: bash
       id: check_container_runner
@@ -82,37 +86,6 @@ runs:
         # Prune all of the docker images
         docker system prune -af
 
-    - name: Manually resolve download.pytorch.org
-      shell: bash
-      continue-on-error: true
-      run: |
-        set +e
-        set -x
-
-        PT_DOMAIN=download.pytorch.org
-        # TODO: Flaky access to download.pytorch.org https://github.com/pytorch/pytorch/issues/100400,
-        # cleaning this up once the issue is fixed. There are more than one resolved IP here, the last
-        # one is returned at random
-        RESOLVED_IP=$(dig -4 +short "${PT_DOMAIN}" | tail -n1)
-
-        if [ -z "${RESOLVED_IP}" ]; then
-          echo "Couldn't resolve ${PT_DOMAIN}, retrying with Google DNS..."
-          RESOLVED_IP=$(dig -4 +short "${PT_DOMAIN}" @8.8.8.8 | tail -n1)
-
-          if [ -z "${RESOLVED_IP}" ]; then
-            echo "Couldn't resolve ${PT_DOMAIN}, exiting..."
-            exit 1
-          fi
-        fi
-
-        if grep -r "${PT_DOMAIN}" /etc/hosts; then
-          # Clean up any old records first
-          sudo sed -i "/${PT_DOMAIN}/d" /etc/hosts
-        fi
-
-        echo "${RESOLVED_IP} ${PT_DOMAIN}" | sudo tee -a /etc/hosts
-        cat /etc/hosts
-
     - name: Check that the docker daemon is running
       shell: bash
       continue-on-error: true

.github/actions/test-pytorch-binary/action.yml

@@ -33,10 +33,6 @@ runs:
         )
 
         echo "CONTAINER_NAME=${container_name}" >> "$GITHUB_ENV"
-        if [[ "${GPU_ARCH_TYPE}" != "rocm" && "${BUILD_ENVIRONMENT}" != "linux-aarch64-binary-manywheel" && "${BUILD_ENVIRONMENT}" != "linux-s390x-binary-manywheel" && "${GPU_ARCH_TYPE}" != "xpu" ]]; then
-          # Propagate download.pytorch.org IP to container. This is only needed on Linux non aarch64 runner
-          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" bash -c "/bin/cat >> /etc/hosts"
-        fi
 
         docker exec -t -w "${PYTORCH_ROOT}" "${container_name}" bash -c "bash .circleci/scripts/binary_populate_env.sh"
         # Generate test script

.github/pytorch-probot.yml

@@ -30,6 +30,7 @@ ciflow_push_tags:
 - ciflow/riscv64
 - ciflow/rocm
 - ciflow/rocm-mi300
+- ciflow/rocm-mi355
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.github/scripts/github_utils.py

@@ -18,6 +18,7 @@ class GitHubComment:
     body_text: str
     created_at: str
     author_login: str
+    author_url: Optional[str]
     author_association: str
     editor_login: Optional[str]
     database_id: int

.github/scripts/test_check_labels.py

@@ -38,6 +38,7 @@ def mock_get_comments() -> list[GitHubComment]:
             body_text="mock_body_text",
             created_at="",
             author_login="",
+            author_url=None,
             author_association="",
             editor_login=None,
             database_id=1,
@@ -48,6 +49,7 @@ def mock_get_comments() -> list[GitHubComment]:
             body_text=" #" + LABEL_ERR_MSG_TITLE.replace("`", ""),
             created_at="",
             author_login=BOT_AUTHORS[1],
+            author_url=None,
             author_association="",
             editor_login=None,
             database_id=2,

.github/scripts/test_trymerge.py

@@ -32,6 +32,7 @@ from trymerge import (
     main as trymerge_main,
     MandatoryChecksMissingError,
     MergeRule,
+    PostCommentError,
     RE_GHSTACK_DESC,
     read_merge_rules,
     remove_job_name_suffix,
@@ -588,6 +589,23 @@ class TestTryMerge(TestCase):
             self.assertEqual(mock_merge_base, pr.get_merge_base())
             mocked_gh_fetch_merge_base.assert_called_once()
 
+    def test_app_can_revert(self, *args: Any) -> None:
+        pr = GitHubPR("pytorch", "pytorch", 164660)
+        repo = DummyGitRepo()
+        app_comment_id, impostor_comment_id = 3375785595, 3377647892
+        # Check that app can revert
+        self.assertIsNotNone(validate_revert(repo, pr, comment_id=app_comment_id))
+        # But impostor can not
+        self.assertRaises(
+            PostCommentError,
+            lambda: validate_revert(repo, pr, comment_id=impostor_comment_id),
+        )
+        # Despite it's name being the name of the bot
+        self.assertEqual(
+            pr.get_comment_by_id(impostor_comment_id).author_login,
+            "pytorch-auto-revert",
+        )
+
 
 @mock.patch("trymerge.gh_graphql", side_effect=mocked_gh_graphql)
 @mock.patch("trymerge.gh_fetch_merge_base", return_value="")

.github/scripts/trymerge.py

@@ -234,6 +234,7 @@ query ($owner: String!, $name: String!, $number: Int!) {
           createdAt
           author {
             login
+            url
           }
           authorAssociation
           editor {
@@ -1093,6 +1094,7 @@ class GitHubPR:
             body_text=node["bodyText"],
             created_at=node["createdAt"] if "createdAt" in node else "",
             author_login=node["author"]["login"],
+            author_url=node["author"].get("url", None),
             author_association=node["authorAssociation"],
             editor_login=editor["login"] if editor else None,
             database_id=node["databaseId"],
@@ -2029,6 +2031,11 @@ def validate_revert(
     # For some reason, one can not be a member of private repo, only CONTRIBUTOR
     if pr.is_base_repo_private():
         allowed_reverters.append("CONTRIBUTOR")
+    # Special case the pytorch-auto-revert app, whose does not have association
+    # But should be able to issue revert command
+    if comment.author_url == "https://github.com/apps/pytorch-auto-revert":
+        allowed_reverters.append("NONE")
+
     if author_association not in allowed_reverters:
         raise PostCommentError(
             f"Will not revert as @{author_login} is not one of "

.github/workflows/_linux-test.yml

@@ -389,8 +389,6 @@ jobs:
             "${DOCKER_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
-          # Propagate download.pytorch.org IP to container
-          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
           echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
 
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then

.github/workflows/h100-distributed.yml

@@ -37,7 +37,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.12xlarge"
+      runner: "linux.c7i.12xlarge"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90-dist
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '9.0'

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -2,7 +2,7 @@ name: inductor-perf-nightly-h100
 
 on:
   schedule:
-    - cron: 15 0,12 * * 1-6
+    - cron: 15 0 * * 1-6
     - cron: 0 7 * * 0
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs

.github/workflows/inductor-perf-test-nightly-macos.yml

@@ -63,6 +63,7 @@ jobs:
       # Same as the build job
       python-version: 3.12.7
       test-matrix: ${{ needs.macos-perf-py3-arm64-build.outputs.test-matrix }}
+      timeout-minutes: 300
       disable-monitor: false
       monitor-log-interval: 15
       monitor-data-collect-interval: 4

.github/workflows/rocm-mi355.yml

@@ -1,6 +1,9 @@
 name: rocm-mi355
 
 on:
+  push:
+    tags:
+      - ciflow/rocm-mi355/*
   workflow_dispatch:
   schedule:
     - cron: 30 11,1 * * *  # about 4:30am PDT and 6:30pm PDT
@@ -64,5 +67,7 @@ jobs:
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
+      tests-to-include: >-
+                        ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor test_matmul_cuda test_scaled_matmul_cuda'
+                           || '' }}
     secrets: inherit

.github/workflows/rocm.yml

@@ -59,3 +59,29 @@ jobs:
       docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-rocm-py3_10-gfx1100-test:
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3_10-gfx1100
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+        ]}
+      tests-to-include: >
+         test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
+         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
+         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
+         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
+         inductor/test_flex_attention inductor/test_max_autotune
+    secrets: inherit

.github/workflows/trunk.yml

@@ -56,7 +56,7 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       build-generates-artifacts: false
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.4xlarge"
+      runner: "linux.c7i.4xlarge"
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 1 },
@@ -249,3 +249,14 @@ jobs:
       docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-py3_10-gcc11-full-debug-build-only:
+    name: linux-jammy-py3.10-gcc11-full-debug-build-only
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.2xlarge.memory
+      build-environment: linux-jammy-py3.10-gcc11-full-debug-build-only
+      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
+    secrets: inherit

.github/workflows/xpu.yml

@@ -35,7 +35,7 @@ jobs:
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-1-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-1-py3
-      runner: linux.12xlarge
+      runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
@@ -56,7 +56,7 @@ jobs:
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3
-      runner: linux.12xlarge
+      runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },

.gitignore

@@ -88,7 +88,7 @@ torch_compile_debug/
 # Listed manually because some files in this directory are not generated
 torch/testing/_internal/generated/annotated_fn_args.py
 torch/testing/_internal/data/*.pt
-torch/csrc/api/include/torch/version.h
+torch/headeronly/version.h
 torch/csrc/cudnn/cuDNN.cpp
 torch/csrc/generated
 torch/csrc/generic/TensorMethods.cpp

.lintrunner.toml

@@ -28,7 +28,7 @@ exclude_patterns = [
     'torch/lib/**',
     'venv/**',
     '**/*.pyi',
-    "tools/experimental/dynamic_shapes/torchfuzz/**",
+    "tools/experimental/torchfuzz/**",
     'tools/test/test_selective_build.py',
 ]
 command = [
@@ -198,7 +198,7 @@ exclude_patterns = [
     'tools/test/gen_operators_yaml_test.py',
     'tools/test/gen_oplist_test.py',
     'tools/test/test_selective_build.py',
-    'tools/experimental/dynamic_shapes/torchfuzz/**',
+    'tools/experimental/torchfuzz/**',
 ]
 command = [
     'python3',

BUILD.bazel

@@ -13,6 +13,9 @@ load(":build_variables.bzl", "jit_core_sources", "lazy_tensor_ts_sources", "libt
 load(":ufunc_defs.bzl", "aten_ufunc_generated_cpu_kernel_sources", "aten_ufunc_generated_cpu_sources", "aten_ufunc_generated_cuda_sources")
 load("//:tools/bazel.bzl", "rules")
 
+# Export files for use by torch/headeronly (where version.h generation now lives)
+exports_files(["version.txt"])
+
 define_targets(rules = rules)
 
 COMMON_COPTS = [
@@ -690,7 +693,9 @@ cc_library(
             "torch/csrc/*/generated/*.h",
             "torch/csrc/jit/serialization/mobile_bytecode_generated.h",
         ] + torch_cuda_headers,
-    ) + GENERATED_AUTOGRAD_CPP + [":version_h"],
+    ) + GENERATED_AUTOGRAD_CPP + [
+        "//torch/headeronly:version_h",
+    ],
     includes = [
         "third_party/kineto/libkineto/include",
         "torch/csrc",

CMakeLists.txt

@@ -388,9 +388,9 @@ cmake_dependent_option(USE_PRIORITIZED_TEXT_FOR_LD "Use prioritized text linker
 
 option(USE_MIMALLOC "Use mimalloc" OFF)
 # Enable third party mimalloc library to improve memory allocation performance
-# on Windows.
+# on Windows and AArch64.
 option(USE_MIMALLOC_ON_MKL "Use mimalloc on MKL" OFF)
-if(WIN32)
+if(WIN32 OR (CPU_AARCH64 AND NOT APPLE))
   set(USE_MIMALLOC ON)
 
   # Not enable USE_MIMALLOC_ON_MKL due to it caused issue:

aten/src/ATen/BlasBackend.h

@@ -28,4 +28,19 @@ inline std::ostream& operator<<(std::ostream& stream, at::BlasBackend backend) {
   return stream << BlasBackendToString(backend);
 }
 
+namespace blas {
+
+enum class ScalingType : std::uint8_t {
+  TensorWise, // fp32 scales
+  RowWise, // fp32 scales
+  BlockWise1x16, // fp8_e4m3fn scales
+  BlockWise1x32, // fp8_e8m0fnu scales
+  BlockWise1x128, // fp32 scales
+  BlockWise128x128, // fp32 scales
+};
+
+enum class SwizzleType : std::uint8_t { NO_SWIZZLE = 0, SWIZZLE_32_4_4 = 1 };
+
+} // namespace blas
+
 } // namespace at

aten/src/ATen/CPUApplyUtils.h

@@ -144,8 +144,7 @@ inline std::string _all_equal_numel_error(at::ArrayRef<Tensor> tensors) {
 inline bool _apply_preamble(ArrayRef<Tensor> tensors) {
   checkDeviceType("CPU_tensor_apply", tensors, kCPU);
   checkLayout("CPU_tensor_apply", tensors, kStrided);
-  if (!_all_equal_numel(tensors))
-    TORCH_CHECK(false, _all_equal_numel_error(tensors));
+  TORCH_CHECK(_all_equal_numel(tensors), _all_equal_numel_error(tensors));
   // An empty tensor has no elements
   for (auto& t : tensors)
     if (t.numel() == 0)

aten/src/ATen/Context.cpp

@@ -483,8 +483,8 @@ at::BlasBackend Context::blasPreferredBackend() {
 #if ROCM_VERSION >= 60300
           "gfx1100", "gfx1101", "gfx1200", "gfx1201", "gfx908",
 #endif
-#if ROCM_VERSION >= 60500
-          "gfx950"
+#if ROCM_VERSION >= 70000
+          "gfx950", "gfx1150", "gfx1151"
 #endif
       };
       for (auto index: c10::irange(detail::getCUDAHooks().deviceCount())) {
@@ -587,20 +587,33 @@ void Context::setROCmFAPreferredBackend(at::ROCmFABackend b) {
   rocm_fa_preferred_backend = b;
 }
 
