Fix indentation

* Added a local `main_stream` copy of the global VSLStream
    and changed the kernel logic to immediately advance the global
    stream by the number of generated elements upon obtaining the copy.
    This fixes an issue where previously calling these kernels in
    multiple concurrent python threads could yield identical generated
    sequences (since the advance in the global stream happened only
    after the generation step was finished).

.ci/magma-rocm/README.md

@@ -30,6 +30,7 @@ into a tarball, with the following structure:
 More specifically, `build_magma.sh` copies over the relevant files from the `package_files` directory depending on the ROCm version.
 Outputted binaries should be in the `output` folder.
 
+
 ## Pushing
 
 Packages can be uploaded to an S3 bucket using:

.ci/pytorch/common_utils.sh

@@ -96,6 +96,7 @@ function pip_build_and_install() {
     python3 -m pip wheel \
       --no-build-isolation \
       --no-deps \
+      --no-use-pep517 \
       -w "${wheel_dir}" \
       "${build_target}"
   fi
@@ -307,28 +308,6 @@ function install_torchao() {
   pip_build_and_install "git+https://github.com/pytorch/ao.git@${commit}" dist/ao
 }
 
-function install_flash_attn_cute() {
-  echo "Installing FlashAttention CuTe from GitHub..."
-  # Grab latest main til we have a pinned commit
-  local flash_attn_commit
-  flash_attn_commit=$(git ls-remote https://github.com/Dao-AILab/flash-attention.git HEAD | cut -f1)
-
-  # Clone the repo to a temporary directory
-  rm -rf flash-attention-build
-  git clone --depth 1 --recursive https://github.com/Dao-AILab/flash-attention.git flash-attention-build
-
-  pushd flash-attention-build
-  git checkout "${flash_attn_commit}"
-
-  # Install only the 'cute' sub-directory
-  pip_install -e flash_attn/cute/
-  popd
-
-  # remove the local repo
-  rm -rf flash-attention-build
-  echo "FlashAttention CuTe installation complete."
-}
-
 function print_sccache_stats() {
   echo 'PyTorch Build Statistics'
   sccache --show-stats

.ci/pytorch/smoke_test/check_binary_symbols.py

@@ -100,337 +100,6 @@ def check_lib_statically_linked_libstdc_cxx_abi_symbols(lib: str) -> None:
         )
 
 
-def _compile_and_extract_symbols(
-    cpp_content: str, compile_flags: list[str], exclude_list: list[str] | None = None
-) -> list[str]:
-    """
-    Helper to compile a C++ file and extract all symbols.
-
-    Args:
-        cpp_content: C++ source code to compile
-        compile_flags: Compilation flags
-        exclude_list: List of symbol names to exclude. Defaults to ["main"].
-
-    Returns:
-        List of all symbols found in the object file (excluding those in exclude_list).
-    """
-    import subprocess
-    import tempfile
-
-    if exclude_list is None:
-        exclude_list = ["main"]
-
-    with tempfile.TemporaryDirectory() as tmpdir:
-        tmppath = Path(tmpdir)
-        cpp_file = tmppath / "test.cpp"
-        obj_file = tmppath / "test.o"
-
-        cpp_file.write_text(cpp_content)
-
-        result = subprocess.run(
-            compile_flags + [str(cpp_file), "-o", str(obj_file)],
-            capture_output=True,
-            text=True,
-            timeout=60,
-        )
-
-        if result.returncode != 0:
-            raise RuntimeError(f"Compilation failed: {result.stderr}")
-
-        symbols = get_symbols(str(obj_file))
-
-        # Return all symbol names, excluding those in the exclude list
-        return [name for _addr, _stype, name in symbols if name not in exclude_list]
-
-
-def check_stable_only_symbols(install_root: Path) -> None:
-    """
-    Test TORCH_STABLE_ONLY and TORCH_TARGET_VERSION by compiling test code and comparing symbol counts.
-
-    This approach tests:
-    1. WITHOUT macros -> many torch symbols exposed
-    2. WITH TORCH_STABLE_ONLY -> zero torch symbols (all hidden)
-    3. WITH TORCH_TARGET_VERSION -> zero torch symbols (all hidden)
-    4. WITH both macros -> zero torch symbols (all hidden)
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    test_cpp_content = """
-// Main torch C++ API headers
-#include <torch/torch.h>
-#include <torch/all.h>
-
-// ATen tensor library
-#include <ATen/ATen.h>
-
-// Core c10 headers (commonly used)
-#include <c10/core/Device.h>
-#include <c10/core/DeviceType.h>
-#include <c10/core/ScalarType.h>
-#include <c10/core/TensorOptions.h>
-#include <c10/util/Optional.h>
-
-int main() { return 0; }
-"""
-
-    base_compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",  # Compile only, don't link
-    ]
-
-    # Compile WITHOUT any macros
-    symbols_without = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=base_compile_flags,
-    )
-
-    # We expect constexpr symbols, inline functions used by other headers etc.
-    # to produce symbols
-    num_symbols_without = len(symbols_without)
-    print(f"Found {num_symbols_without} symbols without any macros defined")
-    assert num_symbols_without != 0, (
-        "Expected a non-zero number of symbols without any macros"
-    )
-
-    # Compile WITH TORCH_STABLE_ONLY (expect 0 symbols)
-    compile_flags_with_stable_only = base_compile_flags + ["-DTORCH_STABLE_ONLY"]
-
-    symbols_with_stable_only = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_stable_only,
-    )
-
-    num_symbols_with_stable_only = len(symbols_with_stable_only)
-    assert num_symbols_with_stable_only == 0, (
-        f"Expected no symbols with TORCH_STABLE_ONLY macro, but found {num_symbols_with_stable_only}"
-    )
-
-    # Compile WITH TORCH_TARGET_VERSION (expect 0 symbols)
-    compile_flags_with_target_version = base_compile_flags + [
-        "-DTORCH_TARGET_VERSION=1"
-    ]
-
-    symbols_with_target_version = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_target_version,
-    )
-
-    num_symbols_with_target_version = len(symbols_with_target_version)
-    assert num_symbols_with_target_version == 0, (
-        f"Expected no symbols with TORCH_TARGET_VERSION macro, but found {num_symbols_with_target_version}"
-    )
-
-    # Compile WITH both macros (expect 0 symbols)
-    compile_flags_with_both = base_compile_flags + [
-        "-DTORCH_STABLE_ONLY",
-        "-DTORCH_TARGET_VERSION=1",
-    ]
-
-    symbols_with_both = _compile_and_extract_symbols(
-        cpp_content=test_cpp_content,
-        compile_flags=compile_flags_with_both,
-    )
-
-    num_symbols_with_both = len(symbols_with_both)
-    assert num_symbols_with_both == 0, (
-        f"Expected no symbols with both macros, but found {num_symbols_with_both}"
-    )
-
-
-def check_stable_api_symbols(install_root: Path) -> None:
-    """
-    Test that stable API headers