-bool Context::allowFP16ReductionCuBLAS() const {
+CuBLASReductionOption Context::allowFP16ReductionCuBLAS() const {
   return allow_fp16_reduction_cublas;
 }
 
-void Context::setAllowFP16ReductionCuBLAS(bool b) {
-  allow_fp16_reduction_cublas = b;
+CuBLASReductionOption inline get_reduction_option(bool allow_reduced_precision, bool allow_splitk) {
+  TORCH_CHECK(
+      !(allow_reduced_precision && !allow_splitk),
+      "allow_splitk=False is not supported when reduced precision reductions are enabled");
+  if (allow_reduced_precision) {
+    return CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  } else if (allow_splitk) {
+    return CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK;
+  } else {
+    return CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK;
+  }
 }
 
-bool Context::allowBF16ReductionCuBLAS() const {
+void Context::setAllowFP16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_fp16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
+}
+
+CuBLASReductionOption Context::allowBF16ReductionCuBLAS() const {
   return allow_bf16_reduction_cublas;
 }
 
-void Context::setAllowBF16ReductionCuBLAS(bool b) {
-  allow_bf16_reduction_cublas = b;
+void Context::setAllowBF16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_bf16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
 }
 
 bool Context::allowFP16AccumulationCuBLAS() const {

aten/src/ATen/Context.h

@@ -38,6 +38,12 @@ namespace at {
 class Tensor;
 
 enum class TORCH_API Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };
+
+enum class CuBLASReductionOption : uint8_t {
+  AllowReducedPrecisionWithSplitK = 0,
+  DisallowReducedPrecisionAllowSplitK = 1,
+  DisallowReducedPrecisionDisallowSplitK = 2,
+};
 enum class TORCH_API Float32Backend { GENERIC, CUDA, MKLDNN };
 enum class TORCH_API Float32Op { ALL, CONV, RNN, MATMUL };
 enum class TORCH_API Float32Precision { NONE, IEEE, TF32, BF16 };
@@ -220,15 +226,15 @@ class TORCH_API Context {
   bool userEnabledMkldnn() const;
   void setUserEnabledMkldnn(bool e);
   bool benchmarkCuDNN() const;
-  void setBenchmarkCuDNN(bool);
+  void setBenchmarkCuDNN(bool /*b*/);
   int benchmarkLimitCuDNN() const;
-  void setBenchmarkLimitCuDNN(int);
+  void setBenchmarkLimitCuDNN(int /*b*/);
   bool immediateMiopen() const;
-  void setImmediateMiopen(bool);
+  void setImmediateMiopen(bool /*b*/);
   bool deterministicCuDNN() const;
-  void setDeterministicCuDNN(bool);
+  void setDeterministicCuDNN(bool /*b*/);
   bool deterministicMkldnn() const;
-  void setDeterministicMkldnn(bool);
+  void setDeterministicMkldnn(bool /*b*/);
   bool userEnabledNNPACK() const;
   void setUserEnabledNNPACK(bool e);
 
@@ -246,32 +252,32 @@ class TORCH_API Context {
   void setSDPPriorityOrder(const std::vector<int64_t>& order);
   std::array<at::SDPBackend, at::num_sdp_backends> sDPPriorityOrder();
 
-  void setSDPUseFlash(bool);
+  void setSDPUseFlash(bool /*e*/);
   bool userEnabledFlashSDP() const;
 
-  void setSDPUseMemEfficient(bool);
+  void setSDPUseMemEfficient(bool /*e*/);
   bool userEnabledMemEfficientSDP() const;
 
-  void setSDPUseMath(bool);
+  void setSDPUseMath(bool /*e*/);
   bool userEnabledMathSDP() const;
 
-  void setSDPUseCuDNN(bool);
+  void setSDPUseCuDNN(bool /*e*/);
   bool userEnabledCuDNNSDP() const;
 
-  void setAllowFP16BF16ReductionMathSDP(bool);
+  void setAllowFP16BF16ReductionMathSDP(bool /*e*/);
   bool allowFP16BF16ReductionMathSDP() const;
 
-  void setSDPUseOverrideable(bool);
+  void setSDPUseOverrideable(bool /*e*/);
   bool userEnabledOverrideableSDP() const;
 
   at::LinalgBackend linalgPreferredBackend() const;
-  void setLinalgPreferredBackend(at::LinalgBackend);
+  void setLinalgPreferredBackend(at::LinalgBackend /*b*/);
 
   at::BlasBackend blasPreferredBackend();
-  void setBlasPreferredBackend(at::BlasBackend);
+  void setBlasPreferredBackend(at::BlasBackend /*b*/);
 
   at::ROCmFABackend getROCmFAPreferredBackend();
-  void setROCmFAPreferredBackend(at::ROCmFABackend);
+  void setROCmFAPreferredBackend(at::ROCmFABackend /*b*/);
 
   // Note [Enabling Deterministic Operations]
   // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -304,9 +310,9 @@ class TORCH_API Context {
 
   bool deterministicAlgorithms() const;
   bool deterministicAlgorithmsWarnOnly() const;
-  void setDeterministicAlgorithms(bool, bool);
+  void setDeterministicAlgorithms(bool /*b*/, bool /*warn_only*/);
   bool deterministicFillUninitializedMemory() const;
-  void setDeterministicFillUninitializedMemory(bool);
+  void setDeterministicFillUninitializedMemory(bool /*b*/);
 
   // Note [Writing Nondeterministic Operations]
   // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -350,19 +356,23 @@ class TORCH_API Context {
       Float32Op op,
       Float32Precision p);
   bool allowTF32CuDNN(std::optional<Float32Op> op = std::nullopt) const;
-  void setAllowTF32CuDNN(bool);
+  void setAllowTF32CuDNN(bool /*b*/);
   bool allowTF32OneDNN() const;
-  void setAllowTF32OneDNN(bool);
+  void setAllowTF32OneDNN(bool /*b*/);
   bool allowTF32CuBLAS() const;
-  void setAllowTF32CuBLAS(bool);
+  void setAllowTF32CuBLAS(bool /*b*/);
   Float32MatmulPrecision float32MatmulPrecision() const;
   Float32Precision float32Precision(Float32Backend backend, Float32Op op) const;
-  bool allowFP16ReductionCuBLAS() const;
-  void setAllowFP16ReductionCuBLAS(bool);
-  bool allowBF16ReductionCuBLAS() const;
-  void setAllowBF16ReductionCuBLAS(bool);
+  CuBLASReductionOption allowFP16ReductionCuBLAS() const;
+  void setAllowFP16ReductionCuBLAS(
+      bool allow_reduced_precision,
+      bool allow_splitk = true);
+  CuBLASReductionOption allowBF16ReductionCuBLAS() const;
+  void setAllowBF16ReductionCuBLAS(
+      bool allow_reduced_precision,
+      bool allow_splitk = true);
   bool allowFP16AccumulationCuBLAS() const;
-  void setAllowFP16AccumulationCuBLAS(bool);
+  void setAllowFP16AccumulationCuBLAS(bool /*b*/);
 
   // Matmuls can use a so-called "persistent" kernel which launches one CUDA
   // block for each SM on the GPU, and each block then iterates over multiple
@@ -374,7 +384,7 @@ class TORCH_API Context {
   // to make matmuls target only a subset of the SMs, so they can fully schedule
   // even next to a comms kernel, and only be a few percent slower.
   std::optional<int32_t> _SMCarveout_EXPERIMENTAL() const;
-  void _setSMCarveout_EXPERIMENTAL(std::optional<int32_t>);
+  void _setSMCarveout_EXPERIMENTAL(std::optional<int32_t> /*c*/);
 
   at::QEngine qEngine() const;
   void setQEngine(at::QEngine e);
@@ -395,7 +405,7 @@ class TORCH_API Context {
   void setDefaultMobileCPUAllocator();
   void unsetDefaultMobileCPUAllocator();
   bool allowFP16ReductionCPU() const;
-  void setAllowFP16ReductionCPU(bool);
+  void setAllowFP16ReductionCPU(bool /*b*/);
 
   // Preserved for BC
   void lazyInitCUDA() {
@@ -452,8 +462,10 @@ class TORCH_API Context {
       : at::Float32MatmulPrecision::HIGHEST;
   int benchmark_limit_cudnn = 10;
   bool allow_tf32_cudnn = true;
-  bool allow_fp16_reduction_cublas = true;
-  bool allow_bf16_reduction_cublas = true;
+  CuBLASReductionOption allow_fp16_reduction_cublas =
+      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  CuBLASReductionOption allow_bf16_reduction_cublas =
+      CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
   bool allow_fp16_accumulation_cublas = false;
   std::optional<int32_t> sm_carveout = std::nullopt;
   bool enabled_mkldnn = true;

aten/src/ATen/EmptyTensor.h

@@ -16,8 +16,8 @@ inline void check_size_nonnegative(ArrayRef<int64_t> size) {
 
 inline void check_size_nonnegative(ArrayRef<c10::SymInt> size) {
   for (const auto& x : size) {
-    TORCH_CHECK(
-        x.expect_size(__FILE__, __LINE__),
+    TORCH_SYM_CHECK(
+        x.sym_ge(0),
         "Trying to create tensor with negative dimension ",
         x,
         ": ",

aten/src/ATen/MapAllocator.cpp

@@ -62,7 +62,7 @@ constexpr const char* unknown_eventname = "eventname not specified";
 #endif
 }  // namespace (anonymous)
 
-MapAllocator::MapAllocator(WithFd, std::string_view filename, int fd, int flags, size_t size)
+MapAllocator::MapAllocator(WithFd /*unused*/, std::string_view filename, int fd, int flags, size_t size)
   : filename_(filename.empty() ? unknown_filename : filename)
   , size_(0) // to be filled later
 #ifdef _WIN32
@@ -494,7 +494,7 @@ RefcountedMapAllocator::RefcountedMapAllocator(const char *filename, int flags,
 
     initializeAlloc();
 }
-RefcountedMapAllocator::RefcountedMapAllocator(WithFd, const char *filename, int fd, int flags, size_t size)
+RefcountedMapAllocator::RefcountedMapAllocator(WithFd /*unused*/, const char *filename, int fd, int flags, size_t size)
   : RefcountedMapAllocatorArgCheck(flags)
   , MapAllocator(WITH_FD, filename, flags, fd, size + map_alloc_alignment) {
 
@@ -614,7 +614,7 @@ at::DataPtr MapAllocator::makeDataPtr(std::string_view filename, int flags, size
   return {context->data(), context, &deleteMapAllocator, at::DeviceType::CPU};
 }
 
-at::DataPtr MapAllocator::makeDataPtr(WithFd, const char *filename, int fd, int flags, size_t size, size_t* actual_size_out) {
+at::DataPtr MapAllocator::makeDataPtr(WithFd /*unused*/, const char *filename, int fd, int flags, size_t size, size_t* actual_size_out) {
   auto* context = new MapAllocator(WITH_FD, filename, fd, flags, size);
   if (actual_size_out) *actual_size_out = context->size();
   return {context->data(), context, &deleteMapAllocator, at::DeviceType::CPU};
@@ -626,7 +626,7 @@ at::DataPtr RefcountedMapAllocator::makeDataPtr(const char *filename, int flags,
   return {context->data(), context, &deleteRefcountedMapAllocator, at::DeviceType::CPU};
 }
 
-at::DataPtr RefcountedMapAllocator::makeDataPtr(WithFd, const char *filename, int fd, int flags, size_t size, size_t* actual_size_out) {
+at::DataPtr RefcountedMapAllocator::makeDataPtr(WithFd /*unused*/, const char *filename, int fd, int flags, size_t size, size_t* actual_size_out) {
   auto* context = new RefcountedMapAllocator(WITH_FD, filename, fd, flags, size);
   if (actual_size_out) *actual_size_out = context->size() - map_alloc_alignment;
   return {context->data(), context, &deleteRefcountedMapAllocator, at::DeviceType::CPU};

aten/src/ATen/MapAllocator.h

@@ -25,7 +25,7 @@ class TORCH_API MapAllocator {
  public:
   MapAllocator(std::string_view filename, int flags, size_t size);
   MapAllocator(
-      WithFd,
+      WithFd /*unused*/,
       std::string_view filename,
       int fd,
       int flags,
@@ -59,14 +59,14 @@ class TORCH_API MapAllocator {
     return flags_;
   }
 
-  static MapAllocator* fromDataPtr(const at::DataPtr&);
+  static MapAllocator* fromDataPtr(const at::DataPtr& /*dptr*/);
   static at::DataPtr makeDataPtr(
       std::string_view filename,
       int flags,
       size_t size,
       size_t* actual_size_out);
   static at::DataPtr makeDataPtr(
-      WithFd,
+      WithFd /*unused*/,
       const char* filename,
       int fd,
       int flags,
@@ -105,13 +105,13 @@ class TORCH_API RefcountedMapAllocator : private RefcountedMapAllocatorArgCheck,
  public:
   RefcountedMapAllocator(const char* filename, int flags, size_t size);
   RefcountedMapAllocator(
-      WithFd,
+      WithFd /*unused*/,
       const char* filename,
       int fd,
       int flags,
       size_t size);
 
-  static RefcountedMapAllocator* fromDataPtr(const at::DataPtr&);
+  static RefcountedMapAllocator* fromDataPtr(const at::DataPtr& /*dptr*/);
   RefcountedMapAllocator(const RefcountedMapAllocator&) = delete;
   RefcountedMapAllocator(RefcountedMapAllocator&&) = delete;
   RefcountedMapAllocator& operator=(const RefcountedMapAllocator&) = delete;
@@ -122,7 +122,7 @@ class TORCH_API RefcountedMapAllocator : private RefcountedMapAllocatorArgCheck,
       size_t size,
       size_t* actual_size_out);
   static at::DataPtr makeDataPtr(
-      WithFd,
+      WithFd /*unused*/,
       const char* filename,
       int fd,
       int flags,

aten/src/ATen/NestedTensorImpl.cpp

@@ -273,7 +273,7 @@ c10::SymInt NestedTensorImpl::sym_numel_custom() const {
   return NestedTensorImpl::numel_custom();
 }
 
-c10::SymBool NestedTensorImpl::sym_is_contiguous_custom(MemoryFormat) const {
+c10::SymBool NestedTensorImpl::sym_is_contiguous_custom(MemoryFormat /*memory_format*/) const {
   return nested_tensor_impl_is_contiguous(this);
 }
 IntArrayRef NestedTensorImpl::sizes_custom() const {

aten/src/ATen/NestedTensorImpl.h

@@ -115,7 +115,8 @@ struct TORCH_API NestedTensorImpl : public c10::TensorImpl {
   // with real implementations
   int64_t numel_custom() const override;
   c10::SymInt sym_numel_custom() const override;
-  c10::SymBool sym_is_contiguous_custom(MemoryFormat) const override;
+  c10::SymBool sym_is_contiguous_custom(
+      MemoryFormat /*memory_format*/) const override;
   int64_t size_custom(int64_t d) const override {
     return this->size(d);
   }

aten/src/ATen/Parallel.h

@@ -14,7 +14,7 @@ inline int64_t divup(int64_t x, int64_t y) {
 TORCH_API void init_num_threads();
 
 // Sets the number of threads to be used in parallel region
-TORCH_API void set_num_threads(int);
+TORCH_API void set_num_threads(int /*nthreads*/);
 
 // Returns the maximum number of threads that may be used in a parallel region
 TORCH_API int get_num_threads();
@@ -37,7 +37,7 @@ inline void lazy_init_num_threads() {
   }
 }
 
-TORCH_API void set_thread_num(int);
+TORCH_API void set_thread_num(int /*id*/);
 
 class TORCH_API ThreadIdGuard {
  public:
@@ -130,7 +130,7 @@ inline scalar_t parallel_reduce(
 TORCH_API std::string get_parallel_info();
 
 // Sets number of threads used for inter-op parallelism
-TORCH_API void set_num_interop_threads(int);
+TORCH_API void set_num_interop_threads(int /*nthreads*/);
 
 // Returns the number of threads used for inter-op parallelism
 TORCH_API size_t get_num_interop_threads();

aten/src/ATen/SparseCsrTensorImpl.cpp

@@ -252,7 +252,7 @@ void SparseCsrTensorImpl::set_stride(int64_t dim, int64_t new_stride) {
 void SparseCsrTensorImpl::set_storage_offset(int64_t storage_offset) {
   TORCH_CHECK(false, "Sparse ", at::sparse_csr::layoutToString(layout_, /*upper=*/true), " tensors do not have set_storage_offset.");
 }
-c10::SymBool SparseCsrTensorImpl::sym_is_contiguous_custom(MemoryFormat) const {
+c10::SymBool SparseCsrTensorImpl::sym_is_contiguous_custom(MemoryFormat /*memory_format*/) const {
   TORCH_CHECK(false, "Sparse ", at::sparse_csr::layoutToString(layout_, /*upper=*/true), " tensors do not have is_contiguous");
 }
 } // namespace at

aten/src/ATen/SparseCsrTensorImpl.h

@@ -32,10 +32,10 @@ struct TORCH_API SparseCsrTensorImpl : public TensorImpl {
 
  public:
   explicit SparseCsrTensorImpl(
-      at::DispatchKeySet,
+      at::DispatchKeySet /*key_set*/,
       at::Device device,
       Layout layout,
-      const caffe2::TypeMeta);
+      const caffe2::TypeMeta /*data_type*/);
 
   void resize_(int64_t nnz, IntArrayRef size);
   void resize_and_clear_(
@@ -86,7 +86,8 @@ struct TORCH_API SparseCsrTensorImpl : public TensorImpl {
  protected:
   IntArrayRef strides_custom() const override;
   SymIntArrayRef sym_strides_custom() const override;
-  SymBool sym_is_contiguous_custom(MemoryFormat) const override;
+  SymBool sym_is_contiguous_custom(
+      MemoryFormat /*memory_format*/) const override;
 
  public:
   void set_size(int64_t dim, int64_t new_size) override;

aten/src/ATen/SparseTensorImpl.h

@@ -46,7 +46,9 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
 
  public:
   // Public for now...
-  explicit SparseTensorImpl(at::DispatchKeySet, const caffe2::TypeMeta);
+  explicit SparseTensorImpl(
+      at::DispatchKeySet /*key_set*/,
+      const caffe2::TypeMeta /*data_type*/);
 
   void release_resources() override;
 
@@ -229,14 +231,14 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
   }
 
   void resize_(int64_t sparse_dim, int64_t dense_dim, ArrayRef<int64_t> size) {
-    return _resize_(sparse_dim, dense_dim, size);
+    _resize_(sparse_dim, dense_dim, size);
   }
 
   void resize_(
       int64_t sparse_dim,
       int64_t dense_dim,
       ArrayRef<c10::SymInt> size) {
-    return _resize_(sparse_dim, dense_dim, size);
+    _resize_(sparse_dim, dense_dim, size);
   }
 
   // NOTE: this function will resize the sparse tensor and also set `indices`
@@ -384,8 +386,8 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
 
  private:
   explicit SparseTensorImpl(
-      at::DispatchKeySet,
-      const caffe2::TypeMeta,
+      at::DispatchKeySet /*key_set*/,
+      const caffe2::TypeMeta /*data_type*/,
       at::Tensor indices,
       at::Tensor values);
 

aten/src/ATen/TensorIndexing.cpp

@@ -59,7 +59,7 @@ static inline void set_item(const Tensor& self, ArrayRef<TensorIndex> indices, c
     }
   }
 
-  return set_item(self, indices, value);
+  set_item(self, indices, value);
 }
 
 } // namespace indexing

aten/src/ATen/TensorIndexing.h

@@ -112,10 +112,10 @@ TORCH_API std::ostream& operator<<(std::ostream& stream, const Slice& slice);
 // `torch.tensor([1, 2])`) | `torch::tensor({1, 2})`
 struct TORCH_API TensorIndex final {
   // Case 1: `at::indexing::None`
-  TensorIndex(std::nullopt_t) : type_(TensorIndexType::None) {}
+  TensorIndex(std::nullopt_t /*unused*/) : type_(TensorIndexType::None) {}
 
   // Case 2: "..." / `at::indexing::Ellipsis`
-  TensorIndex(at::indexing::EllipsisIndexType)
+  TensorIndex(at::indexing::EllipsisIndexType /*unused*/)
       : type_(TensorIndexType::Ellipsis) {}
   TensorIndex(const char* str) : TensorIndex(at::indexing::Ellipsis) {
     TORCH_CHECK_VALUE(

aten/src/ATen/TensorIterator.cpp

@@ -765,7 +765,8 @@ void TensorIteratorBase::for_each(loop2d_t loop, int64_t grain_size) {
   if (numel == 0) {
     return;
   } else if (numel < grain_size || at::get_num_threads() == 1) {
-    return serial_for_each(loop, {0, numel});
+    serial_for_each(loop, {0, numel});
+    return;
   } else {
     at::parallel_for(0, numel, grain_size, [&](int64_t begin, int64_t end) {
       serial_for_each(loop, {begin, end});

aten/src/ATen/TensorIterator.h

@@ -250,7 +250,7 @@ struct TORCH_API TensorIteratorBase : public impl::MetaBase {
   using PtrVector = SmallVector<char*, 4>;
   using StrideVector = SmallVector<int64_t, 6>;
 
-  void build(TensorIteratorConfig&);
+  void build(TensorIteratorConfig& /*config*/);
 
   // The inner-loop function operates on the fastest moving dimension. It
   // implements element-wise operations in terms of 1-d strided tensors.
@@ -618,20 +618,20 @@ struct TORCH_API TensorIteratorBase : public impl::MetaBase {
 #undef TORCH_DISALLOW_TEMPORARIES
  protected:
   // Mutable reference as it moves tensors out of TensorIteratorConfig
-  void populate_operands(TensorIteratorConfig&);
+  void populate_operands(TensorIteratorConfig& /*config*/);
   void mark_outputs();
-  void mark_resize_outputs(const TensorIteratorConfig&);
-  void compute_mem_overlaps(const TensorIteratorConfig&);
-  void compute_shape(const TensorIteratorConfig&);
-  void compute_strides(const TensorIteratorConfig&);
+  void mark_resize_outputs(const TensorIteratorConfig& /*config*/);
+  void compute_mem_overlaps(const TensorIteratorConfig& /*config*/);
+  void compute_shape(const TensorIteratorConfig& /*config*/);
+  void compute_strides(const TensorIteratorConfig& /*config*/);
   void reorder_dimensions();
   void permute_dimensions(IntArrayRef perm);
-  void compute_types(const TensorIteratorConfig&);
+  void compute_types(const TensorIteratorConfig& /*config*/);
   ScalarType compute_common_dtype();
   void allocate_or_resize_outputs();
-  bool fast_set_up(const TensorIteratorConfig&);
-  FastSetupType compute_fast_setup_type(const TensorIteratorConfig&);
-  void compute_names(const TensorIteratorConfig&);
+  bool fast_set_up(const TensorIteratorConfig& /*config*/);
+  FastSetupType compute_fast_setup_type(const TensorIteratorConfig& /*config*/);
+  void compute_names(const TensorIteratorConfig& /*config*/);
   void propagate_names_to_outputs();
   void coalesce_dimensions();
 

aten/src/ATen/Utils.h

@@ -20,7 +20,7 @@
 
 namespace at {
 
-TORCH_API int _crash_if_asan(int);
+TORCH_API int _crash_if_asan(int /*arg*/);
 
 // Converts a TensorList (i.e. ArrayRef<Tensor> to vector of TensorImpl*)
 // NB: This is ONLY used by legacy TH bindings, and ONLY used by cat.

aten/src/ATen/Version.cpp

@@ -103,9 +103,7 @@ std::string get_cpu_capability() {
 #elif defined(HAVE_ZVECTOR_CPU_DEFINITION)
     case native::CPUCapability::ZVECTOR:
       return "Z VECTOR";
-#elif defined(HAVE_SVE_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
-    case native::CPUCapability::SVE128:
-      return "SVE128";
+#elif defined(HAVE_SVE256_CPU_DEFINITION) && defined(HAVE_ARM_BF16_CPU_DEFINITION)
     case native::CPUCapability::SVE256:
       return "SVE256";
 #else

aten/src/ATen/autocast_mode.cpp

@@ -148,7 +148,7 @@ Tensor cached_cast(at::ScalarType to_type, const Tensor& arg, DeviceType device_
 Banned functions
 *******************************/
 
-static Tensor binary_cross_entropy_banned(const Tensor &, const Tensor &, const std::optional<Tensor>&, int64_t) {
+static Tensor binary_cross_entropy_banned(const Tensor & /*unused*/, const Tensor & /*unused*/, const std::optional<Tensor>& /*unused*/, int64_t /*unused*/) {
   TORCH_CHECK(false, "torch.nn.functional.binary_cross_entropy and torch.nn.BCELoss are unsafe to autocast.\n"
            "Many models use a sigmoid layer right before the binary cross entropy layer.\n"
            "In this case, combine the two layers using torch.nn.functional.binary_cross_entropy_with_logits\n"

aten/src/ATen/core/NamedTensor.cpp

@@ -49,7 +49,7 @@ static void check_unique_names(DimnameList names) {
 }
 
 void check_names_valid_for(const TensorBase& tensor, DimnameList names) {
-  return impl::check_names_valid_for(tensor.unsafeGetTensorImpl(), names);
+  impl::check_names_valid_for(tensor.unsafeGetTensorImpl(), names);
 }
 
 void check_names_valid_for(size_t tensor_dim, DimnameList names) {

aten/src/ATen/core/NamedTensor.h

@@ -27,11 +27,11 @@ struct TORCH_API NamedTensorMeta final : public c10::NamedTensorMetaInterface {
     HasNonWildcard
   };
 
-  explicit NamedTensorMeta(HAS_NON_WILDCARD, DimnameList names)
+  explicit NamedTensorMeta(HAS_NON_WILDCARD /*unused*/, DimnameList names)
     : names_(names.vec()) {
     check_invariants();
   }
-  explicit NamedTensorMeta(HAS_NON_WILDCARD, std::vector<Dimname>&& names)
+  explicit NamedTensorMeta(HAS_NON_WILDCARD /*unused*/, std::vector<Dimname>&& names)
     : names_(std::move(names)) {
     check_invariants();
   }
@@ -52,13 +52,13 @@ struct TORCH_API NamedTensorMeta final : public c10::NamedTensorMetaInterface {
       std::any_of(names_.begin(), names_.end(), [](const Dimname& n) { return !n.isWildcard(); }));
   }
 
-  void set_names(HAS_NON_WILDCARD, DimnameList new_names) {
+  void set_names(HAS_NON_WILDCARD /*unused*/, DimnameList new_names) {
     TORCH_INTERNAL_ASSERT(new_names.size() == names_.size());
     std::copy(new_names.begin(), new_names.end(), names_.begin());
     check_invariants();
   }
 
-  void set_names(HAS_NON_WILDCARD, std::vector<Dimname>&& new_names) {
+  void set_names(HAS_NON_WILDCARD /*unused*/, std::vector<Dimname>&& new_names) {
     TORCH_INTERNAL_ASSERT(new_names.size() == names_.size());
     names_ = std::move(new_names);
     check_invariants();

aten/src/ATen/core/PythonOpRegistrationTrampoline.h

@@ -13,7 +13,7 @@ class TORCH_API PythonOpRegistrationTrampoline final {
 public:
   //  Returns true if you successfully registered yourself (that means
   //  you are in the hot seat for doing the operator registrations!)
-  static bool registerInterpreter(c10::impl::PyInterpreter*);
+  static bool registerInterpreter(c10::impl::PyInterpreter* /*interp*/);
 
   // Returns nullptr if no interpreter has been registered yet.
   static c10::impl::PyInterpreter* getInterpreter();

aten/src/ATen/core/Tensor.cpp

@@ -138,7 +138,7 @@ void Tensor::_backward(TensorList inputs,
         const std::optional<Tensor>& gradient,
         std::optional<bool> keep_graph,
         bool create_graph) const {
-  return impl::GetVariableHooks()->_backward(*this, inputs, gradient, keep_graph, create_graph);
+  impl::GetVariableHooks()->_backward(*this, inputs, gradient, keep_graph, create_graph);
 }
 
 const TensorBase& TensorBase::requires_grad_(bool _requires_grad) const {
@@ -173,4 +173,12 @@ unsigned TensorBase::_register_hook(std::function<TensorBase(const TensorBase&)>
   return impl::GetVariableHooks()->_register_hook(*this, std::move(hook));
 }
 
+std::optional<ScalarType> TensorBase::grad_dtype() const {
+  return impl::GetVariableHooks()->grad_dtype(*this);
+}
+
+void TensorBase::set_grad_dtype(const std::optional<ScalarType>& grad_dtype) const {
+  return impl::GetVariableHooks()->set_grad_dtype(*this, grad_dtype);
+}
+
 } // namespace at

aten/src/ATen/core/TensorBase.h

@@ -100,7 +100,7 @@ class TORCH_API TensorBase {
   // Create a Tensor with a +0 reference count. Special care must be
   // taken to avoid decrementing this reference count at destruction
   // time. Intended to support MaybeOwnedTraits<Tensor>.
-  explicit TensorBase(unsafe_borrow_t, const TensorBase& rhs)
+  explicit TensorBase(unsafe_borrow_t /*unused*/, const TensorBase& rhs)
       : impl_(c10::intrusive_ptr<at::TensorImpl, UndefinedTensorImpl>(rhs.impl_.get(), c10::raw::DontIncreaseRefcount{})) {}
   friend MaybeOwnedTraits<TensorBase>;
 
@@ -930,6 +930,10 @@ public:
 
   const TensorBase& requires_grad_(bool _requires_grad=true) const;
 
+  std::optional<ScalarType> grad_dtype() const;
+
+  void set_grad_dtype(const std::optional<ScalarType>& grad_dtype) const;
+
   // View Variables
   //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -950,7 +954,7 @@ protected:
   c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;
 
 private:
-  TensorBase __dispatch_contiguous(c10::MemoryFormat) const;
+  TensorBase __dispatch_contiguous(c10::MemoryFormat /*memory_format*/) const;
 };
 
 inline DeviceIndex get_device(const TensorBase& self) {

aten/src/ATen/core/VariableHooksInterface.h

@@ -68,6 +68,8 @@ struct TORCH_API VariableHooksInterface {
       const c10::OperatorHandle& op,
       c10::DispatchKeySet dispatch_keys,
       torch::jit::Stack* stack) const = 0;
+  virtual std::optional<c10::ScalarType> grad_dtype(const TensorBase&) const = 0;
+  virtual void set_grad_dtype(const TensorBase&, const std::optional<c10::ScalarType>&) const = 0;
 };
 
 TORCH_API void SetVariableHooks(VariableHooksInterface* hooks);

aten/src/ATen/core/boxing/BoxedKernel.h

@@ -18,10 +18,10 @@ class KernelFunction;
 // implementation notes; notably, this does NOT actually go through the
 // boxing/unboxing codepath.
 TORCH_API void fallthrough_kernel(
-    OperatorKernel*,
-    const OperatorHandle&,
-    DispatchKeySet,
-    Stack*);
+    OperatorKernel* /*unused*/,
+    const OperatorHandle& /*unused*/,
+    DispatchKeySet /*unused*/,
+    Stack* /*unused*/);
 
 // Note [Ambiguity in AutogradOther kernel]
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -62,10 +62,10 @@ TORCH_API void fallthrough_kernel(
 // than arbitrarily pick one or the other, we just register a kernel that raises
 // an error and let the user decide how to proceed.
 TORCH_API void ambiguous_autogradother_kernel(
-    OperatorKernel*,
-    const OperatorHandle&,
-    DispatchKeySet,
-    Stack*);
+    OperatorKernel* /*unused*/,
+    const OperatorHandle& /*op*/,
+    DispatchKeySet /*unused*/,
+    Stack* /*unused*/);
 
 // Note [named_not_supported_kernel]
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -75,10 +75,10 @@ TORCH_API void ambiguous_autogradother_kernel(
 // give a good error message in cases when boxing is not supported).  When
 // boxing is universally supported this can be removed.
 [[noreturn]] TORCH_API void named_not_supported_kernel(
-    OperatorKernel*,
-    const OperatorHandle&,
-    DispatchKeySet,
-    Stack*);
+    OperatorKernel* /*unused*/,
+    const OperatorHandle& /*op*/,
+    DispatchKeySet /*unused*/,
+    Stack* /*unused*/);
 
 /**
  * BoxedKernel is similar to a std::function storing a boxed kernel.
@@ -185,16 +185,16 @@ class TORCH_API BoxedKernel final {
 
   template <BoxedKernelFunction* func>
   static void make_boxed_function(
-      OperatorKernel*,
+      OperatorKernel* /*unused*/,
       const OperatorHandle& opHandle,
-      DispatchKeySet,
+      DispatchKeySet /*unused*/,
       Stack* stack);
 
   template <BoxedKernelFunction_withDispatchKeys* func>
   static void make_boxed_function(
-      OperatorKernel*,
+      OperatorKernel* /*unused*/,
       const OperatorHandle& opHandle,
-      DispatchKeySet,
+      DispatchKeySet /*ks*/,
       Stack* stack);
 
   explicit BoxedKernel(

aten/src/ATen/core/boxing/BoxedKernel_impl.h

@@ -11,9 +11,9 @@ inline BoxedKernel::BoxedKernel(
 
 template <BoxedKernel::BoxedKernelFunction* func>
 inline void BoxedKernel::make_boxed_function(
-    OperatorKernel*,
+    OperatorKernel* /*unused*/,
     const OperatorHandle& opHandle,
-    DispatchKeySet,
+    DispatchKeySet /*unused*/,
     Stack* stack) {
   // Note that we're dropping the DispatchKeySet argument.
   // See Note [Plumbing Keys Through The Dispatcher 2] for details.
@@ -22,7 +22,7 @@ inline void BoxedKernel::make_boxed_function(
 
 template <BoxedKernel::BoxedKernelFunction_withDispatchKeys* func>
 inline void BoxedKernel::make_boxed_function(
-    OperatorKernel*,
+    OperatorKernel* /*unused*/,
     const OperatorHandle& opHandle,
     DispatchKeySet ks,
     Stack* stack) {

aten/src/ATen/core/boxing/KernelFunction.cpp

@@ -10,7 +10,7 @@ namespace c10 {
 // be handled specially.  Its semantics is that it redispatches to the
 // *next* dispatch key that would have been processed, skipping the current
 // one.
-void fallthrough_kernel(OperatorKernel*, const OperatorHandle&, DispatchKeySet, Stack*) {
+void fallthrough_kernel(OperatorKernel* /*unused*/, const OperatorHandle& /*unused*/, DispatchKeySet /*unused*/, Stack* /*unused*/) {
   TORCH_INTERNAL_ASSERT(0,
     "fallthrough_kernel was executed but it should have been short-circuited by the dispatcher. "
     "This could occur if you registered a fallthrough kernel as a override for a specific operator "
@@ -19,7 +19,7 @@ void fallthrough_kernel(OperatorKernel*, const OperatorHandle&, DispatchKeySet,
     "let us know in the bug tracker.");
 }
 
-void ambiguous_autogradother_kernel(OperatorKernel*, const OperatorHandle& op, DispatchKeySet, Stack*) {
+void ambiguous_autogradother_kernel(OperatorKernel* /*unused*/, const OperatorHandle& op, DispatchKeySet /*unused*/, Stack* /*unused*/) {
   TORCH_INTERNAL_ASSERT(0,
     op.operator_name(), " has kernels registered to both CompositeImplicitAutograd and a backend mapped to AutogradOther. "
     "This makes the backend kernel unreachable; the dispatcher will always prefer the CompositeImplicitAutograd lowering "
@@ -32,7 +32,7 @@ void ambiguous_autogradother_kernel(OperatorKernel*, const OperatorHandle& op, D
     "\nCanonical state\n~~~~~~~~~~~\n", op.dumpState(), "\n\n");
 }
 
-void named_not_supported_kernel(OperatorKernel*, const OperatorHandle& op, DispatchKeySet, Stack*) {
+void named_not_supported_kernel(OperatorKernel* /*unused*/, const OperatorHandle& op, DispatchKeySet /*unused*/, Stack* /*unused*/) {
   // DO NOT LOOK AT STACK, YOU HAVE SHORT CIRCUITED BOXING
   // See Note [named_not_supported_kernel]
   TORCH_CHECK(0,

aten/src/ATen/core/boxing/KernelFunction.h

@@ -229,7 +229,7 @@ class TORCH_API KernelFunction final {
    * &unboxed_func>();
    */
   template <class FuncPtr, bool AllowLegacyTypes = false>
-  static KernelFunction makeFromUnboxedFunction(FuncPtr);
+  static KernelFunction makeFromUnboxedFunction(FuncPtr /*func_ptr*/);
 
   /**
    * Create a KernelFunction from an unboxed function.
@@ -271,7 +271,7 @@ class TORCH_API KernelFunction final {
 
   std::string dumpState() const;
   // For testing internal invariants only
-  bool _equalsBoxedAndUnboxed(const KernelFunction&) const;
+  bool _equalsBoxedAndUnboxed(const KernelFunction& /*other*/) const;
 
   // Register a token to be invalidated when this KernelFunction is destroyed
   void registerToken(std::weak_ptr<KernelToken> token) const;

aten/src/ATen/core/boxing/impl/boxing.h

@@ -131,7 +131,7 @@ C10_ALWAYS_INLINE_UNLESS_MOBILE void boxToStack(
   new (dest++) IValue(options.pinned_memory());
 }
 
-inline void boxArgsToStack(IValue*&) {}
+inline void boxArgsToStack(IValue*& /*unused*/) {}
 
 template <typename T, typename... Args>
 C10_ALWAYS_INLINE_UNLESS_MOBILE void boxArgsToStack(
@@ -185,7 +185,7 @@ struct PopResult<std::tuple<Types...>> final {
   template <size_t... indices>
   static Result pop_to_tuple_impl(
       Stack& stack,
-      std::index_sequence<indices...>) {
+      std::index_sequence<indices...> /*unused*/) {
     return std::make_tuple((std::move(stack[indices]).template to<Types>())...);
   }
 };

aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h

@@ -561,7 +561,7 @@ struct wrap_kernel_functor_unboxed_<
   // doesn't use &&
   static ReturnType call(
       OperatorKernel* functor,
-      DispatchKeySet,
+      DispatchKeySet /*unused*/,
       ParameterTypes... args) {
     KernelFunctor* functor_ = static_cast<KernelFunctor*>(functor);
     // Note [Plumbing Keys Through The Dispatcher 2]
@@ -629,8 +629,8 @@ call_functor_with_args_from_stack_(
     OperatorKernel* functor,
     DispatchKeySet dispatchKeySet,
     Stack* stack,
-    std::index_sequence<ivalue_arg_indices...>,
-    guts::typelist::typelist<ArgTypes...>*) {
+    std::index_sequence<ivalue_arg_indices...> /*unused*/,
+    guts::typelist::typelist<ArgTypes...>* /*unused*/) {
   (void)(stack); // when sizeof...(ivalue_arg_indices) == 0, this argument would
                  // be unused and we have to silence the compiler warning.
 
@@ -708,7 +708,7 @@ struct push_outputs<std::tuple<OutputTypes...>, AllowDeprecatedTypes> final {
   static void call_(
       std::tuple<OutputTypes...>&& output,
       Stack* stack,
-      std::index_sequence<indices...>) {
+      std::index_sequence<indices...> /*unused*/) {
     torch::jit::push(
         *stack,
         return_to_ivalue<OutputTypes, AllowDeprecatedTypes>::call(
@@ -718,7 +718,7 @@ struct push_outputs<std::tuple<OutputTypes...>, AllowDeprecatedTypes> final {
   static void copy_(
       const std::tuple<OutputTypes...>& output,
       Stack* stack,
-      std::index_sequence<indices...>) {
+      std::index_sequence<indices...> /*unused*/) {
     torch::jit::push(
         *stack,
         return_to_ivalue<OutputTypes, AllowDeprecatedTypes>::copy(
@@ -741,7 +741,7 @@ struct make_boxed_from_unboxed_functor final {
 
   static void call(
       OperatorKernel* functor,
-      const OperatorHandle&,
+      const OperatorHandle& /*unused*/,
       DispatchKeySet dispatchKeySet,
       Stack* stack) {
     using ReturnType =

aten/src/ATen/core/builtin_function.h

@@ -63,13 +63,13 @@ struct BuiltinOpFunction : public Function {
 
   bool call(
       Stack& stack,
-      std::optional<size_t>,
-      c10::function_ref<void(const Code&)>) override {
+      std::optional<size_t> /*unused*/,
+      c10::function_ref<void(const Code&)> /*unused*/) override {
     run(stack);
     return false;
   }
 
-  bool call(Stack& stack, c10::function_ref<void(const mobile::Code&)>)
+  bool call(Stack& stack, c10::function_ref<void(const mobile::Code&)> /*unused*/)
       override {
     run(stack);
     return false;

aten/src/ATen/core/dispatch/DispatchKeyExtractor.h

@@ -80,7 +80,8 @@ struct MultiDispatchKeySet : at::IterArgs<MultiDispatchKeySet> {
       ts = ts | x.key_set();
     }
   }
-  [[noreturn]] void operator()(at::ArrayRef<std::optional<at::Tensor>>) {
+  [[noreturn]] void operator()(
+      at::ArrayRef<std::optional<at::Tensor>> /*unused*/) {
     // Just checking that the handling of Tensor?[] didn't change.
     TORCH_INTERNAL_ASSERT(false);
   }
@@ -95,7 +96,7 @@ struct MultiDispatchKeySet : at::IterArgs<MultiDispatchKeySet> {
     }
   }
   template <typename T>
-  void operator()(const T&) {
+  void operator()(const T& /*unused*/) {
     // do nothing
   }
 };

aten/src/ATen/core/dispatch/Dispatcher.h

@@ -496,7 +496,7 @@ class TORCH_API OperatorHandle {
   }
 
   void checkInvariants() const {
-    return operatorDef_->op.checkInvariants();
+    operatorDef_->op.checkInvariants();
   }
 
   c10::ArrayRef<at::Tag> getTags() const {
@@ -633,7 +633,7 @@ class TypedOperatorHandle<Return(Args...)> final : public OperatorHandle {
 
 namespace detail {
 template <class... Args>
-inline void unused_arg_(const Args&...) {}
+inline void unused_arg_(const Args&... /*unused*/) {}
 
 // CaptureKernelCall is intended to capture return values from Dispatcher
 // unboxed kernel calls. A record function may request to get outputs from the
@@ -932,7 +932,7 @@ inline void Dispatcher::redispatchBoxed(
   }
 #endif
   const auto& kernel = entry.lookup(dispatchKeySet);
-  return kernel.callBoxed(op, dispatchKeySet, stack);
+  kernel.callBoxed(op, dispatchKeySet, stack);
 }
 
 } // namespace c10

aten/src/ATen/core/dispatch/OperatorEntry.h

@@ -105,7 +105,7 @@ class TORCH_API OperatorEntry final {
   // versa that is an error.  (Refcounting for the registrations is
   // handled in the OperatorHandle in Dispatcher)
   void registerSchema(
-      FunctionSchema&&,
+      FunctionSchema&& /*schema*/,
       std::string&& debug,
       std::vector<at::Tag> tags = {});
   void deregisterSchema();

aten/src/ATen/core/dynamic_type.cpp

@@ -177,7 +177,7 @@ bool DynamicType::equals(const Type& rhs) const {
   return equals(*create(rhs));
 }
 
-bool DynamicType::isSubtypeOfExt(const Type& rhs, std::ostream*) const {
+bool DynamicType::isSubtypeOfExt(const Type& rhs, std::ostream* /*why_not*/) const {
   auto other = create(rhs);
   if (tag_ == other->tag_) {
     if (equals(*other)) {
@@ -371,7 +371,7 @@ DynamicTypePtr ivalue::TupleTypeFactory<c10::DynamicType>::create(
 }
 
 DynamicTypePtr ivalue::TupleTypeFactory<c10::DynamicType>::fallback(
-    const Type&) {
+    const Type& /*unused*/) {
   TORCH_INTERNAL_ASSERT_DEBUG_ONLY(false);
   return nullptr;
 }

aten/src/ATen/core/dynamic_type.h

@@ -138,8 +138,8 @@ class DynamicType : public SharedType {
 
   struct Arguments {
     Arguments() = default;
-    Arguments(c10::ArrayRef<TypePtr>);
-    Arguments(const std::vector<std::string_view>&, c10::ArrayRef<TypePtr>);
+    Arguments(c10::ArrayRef<TypePtr> /*args*/);
+    Arguments(const std::vector<std::string_view>& /*names*/, c10::ArrayRef<TypePtr> /*args*/);
     std::vector<LabeledDynamicType> elems;
   };
 
@@ -156,15 +156,15 @@ class DynamicType : public SharedType {
   static const TypeKind Kind = TypeKind::DynamicType;
   static TORCH_API DynamicTypePtr create(Type& ty);
 
-  explicit DynamicType(Tag, Arguments);
-  explicit DynamicType(Tag, std::string_view, Arguments);
+  explicit DynamicType(Tag /*tag*/, Arguments /*arguments*/);
+  explicit DynamicType(Tag /*tag*/, std::string_view /*name*/, Arguments /*arguments*/);
 
   DynamicType(DynamicType&& other) = delete;
   DynamicType(const DynamicType&) = delete;
   DynamicType& operator=(const DynamicType&) = delete;
   DynamicType& operator=(DynamicType&&) = delete;
 
-  TypePtr containedType(size_t) const override;
+  TypePtr containedType(size_t /*i*/) const override;
   size_t containedTypeSize() const override;
   Tag tag() const {
     return tag_;

aten/src/ATen/core/function.h

@@ -96,15 +96,15 @@ struct TORCH_API Function {
   // Overload for server interpreter, a bailout size is needed for graph
   // executor.
   virtual bool call(
-      Stack&,
-      std::optional<size_t>,
-      c10::function_ref<void(const Code&)>) {
+      Stack& /*unused*/,
+      std::optional<size_t> /*unused*/,
+      c10::function_ref<void(const Code&)> /*unused*/) {
     TORCH_INTERNAL_ASSERT_DEBUG_ONLY(false);
     return false;
   }
 
   // Overload for mobile interpreter.
-  virtual bool call(Stack&, c10::function_ref<void(const mobile::Code&)>) {
+  virtual bool call(Stack& /*unused*/, c10::function_ref<void(const mobile::Code&)> /*unused*/) {
     TORCH_INTERNAL_ASSERT_DEBUG_ONLY(false);
     return false;
   }

aten/src/ATen/core/ivalue.h

@@ -847,7 +847,7 @@ struct TORCH_API IValue final {
   IValue(std::optional<T> v);
   template <class T, enable_if_list_is_ivalue_constructible<T> = nullptr>
   IValue(c10::OptionalArrayRef<T> v);
-  IValue(std::nullopt_t);
+  IValue(std::nullopt_t /*unused*/);
 
   // ClassType
   IValue(c10::intrusive_ptr<ivalue::Object> v);

aten/src/ATen/core/ivalue_inl.h

@@ -660,7 +660,7 @@ struct TORCH_API TupleTypeFactory<TupleType> {
 template <>
 struct TORCH_API TupleTypeFactory<c10::DynamicType> {
   static DynamicTypePtr create(const std::vector<TypePtr>& elemTypes);
-  static DynamicTypePtr fallback(const Type&);
+  static DynamicTypePtr fallback(const Type& /*unused*/);
 };
 
 struct TORCH_API Tuple : c10::intrusive_ptr_target {
@@ -1682,7 +1682,7 @@ struct ivalue::EnumHolder : c10::intrusive_ptr_target {
 namespace detail {
 
 struct _guarded_unsigned_long_unique_dummy final {
-  _guarded_unsigned_long_unique_dummy(int64_t){}
+  _guarded_unsigned_long_unique_dummy(int64_t /*unused*/){}
 };
 using _guarded_unsigned_long = std::conditional_t<
     std::is_same_v<unsigned long, uint32_t> ||
@@ -1776,7 +1776,7 @@ template <class Elem>
 // native_functions.yaml still return std::vector.
 // C10_DEPRECATED_MESSAGE("IValues based on std::vector<T> are potentially slow
 // and deprecated. Please use torch::List<T> instead.")
-std::vector<Elem> generic_to(IValue ivalue, _fake_type<std::vector<Elem>>) {
+std::vector<Elem> generic_to(IValue ivalue, _fake_type<std::vector<Elem>> /*unused*/) {
   // We need to do a deep copy of the vector because there might be other
   // references to this same IValue that also use the list. We can't just
   // move the elements out.
@@ -1826,18 +1826,18 @@ c10::intrusive_ptr<T> IValue::toCustomClass() const& {
 }
 
 template <typename T>
-T generic_to(IValue ivalue, _fake_type<T>) {
+T generic_to(IValue ivalue, _fake_type<T> /*unused*/) {
   using ElemType = typename std::remove_pointer<T>::type::element_type;
   return std::move(ivalue).template toCustomClass<ElemType>();
 }
 
 template <typename T>
-tagged_capsule<T> generic_to(IValue ivalue, _fake_type<tagged_capsule<T>>) {
+tagged_capsule<T> generic_to(IValue ivalue, _fake_type<tagged_capsule<T>> /*unused*/) {
   return tagged_capsule<T>{std::move(ivalue)};
 }
 
 template <typename Elem>
-c10::List<Elem> generic_to(IValue ivalue, _fake_type<c10::List<Elem>>) {
+c10::List<Elem> generic_to(IValue ivalue, _fake_type<c10::List<Elem>> /*unused*/) {
   return impl::toTypedList<Elem>(std::move(ivalue).toList());
 }
 
@@ -1867,7 +1867,7 @@ std::vector<T> createVectorFromList(const c10::List<T>& impl) {
 }
 
 template <typename T>
-OptionalArray<T> generic_to(IValue ivalue, _fake_type<OptionalArray<T>>) {
+OptionalArray<T> generic_to(IValue ivalue, _fake_type<OptionalArray<T>> /*unused*/) {
   if (ivalue.isNone()) {
     return {};
   }
@@ -1880,8 +1880,8 @@ namespace detail {
 template <typename Elem, size_t... I>
 std::array<Elem, sizeof...(I)> generic_to_array(
     IValue ivalue,
-    _fake_type<std::array<Elem, sizeof...(I)>>,
-    std::index_sequence<I...>) {
+    _fake_type<std::array<Elem, sizeof...(I)>> /*unused*/,
+    std::index_sequence<I...> /*unused*/) {
   // We need to do a deep copy of the array because there might be other
   // references to this same IValue that also use the list. We can't just
   // move the elements out.
@@ -1906,7 +1906,7 @@ std::array<Elem, N> generic_to(
 template <typename Key, typename Value>
 c10::Dict<Key, Value> generic_to(
     IValue ivalue,
-    _fake_type<c10::Dict<Key, Value>>) {
+    _fake_type<c10::Dict<Key, Value>> /*unused*/) {
   return impl::toTypedDict<Key, Value>(std::move(ivalue).toGenericDict());
 }
 
@@ -1915,7 +1915,7 @@ C10_DEPRECATED_MESSAGE(
     "IValues based on std::unordered_map are slow and deprecated. Please use c10::Dict<K, V> instead.")
 std::unordered_map<K, V> generic_to(
     IValue ivalue,
-    _fake_type<std::unordered_map<K, V>>) {
+    _fake_type<std::unordered_map<K, V>> /*unused*/) {
   std::unordered_map<K, V> specialized_dict;
 
   for (const auto& item : std::move(ivalue).toGenericDict()) {
@@ -1926,7 +1926,7 @@ std::unordered_map<K, V> generic_to(
 }
 
 template <typename T>
-std::optional<T> generic_to(IValue ivalue, _fake_type<std::optional<T>>) {
+std::optional<T> generic_to(IValue ivalue, _fake_type<std::optional<T>> /*unused*/) {
   if (ivalue.isNone()) {
     return std::nullopt;
   }
@@ -1937,7 +1937,7 @@ namespace detail {
 template <typename Tuple, std::size_t... INDEX>
 Tuple generic_to_tuple_impl(
     const ivalue::TupleElements& t,
-    std::index_sequence<INDEX...>) {
+    std::index_sequence<INDEX...> /*unused*/) {
   return std::make_tuple(
       t[INDEX].to<typename std::tuple_element<INDEX, Tuple>::type>()...);
 }
@@ -1951,7 +1951,7 @@ template <
             std::is_lvalue_reference<Args>...,
             std::negation<std::is_constructible<IValue, Args>>...>,
         std::nullptr_t> = nullptr>
-std::tuple<Args...> generic_to(const IValue& ivalue, _fake_type<std::tuple<Args...>>) {
+std::tuple<Args...> generic_to(const IValue& ivalue, _fake_type<std::tuple<Args...>> /*unused*/) {
   const auto& vals = ivalue.toTupleRef().elements();
   TORCH_CHECK(vals.size() == sizeof...(Args));
   return detail::generic_to_tuple_impl<std::tuple<Args...>>(vals, Indices{});
@@ -2311,7 +2311,7 @@ inline IValue::IValue(std::optional<T> v) : IValue() {
   }
 }
 
-inline IValue::IValue(std::nullopt_t) : IValue() {}
+inline IValue::IValue(std::nullopt_t /*unused*/) : IValue() {}
 
 inline IValue::IValue(c10::intrusive_ptr<ivalue::Object> v)
     : tag(Tag::Object) {
@@ -2482,15 +2482,15 @@ namespace ivalue {
 namespace detail {
 
 template <typename T>
-IValue from_(T&& x, std::true_type) {
+IValue from_(T&& x, std::true_type /*unused*/) {
   return IValue(std::forward<T>(x));
 }
 template <typename T>
-IValue from_(c10::intrusive_ptr<T> x, std::false_type) {
+IValue from_(c10::intrusive_ptr<T> x, std::false_type /*unused*/) {
   return IValue(std::move(x));
 }
 template <typename T>
-IValue from_(T&& /*x*/, std::false_type) {
+IValue from_(T&& /*x*/, std::false_type /*unused*/) {
   static_assert(
       guts::false_t<T>::value,
       "You are calling from with a type that it doesn't support, and isn't a potential custom class (ie: is an intrusive_ptr)");
@@ -2546,19 +2546,19 @@ struct MaybeOwnedTraits<IValue> {
     return &borrow;
   }
 
-  static bool debugBorrowIsValid(const borrow_type&) {
+  static bool debugBorrowIsValid(const borrow_type& /*unused*/) {
     return true;
   }
 };
 
 template <>
 struct IValue::TagType<c10::Type> {
-  static TORCH_API c10::TypePtr get(const IValue&);
+  static TORCH_API c10::TypePtr get(const IValue& /*v*/);
 };
 
 template <>
 struct IValue::TagType<c10::DynamicType> {
-  static TORCH_API c10::TypePtr get(const IValue&);
+  static TORCH_API c10::TypePtr get(const IValue& /*v*/);
 };
 
 template <typename T>

aten/src/ATen/core/op_registration/infer_schema.h

@@ -44,7 +44,7 @@ constexpr int checkStaticTypes() {
 }
 
 template <typename... Ts, size_t... Is>
-constexpr std::array<ArgumentDef, sizeof...(Ts)> createArgumentVectorFromTypes(std::index_sequence<Is...>) {
+constexpr std::array<ArgumentDef, sizeof...(Ts)> createArgumentVectorFromTypes(std::index_sequence<Is...> /*unused*/) {
   return (
     // Check types for common errors
     checkStaticTypes<Ts...>(),

aten/src/ATen/core/operator_name.h

@@ -83,7 +83,7 @@ inline bool operator!=(const OperatorName& lhs, const OperatorName& rhs) {
 }
 
 TORCH_API std::string toString(const OperatorName& opName);
-TORCH_API std::ostream& operator<<(std::ostream&, const OperatorName&);
+TORCH_API std::ostream& operator<<(std::ostream& /*os*/, const OperatorName& /*opName*/);
 
 } // namespace c10
 

aten/src/ATen/core/type_ptr.h

@@ -16,7 +16,7 @@ class SingletonTypePtr {
   /* implicit */ SingletonTypePtr(T* p) : repr_(p) {}
 
   // We need this to satisfy Pybind11, but it shouldn't be hit.
-  explicit SingletonTypePtr(std::shared_ptr<T>) { TORCH_CHECK(false); }
+  explicit SingletonTypePtr(std::shared_ptr<T> /*unused*/) { TORCH_CHECK(false); }
 
   using element_type = typename std::shared_ptr<T>::element_type;
 

aten/src/ATen/cpu/vec/functional_base.h

@@ -102,31 +102,8 @@ struct VecReduceAllSIMD<float, Op> {
 #endif // defined(__GNUC__) && (__GNUC__ > 5) && !defined(_MSC_VER) &&
        // !defined(C10_MOBILE)
 
-#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
-#if defined(CPU_CAPABILITY_SVE256)
-template <typename Op>
-struct VecReduceAllSIMD<float, Op> {
-  static inline float apply(
-      const Op& vec_fun,
-      const Vectorized<float>& acc_vec) {
-    using Vec = Vectorized<float>;
-    Vec v = acc_vec;
-    // 128-bit shuffle
-    svuint32_t ind = svdupq_n_u32(4, 5, 6, 7);
-    Vec v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    // 64-bit shuffle
-    ind = svdupq_n_u32(2, 3, 0, 1);
-    v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    // 32-bit shuffle
-    ind = svdupq_n_u32(1, 0, 2, 3);
-    v1 = svtbl_f32(v, ind);
-    v = vec_fun(v, v1);
-    return svlasta(svpfalse(), v);
-  }
-};
-#else
+#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
+    !defined(CPU_CAPABILITY_SVE)
 template <typename Op>
 struct VecReduceAllSIMD<float, Op> {
   static inline float apply(
@@ -163,8 +140,35 @@ struct VecReduceAllSIMD<float, std::plus<Vectorized<float>>> {
     return vaddvq_f32(acc_vec);
   }
 };
-#endif // defined(CPU_CAPABILITY_SVE256)
 #endif // defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
+       // && !defined(CPU_CAPABILITY_SVE)
+
+#if defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
+    defined(CPU_CAPABILITY_SVE256)
+template <typename Op>
+struct VecReduceAllSIMD<float, Op> {
+  static inline float apply(
+      const Op& vec_fun,
+      const Vectorized<float>& acc_vec) {
+    using Vec = Vectorized<float>;
+    Vec v = acc_vec;
+    // 128-bit shuffle
+    svuint32_t ind = svdupq_n_u32(4, 5, 6, 7);
+    Vec v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    // 64-bit shuffle
+    ind = svdupq_n_u32(2, 3, 0, 1);
+    v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    // 32-bit shuffle
+    ind = svdupq_n_u32(1, 0, 2, 3);
+    v1 = svtbl_f32(v, ind);
+    v = vec_fun(v, v1);
+    return svlasta(svpfalse(), v);
+  }
+};
+#endif // defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__)
+       // && defined(CPU_CAPABILITY_SVE256)
 
 template <typename scalar_t, typename Op>
 inline scalar_t vec_reduce_all(

aten/src/ATen/cpu/vec/sve/sve_helper.h

@@ -1,21 +1,9 @@
 #pragma once
 
 #include <ATen/cpu/vec/intrinsics.h>
-#include <c10/macros/Macros.h>
-#include <cstdint>
 
 #include <ATen/cpu/vec/vec_base.h>
 
-#if defined(__aarch64__) &&                     \
-    (defined(AT_BUILD_ARM_VEC256_WITH_SLEEF) || \
-     defined(AT_BUILD_ARM_VECSVE_WITH_SLEEF))
-#define SLEEF_STATIC_LIBS
-#include <sleef.h>
-#define USE_SLEEF(sleef_code, non_sleef_code) sleef_code
-#else
-#define USE_SLEEF(sleef_code, non_sleef_code) non_sleef_code
-#endif
-
 #if defined(CPU_CAPABILITY_SVE)
 
 // Define the data type of VLS(vector-length specific).

aten/src/ATen/cpu/vec/sve/vec_bfloat16.h

@@ -2,6 +2,7 @@
 
 #include <ATen/cpu/vec/intrinsics.h>
 #include <ATen/cpu/vec/sve/sve_helper.h>
+#include <ATen/cpu/vec/sve/vec_common_sve.h>
 #include <ATen/cpu/vec/sve/vec_float.h>
 #include <ATen/cpu/vec/vec_base.h>
 #include <c10/util/bit_cast.h>
@@ -307,8 +308,8 @@ Vectorized<c10::BFloat16> inline operator/(
 }
 
 inline Vectorized<BFloat16>::Vectorized() {
-  const short zero = 0;
-  values = svdup_n_bf16(c10::bit_cast<bfloat16_t>(zero));
+  auto vals_f = svdup_n_f32(0);
+  values = convert_float_bfloat16(vals_f, vals_f);
 }
 
 inline Vectorized<BFloat16>::Vectorized(int val) {

aten/src/ATen/cpu/vec/sve/vec_common_sve.h

@@ -8,48 +8,13 @@
 #include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec_base.h>
 
-#ifdef CPU_CAPABILITY_SVE128
-
-#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_half_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_convert.h>
-
-#include <ATen/cpu/vec/sve/vec_qint.h>
-
-#elif defined(CPU_CAPABILITY_SVE)
-
-#include <ATen/cpu/vec/sve/vec_float.h>
-
+#if defined(CPU_CAPABILITY_SVE)
 #include <ATen/cpu/vec/sve/vec_bfloat16.h>
-
 #include <ATen/cpu/vec/sve/vec_double.h>
+#include <ATen/cpu/vec/sve/vec_float.h>
 #include <ATen/cpu/vec/sve/vec_int.h>
-
 #include <ATen/cpu/vec/sve/vec_qint.h>
-
-#include <ATen/cpu/vec/vec256/vec256_half.h>
-
-#include <ATen/cpu/vec/vec256/vec256_convert.h>
-
-#else // NEON
-
-#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_half_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
-
-#include <ATen/cpu/vec/vec128/vec128_convert.h>
-
-#include <ATen/cpu/vec/vec256/vec256_qint.h>
-
-#endif // defined(CPU_CAPABILITY_SVE128)
-
-#include <ATen/cpu/vec/functional.h>
+#endif
 
 namespace at::vec {
 // Note [CPU_CAPABILITY namespace]
@@ -83,6 +48,12 @@ DEFINE_SVE_CAST(int32_t, s32, float, f32)
 DEFINE_SVE_CAST(int16_t, s16, float, f32)
 DEFINE_SVE_CAST(float, f32, double, f64)
 
+#ifdef __ARM_FEATURE_BF16
+DEFINE_SVE_CAST(int64_t, s64, c10::BFloat16, bf16)
+DEFINE_SVE_CAST(int32_t, s32, c10::BFloat16, bf16)
+DEFINE_SVE_CAST(int16_t, s16, c10::BFloat16, bf16)
+#endif // __ARM_FEATURE_BF16
+
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ GATHER ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 template <int64_t scale = 1>
@@ -202,11 +173,9 @@ std::pair<
   // group cols crossing lanes:
   //   return {a0, b0, a1, b1, a2, b2, a3, b3}
   //          {a4, b4, a5, b5, a6, b6, a7, b7}
-  svbfloat16_t aReg = a;
-  svbfloat16_t bReg = b;
-  Vectorized<c10::BFloat16> c = svzip1_bf16(aReg, bReg);
-  Vectorized<c10::BFloat16> d = svzip2_bf16(aReg, bReg);
-  return std::make_pair(c, d);
+  return std::make_pair(
+      Vectorized<c10::BFloat16>(svzip1_bf16(a, b)),
+      Vectorized<c10::BFloat16>(svzip2_bf16(a, b)));
 }
 #endif // __ARM_FEATURE_BF16
 
@@ -255,27 +224,12 @@ std::pair<
   // swap lanes:
   //   return {a0, a1, a2, a3, a4, a5, a6, a7}
   //          {b0, b1, b2, b3, b4, b5, b6, b7}
-  svbfloat16_t aReg = a;
-  svbfloat16_t bReg = b;
-  Vectorized<c10::BFloat16> c = svuzp1_bf16(aReg, bReg);
-  Vectorized<c10::BFloat16> d = svuzp2_bf16(aReg, bReg);
-  return std::make_pair(c, d);
+  return std::make_pair(
+      Vectorized<c10::BFloat16>(svuzp1_bf16((svbfloat16_t)a, (svbfloat16_t)b)),
+      Vectorized<c10::BFloat16>(svuzp2_bf16((svbfloat16_t)a, (svbfloat16_t)b)));
 }
 #endif // __ARM_FEATURE_BF16
 
-// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ FLIP ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-#define DEFINE_FLIP_FUNC(type, sve_func)                    \
-  inline Vectorized<type> flip(const Vectorized<type>& v) { \
-    return Vectorized<type>(sve_func(v));                   \
-  }
-// Use the macro to define the flip functions
-DEFINE_FLIP_FUNC(float, svrev_f32)
-DEFINE_FLIP_FUNC(double, svrev_f64)
-DEFINE_FLIP_FUNC(int64_t, svrev_s64)
-DEFINE_FLIP_FUNC(int32_t, svrev_s32)
-DEFINE_FLIP_FUNC(int16_t, svrev_s16)
-DEFINE_FLIP_FUNC(int8_t, svrev_s8)
-
 #endif // defined(CPU_CAPABILITY_SVE)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec.h

@@ -1,8 +1,6 @@
 #pragma once
 
-#if defined(__aarch64__)
-#include <ATen/cpu/vec/vec_common_aarch64.h>
-#elif defined(CPU_CAPABILITY_AVX512)
+#if defined(CPU_CAPABILITY_AVX512)
 #include <ATen/cpu/vec/vec512/vec512.h>
 #else
 #include <ATen/cpu/vec/vec128/vec128.h>
@@ -13,34 +11,6 @@ namespace at::vec {
 // See Note [CPU_CAPABILITY namespace]
 inline namespace CPU_CAPABILITY {
 
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
-  stream << val.val_;
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
-  stream << static_cast<int>(val.val_);
-  return stream;
-}
-inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
-  stream << static_cast<unsigned int>(val.val_);
-  return stream;
-}
-
-template <typename T>
-std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
-  T buf[Vectorized<T>::size()];
-  vec.store(buf);
-  stream << "vec[";
-  for (int i = 0; i != Vectorized<T>::size(); i++) {
-    if (i != 0) {
-      stream << ", ";
-    }
-    stream << buf[i];
-  }
-  stream << "]";
-  return stream;
-}
-
 inline Vectorized<bool> convert_to_bool(Vectorized<int8_t> x) {
   __at_align__ bool buffer[x.size()];
   x.ne(Vectorized<int8_t>(0)).store(buffer);

aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h

@@ -2,7 +2,6 @@
 
 // DO NOT DEFINE STATIC DATA IN THIS HEADER!
 // See Note [Do not compile initializers with AVX]
-#include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_reduced_precision_common_neon.h>
 #include <ATen/cpu/vec/vec_base.h>
@@ -263,13 +262,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
             c10::bit_cast<at_bfloat16_t>(val6.x),
             c10::bit_cast<at_bfloat16_t>(val7.x)}) {}
 
-#ifdef CPU_CAPABILITY_SVE128
-  Vectorized(svbfloat16_t v) : Vectorized16(svget_neonq(v)) {}
-  operator svbfloat16_t() const {
-    return svset_neonq(svundef_bf16(), values);
-  }
-#endif
-
   static Vectorized<c10::BFloat16> blendv(
       const Vectorized<c10::BFloat16>& a,
       const Vectorized<c10::BFloat16>& b,
@@ -382,23 +374,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
   Vectorized ge(const Vectorized& other) const;
   Vectorized lt(const Vectorized& other) const;
   Vectorized le(const Vectorized& other) const;
-
-#ifdef CPU_CAPABILITY_SVE128
-
-  template <typename step_t>
-  static Vectorized<BFloat16> arange(
-      BFloat16 base = 0.f,
-      step_t step = static_cast<step_t>(1)) {
-    __at_align__ BFloat16 buffer[size()];
-    for (int64_t i = 0; i < size(); i++) {
-      buffer[i] = base + i * step;
-    }
-    return svget_neonq(
-        svld1_bf16(ptrue, reinterpret_cast<bfloat16_t*>(buffer)));
-  }
-
-#endif // CPU_CAPABILITY_SVE128
-
 }; // Vectorized<c10::BFloat16>
 
 inline std::tuple<Vectorized<float>, Vectorized<float>> convert_bfloat16_float(
@@ -422,24 +397,6 @@ inline Vectorized<c10::BFloat16> convert_float_bfloat16(
   return Vectorized<c10::BFloat16>(at_vcombine_bf16(x1, x2));
 }
 
-inline void load_fp32_from_bf16(const BFloat16* data, Vectorized<float>& out) {
-  __at_align__ float values[Vectorized<float>::size()];
-  for (const auto k : c10::irange(Vectorized<float>::size())) {
-    values[k] = data[k];
-  }
-  out = Vectorized<float>::loadu(values);
-}
-
-inline void load_fp32_from_bf16(
-    const BFloat16* data,
-    Vectorized<float>& out1,
-    Vectorized<float>& out2) {
-  Vectorized<BFloat16> bf16_vec = Vectorized<BFloat16>::loadu(data);
-  auto floats = convert_bfloat16_float(bf16_vec);
-  out1 = std::get<0>(floats);
-  out2 = std::get<1>(floats);
-}
-
 template <typename Op>
 Vectorized<c10::BFloat16> binary_operator_via_float(
     Op op,
@@ -622,12 +579,6 @@ Vectorized<c10::BFloat16> inline fnmsub(
   return -a * b - c;
 }
 
-#else //
-
-CONVERT_NON_VECTORIZED_INIT(BFloat16, bfloat16)
-
-LOAD_FP32_NON_VECTORIZED_INIT(BFloat16, bf16)
-
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec128/vec128_convert.h

@@ -4,7 +4,7 @@
 
 namespace at::vec {
 inline namespace CPU_CAPABILITY {
-#if defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256)
+#if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
 template <typename src_t>
 struct VecConvert<
     float,
@@ -60,7 +60,6 @@ struct VecConvert<float, 1, BFloat16, 1> {
   }
 };
 
-#endif // defined(__aarch64__) && (!defined(CPU_CAPABILITY_SVE) ||
-       // defined(CPU_CAPABILITY_SVE128))
+#endif // defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256)
 } // namespace CPU_CAPABILITY
 } // namespace at::vec

aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h

@@ -4,10 +4,13 @@
 // See Note [Do not compile initializers with AVX]
 
 #include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/sve/sve_helper.h>
 #include <ATen/cpu/vec/vec_base.h>
 #include <c10/util/irange.h>
 
+#if defined(__aarch64__) && defined(AT_BUILD_ARM_VEC256_WITH_SLEEF)
+#include <sleef.h>
+#endif
+
 // Sleef offers vectorized versions of some transcedentals
 // such as sin, cos, tan etc..
 // However for now opting for STL, since we are not building
@@ -32,6 +35,12 @@ inline namespace CPU_CAPABILITY {
 #error "Big endian is not supported."
 #endif
 
+#if defined(AT_BUILD_ARM_VEC256_WITH_SLEEF)
+#define USE_SLEEF(sleef_code, non_sleef_code) sleef_code
+#else
+#define USE_SLEEF(sleef_code, non_sleef_code) non_sleef_code
+#endif
+
 template <int index, bool mask_val>
 struct BlendRegs {
   static float32x4_t impl(
@@ -85,12 +94,6 @@ class Vectorized<float> {
   operator float32x4_t() const {
     return values;
   }
-#ifdef CPU_CAPABILITY_SVE128
-  Vectorized(svfloat32_t v) : values(svget_neonq(v)) {}
-  operator svfloat32_t() const {
-    return svset_neonq(svundef_f32(), values);
-  }
-#endif
   template <int64_t mask>
   static Vectorized<float> blend(
       const Vectorized<float>& a,

aten/src/ATen/cpu/vec/vec128/vec128_half_neon.h

@@ -4,6 +4,7 @@
 // See Note [Do not compile initializers with AVX]
 
 #include <ATen/cpu/vec/intrinsics.h>
+#include <ATen/cpu/vec/vec128/vec128_convert.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_reduced_precision_common_neon.h>
 #include <ATen/cpu/vec/vec_base.h>
@@ -24,6 +25,7 @@ inline namespace CPU_CAPABILITY {
 //    https://bugs.llvm.org/show_bug.cgi?id=45824
 // Most likely we will do aarch32 support with inline asm.
 #if !defined(C10_MOBILE) && defined(__aarch64__)
+
 #ifdef __BIG_ENDIAN__
 #error "Big endian is not supported."
 #endif
@@ -419,24 +421,6 @@ Vectorized<c10::Half> inline operator+(
 #endif
 }
 
-inline void load_fp32_from_fp16(const c10::Half* data, Vectorized<float>& out) {
-  __at_align__ float values[Vectorized<float>::size()];
-  for (const auto k : c10::irange(Vectorized<float>::size())) {
-    values[k] = data[k];
-  }
-  out = Vectorized<float>::loadu(values);
-}
-
-inline void load_fp32_from_fp16(
-    const c10::Half* data,
-    Vectorized<float>& out1,
-    Vectorized<float>& out2) {
-  Vectorized<c10::Half> f16_vec = Vectorized<c10::Half>::loadu(data);
-  auto floats = convert_half_float(f16_vec);
-  out1 = std::get<0>(floats);
-  out2 = std::get<1>(floats);
-}
-
 template <>
 Vectorized<c10::Half> inline operator-(
     const Vectorized<c10::Half>& a,
@@ -672,53 +656,6 @@ Vectorized<c10::Half> inline fnmsub(
   return -a * b - c;
 #endif
 }
-
-#else
-
-#define CONVERT_NON_VECTORIZED_INIT(type, name)                     \
-  inline std::tuple<Vectorized<float>, Vectorized<float>>           \
-      convert_##name##_float(const Vectorized<type>& a) {           \
-    constexpr int64_t K = Vectorized<type>::size();                 \
-    __at_align__ float arr[K];                                      \
-    __at_align__ type arr2[K];                                      \
-    a.store(arr2);                                                  \
-    convert(arr2, arr, K);                                          \
-    return std::make_tuple(                                         \
-        Vectorized<float>::loadu(arr),                              \
-        Vectorized<float>::loadu(arr + Vectorized<float>::size())); \
-  }                                                                 \
-  inline Vectorized<type> convert_float_##name(                     \
-      const Vectorized<float>& a, const Vectorized<float>& b) {     \
-    constexpr int64_t K = Vectorized<type>::size();                 \
-    __at_align__ float arr[K];                                      \
-    __at_align__ type arr2[K];                                      \
-    a.store(arr);                                                   \
-    b.store(arr + Vectorized<float>::size());                       \
-    convert(arr, arr2, K);                                          \
-    return Vectorized<type>::loadu(arr2);                           \
-  }
-
-#define LOAD_FP32_NON_VECTORIZED_INIT(type, name)                           \
-  inline void load_fp32_from_##name(                                        \
-      const type* data, Vectorized<float>& out) {                           \
-    __at_align__ float values[Vectorized<float>::size()];                   \
-    for (const auto k : c10::irange(Vectorized<float>::size())) {           \
-      values[k] = data[k];                                                  \
-    }                                                                       \
-    out = Vectorized<float>::loadu(values);                                 \
-  }                                                                         \
-                                                                            \
-  inline void load_fp32_from_##name(                                        \
-      const type* data, Vectorized<float>& out1, Vectorized<float>& out2) { \
-    load_fp32_from_##name(data, out1);                                      \
-    data += Vectorized<float>::size();                                      \
-    load_fp32_from_##name(data, out2);                                      \
-  }
-
-CONVERT_NON_VECTORIZED_INIT(Half, half)
-
-LOAD_FP32_NON_VECTORIZED_INIT(Half, fp16)
-
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
 
 } // namespace CPU_CAPABILITY

aten/src/ATen/cpu/vec/vec256/vec256.h

@@ -9,16 +9,21 @@
 #if !(                                                 \
     defined(__VSX__) || defined(CPU_CAPABILITY_VSX) || \
     defined(CPU_CAPABILITY_ZVECTOR))
-#include <ATen/cpu/vec/vec256/vec256_double.h>
+#if defined(CPU_CAPABILITY_SVE256)
+#include <ATen/cpu/vec/sve/vec_common_sve.h>
+#else
+// clang-format off
 #include <ATen/cpu/vec/vec256/vec256_float.h>
+#include <ATen/cpu/vec/vec256/vec256_double.h>
 #include <ATen/cpu/vec/vec256/vec256_int.h>
 #include <ATen/cpu/vec/vec256/vec256_qint.h>
+#endif
 #if !defined(CPU_CAPABILITY_SVE256) || !defined(__ARM_FEATURE_BF16)
 #include <ATen/cpu/vec/vec256/vec256_bfloat16.h>
 #endif
-#include <ATen/cpu/vec/vec256/vec256_complex_double.h>
-#include <ATen/cpu/vec/vec256/vec256_complex_float.h>
 #include <ATen/cpu/vec/vec256/vec256_half.h>
+#include <ATen/cpu/vec/vec256/vec256_complex_float.h>
+#include <ATen/cpu/vec/vec256/vec256_complex_double.h>
 // clang-format on
 #elif defined(__VSX__) || defined(CPU_CAPABILITY_VSX)
 #include <ATen/cpu/vec/vec256/vsx/vec256_common_vsx.h>
@@ -51,6 +56,34 @@ namespace at::vec {
 // accessed as `at::vec`.
 inline namespace CPU_CAPABILITY {
 
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
+  stream << val.val_;
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
+  stream << static_cast<int>(val.val_);
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
+  stream << static_cast<unsigned int>(val.val_);
+  return stream;
+}
+
+template <typename T>
+std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
+  T buf[Vectorized<T>::size()];
+  vec.store(buf);
+  stream << "vec[";
+  for (int i = 0; i != Vectorized<T>::size(); i++) {
+    if (i != 0) {
+      stream << ", ";
+    }
+    stream << buf[i];
+  }
+  stream << "]";
+  return stream;
+}
+
 #if defined(CPU_CAPABILITY_AVX2)
 
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CAST (AVX2) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h

@@ -268,7 +268,9 @@ LOAD_FP32_VECTORIZED_INIT(BFloat16, bf16)
 
 #else // defined(CPU_CAPABILITY_AVX2)
 
-#if !(defined(__aarch64__))
+#if !(                                                                      \
+    defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
+    !defined(CPU_CAPABILITY_SVE256))
 CONVERT_NON_VECTORIZED_INIT(BFloat16, bfloat16)
 #endif
 

aten/src/ATen/cpu/vec/vec256/vec256_complex_double.h

@@ -342,19 +342,19 @@ class Vectorized<c10::complex<double>> {
     return _mm256_cmp_pd(values, other.values, _CMP_NEQ_UQ);
   }
   Vectorized<c10::complex<double>> operator<(
-      const Vectorized<c10::complex<double>>&) const {
+      const Vectorized<c10::complex<double>>& /*unused*/) const {
     TORCH_CHECK(false, "not supported for complex numbers");
   }
   Vectorized<c10::complex<double>> operator<=(
-      const Vectorized<c10::complex<double>>&) const {
+      const Vectorized<c10::complex<double>>& /*unused*/) const {
     TORCH_CHECK(false, "not supported for complex numbers");
   }
   Vectorized<c10::complex<double>> operator>(
-      const Vectorized<c10::complex<double>>&) const {
+      const Vectorized<c10::complex<double>>& /*unused*/) const {
     TORCH_CHECK(false, "not supported for complex numbers");
   }
   Vectorized<c10::complex<double>> operator>=(
-      const Vectorized<c10::complex<double>>&) const {
+      const Vectorized<c10::complex<double>>& /*unused*/) const {
     TORCH_CHECK(false, "not supported for complex numbers");
   }
 

aten/src/ATen/cpu/vec/vec256/vec256_half.h

@@ -268,7 +268,9 @@ LOAD_FP32_VECTORIZED_INIT(Half, fp16)
 
 #else // defined(CPU_CAPABILITY_AVX2)
 
-#if !defined(__aarch64__) || defined(CPU_CAPABILITY_SVE256)
+#if !(                                                                      \
+    defined(__aarch64__) && !defined(C10_MOBILE) && !defined(__CUDACC__) && \
+    !defined(CPU_CAPABILITY_SVE256))
 CONVERT_NON_VECTORIZED_INIT(Half, half)
 #endif
 

aten/src/ATen/cpu/vec/vec256/vec256_qint.h

@@ -5,13 +5,6 @@
 
 #include <ATen/cpu/vec/intrinsics.h>
 #include <ATen/cpu/vec/vec_base.h>
-
-#ifdef __aarch64__
-#if defined(CPU_CAPABILITY_SVE128) || !defined(CPU_CAPABILITY_SVE)
-#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
-#endif
-#endif
-
 #include <ATen/native/quantized/AffineQuantizerBase.h>
 
 #include <c10/util/irange.h>
@@ -922,7 +915,7 @@ Vectorized<c10::quint8> inline maximum(
   return a.maximum(b);
 }
 
-#else
+#elif !defined(CPU_CAPABILITY_SVE256)
 
 // NOTE: These are low-performance implementations that we fall back on
 // if we are not building with AVX2. This may not be an issue, because
@@ -1379,18 +1372,12 @@ Vectorized<c10::quint8> inline maximum(
   return a.maximum(b);
 }
 
-#if defined(__aarch64__) && \
-    (defined(CPU_CAPABILITY_SVE128) || !defined(CPU_CAPABILITY_SVE))
+#endif // if defined(CPU_CAPABILITY_AVX2)
+
+#if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
 std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
     at::vec::Vectorized<int8_t> src) {
-
-#ifdef CPU_CAPABILITY_SVE
-  svint8_t x = src;
-  auto s8x8 = vget_low_s8(svget_neonq(x));
-#else
   auto s8x8 = vld1_s8(src.operator const int8_t*());
-#endif
-
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_hi = vmovl_s16(vget_high_s16(s16x8));
@@ -1415,14 +1402,7 @@ std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
 
 Vectorized<float> inline convert_int8_half_register_to_float(
     at::vec::Vectorized<int8_t> src) {
-
-#ifdef CPU_CAPABILITY_SVE
-  svint8_t x = src;
-  auto s8x8 = vget_low_s8(svget_neonq(x));
-#else
   auto s8x8 = vld1_s8(src.operator const int8_t*());
-#endif
-
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_lo = vmovl_s16(vget_low_s16(s16x8));
@@ -1440,8 +1420,5 @@ Vectorized<float> inline convert_int8_half_register_to_float(
 }
 
 #endif
-
-#endif // if defined(CPU_CAPABILITY_AVX2)
-
 } // namespace CPU_CAPABILITY
 } // namespace at::vec

aten/src/ATen/cpu/vec/vec512/vec512.h

@@ -31,6 +31,34 @@ namespace vec {
 // See Note [CPU_CAPABILITY namespace]
 inline namespace CPU_CAPABILITY {
 
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint32& val) {
+  stream << val.val_;
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::qint8& val) {
+  stream << static_cast<int>(val.val_);
+  return stream;
+}
+inline std::ostream& operator<<(std::ostream& stream, const c10::quint8& val) {
+  stream << static_cast<unsigned int>(val.val_);
+  return stream;
+}
+
+template <typename T>
+std::ostream& operator<<(std::ostream& stream, const Vectorized<T>& vec) {
+  T buf[Vectorized<T>::size()];
+  vec.store(buf);
+  stream << "vec[";
+  for (int i = 0; i != Vectorized<T>::size(); i++) {
+    if (i != 0) {
+      stream << ", ";
+    }
+    stream << buf[i];
+  }
+  stream << "]";
+  return stream;
+}
+
 #if defined(CPU_CAPABILITY_AVX512)
 
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CAST (AVX512)

aten/src/ATen/cpu/vec/vec_base.h

@@ -67,7 +67,18 @@ Windows llvm will not have this definition.
 #endif
 #define VECTOR_WIDTH 64
 #define int_vector __m512i
-#elif defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_SVE256)
+#elif defined(__aarch64__) && \
+    !defined(CPU_CAPABILITY_SVE) // CPU_CAPABILITY_AVX512
+// SVE code expects 256-vectors; leave that set for SVE?
+#if defined(__GNUC__)
+#define __at_align__ __attribute__((aligned(16)))
+#elif defined(_WIN32)
+#define __at_align__ __declspec(align(16))
+#else
+#define __at_align__
+#endif
+#define VECTOR_WIDTH 16
+#else // CPU_CAPABILITY_AVX512
 #if defined(__GNUC__)
 #define __at_align__ __attribute__((aligned(32)))
 #elif defined(_WIN32)
@@ -77,27 +88,7 @@ Windows llvm will not have this definition.
 #endif
 #define VECTOR_WIDTH 32
 #define int_vector __m256i
-#elif defined(__aarch64__)
-// Define alignment and vector width for SVE128/Default (e.g., NEON)
-#if defined(__GNUC__)
-#define __at_align__ __attribute__((aligned(16)))
-#elif defined(_WIN32)
-#define __at_align__ __declspec(align(16))
-#else
-#define __at_align__
-#endif
-#define VECTOR_WIDTH 16
-#else
-// Fallback: define default alignment and vector width
-#if defined(__GNUC__)
-#define __at_align__ __attribute__((aligned(32)))
-#elif defined(_WIN32)
-#define __at_align__ __declspec(align(32))
-#else
-#define __at_align__
-#endif
-#define VECTOR_WIDTH 32
-#endif
+#endif // CPU_CAPABILITY_AVX512
 
 namespace at::vec {
 // See Note [CPU_CAPABILITY namespace]

aten/src/ATen/cuda/CUDABlas.cpp

@@ -422,18 +422,34 @@ static inline bool bgemm_internal_cublaslt(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(D
     abType = CUDA_R_16F;
     cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16F;
 #ifndef USE_ROCM
-    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
-      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
-        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
+    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
+    if (fp16_reduction !=
+        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+      uint32_t mask =
+          fp16_reduction ==
+                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
+              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
+                 CUBLASLT_REDUCTION_SCHEME_NONE)
+              : CUBLASLT_REDUCTION_SCHEME_NONE;
+      preference.setAttribute(
+          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
     }
 #endif
   } else if constexpr (std::is_same_v<Dtype, at::BFloat16>) {
     abType = CUDA_R_16BF;
     cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16BF;
 #ifndef USE_ROCM
-    if (!at::globalContext().allowBF16ReductionCuBLAS()) {
-      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
-        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
+    auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
+    if (bf16_reduction !=
+        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+      uint32_t mask =
+          bf16_reduction ==
+                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
+              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
+                 CUBLASLT_REDUCTION_SCHEME_NONE)
+              : CUBLASLT_REDUCTION_SCHEME_NONE;
+      preference.setAttribute(
+          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
     }
 #endif
   } else {
@@ -1120,8 +1136,15 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
   }
   if (prop->major >= 5) {
     cublasMath_t cublas_flags = CUBLAS_DEFAULT_MATH;
-    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
-      cublas_flags = static_cast<cublasMath_t>(cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
+    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
+    TORCH_CHECK(fp16_reduction !=
+        at::CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK,
+          "torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction("
+          "..., allow_splitk=False) requires the cuBLASLt backend");
+    if (fp16_reduction !=
+        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+      cublas_flags = static_cast<cublasMath_t>(
+          cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
     }
     // Disallow fp16 reductions that could lead to unexpected overflow issues.
     TORCH_CUDABLAS_CHECK(cublasSetMathMode(handle, cublas_flags));
@@ -1180,8 +1203,15 @@ inline void gemm_internal_cublas_bfloat16_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DT
   GEMM_CHECK_ARGVALUES(at::BFloat16);
 #ifndef USE_ROCM
   cublasMath_t cublas_flags = CUBLAS_DEFAULT_MATH;
-  if (!at::globalContext().allowBF16ReductionCuBLAS()) {
-    cublas_flags = static_cast<cublasMath_t>(cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
+  auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
+  TORCH_CHECK(bf16_reduction !=
+      at::CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK,
+        "torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction("
+        "..., allow_splitk=False) requires the cuBLASLt backend");
+  if (bf16_reduction !=
+      at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+    cublas_flags = static_cast<cublasMath_t>(
+        cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
   }
 #endif
 #if defined(USE_ROCM)
@@ -1270,7 +1300,7 @@ void gemm_internal<float>(CUDABLAS_GEMM_ARGTYPES(float))
   }
 #if defined(USE_ROCM) && defined(USE_ROCM_CK_GEMM)
   else if (at::globalContext().blasPreferredBackend() == BlasBackend::Ck) {
-    if (at::detail::getCUDAHooks().isGPUArch({"gfx1100"})) { //no CK GEMM version for gfx1100
+    if (at::detail::getCUDAHooks().isGPUArch({"gfx11", "gfx12"})) { //no CK GEMM version
       gemm_internal_cublaslt<float>(CUDABLAS_GEMM_ARGS(float));
     } else{
       at::native::gemm_internal_ck<float>(CUDABLAS_GEMM_ARGS(float));
@@ -1577,18 +1607,34 @@ bool gemm_and_bias(
     abType = CUDA_R_16F;
     cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16F;
 #ifndef USE_ROCM
-    if (!at::globalContext().allowFP16ReductionCuBLAS()) {
-      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
-        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
+    auto fp16_reduction = at::globalContext().allowFP16ReductionCuBLAS();
+    if (fp16_reduction !=
+        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+      uint32_t mask =
+          fp16_reduction ==
+                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
+              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
+                 CUBLASLT_REDUCTION_SCHEME_NONE)
+              : CUBLASLT_REDUCTION_SCHEME_NONE;
+      preference.setAttribute(
+          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
     }
 #endif
   } else if constexpr (std::is_same_v<Dtype, at::BFloat16>) {
     abType = CUDA_R_16BF;
     cType = (std::is_same_v<C_Dtype, float>) ? CUDA_R_32F : CUDA_R_16BF;
 #ifndef USE_ROCM
-    if (!at::globalContext().allowBF16ReductionCuBLAS()) {
-      preference.setAttribute(CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK,
-        CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE | CUBLASLT_REDUCTION_SCHEME_NONE);
+    auto bf16_reduction = at::globalContext().allowBF16ReductionCuBLAS();
+    if (bf16_reduction !=
+        at::CuBLASReductionOption::AllowReducedPrecisionWithSplitK) {
+      uint32_t mask =
+          bf16_reduction ==
+                  at::CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK
+              ? (CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE |
+                 CUBLASLT_REDUCTION_SCHEME_NONE)
+              : CUBLASLT_REDUCTION_SCHEME_NONE;
+      preference.setAttribute(
+          CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, mask);
     }
 #endif
   }
@@ -1815,6 +1861,8 @@ template bool gemm_and_bias(
     int64_t result_ld,
     GEMMAndBiasActivationEpilogue activation);
 
+using at::blas::ScalingType;
+
 int get_scale_mode(ScalingType scaling_type, ScalarType scale_dtype, bool use_fast_accum) {
   switch (scaling_type) {
     case ScalingType::BlockWise1x32:

aten/src/ATen/cuda/CUDABlas.h

@@ -14,6 +14,7 @@
  */
 
 #include <ATen/cuda/CUDAContext.h>
+#include <ATen/BlasBackend.h>
 #include <ATen/OpMathType.h>
 
 namespace at::cuda::blas {
@@ -136,15 +137,6 @@ void int8_gemm(
     int32_t* result_ptr,
     int64_t result_ld);
 
-enum class ScalingType : std::uint8_t {
-  TensorWise,  // fp32 scales
-  RowWise,  // fp32 scales
-  BlockWise1x16,  // fp8_e4m3fn scales
-  BlockWise1x32,  // fp8_e8m0fnu scales
-  BlockWise1x128,  // fp32 scales
-  BlockWise128x128,  // fp32 scales
-};
-
 void scaled_gemm(
     char transa,
     char transb,
@@ -156,13 +148,13 @@ void scaled_gemm(
     int64_t mat1_ld,
     ScalarType mat1_dtype,
     ScalarType mat1_scale_dtype,
-    ScalingType mat1_scaling_type,
+    at::blas::ScalingType mat1_scaling_type,
     const void* mat2_ptr,
     const void* mat2_scale_ptr,
     int64_t mat2_ld,
     ScalarType mat2_dtype,
     ScalarType mat2_scale_dtype,
-    ScalingType mat2_scaling_type,
+    at::blas::ScalingType mat2_scaling_type,
     const void* bias_ptr,
     ScalarType bias_dtype,
     void* result_ptr,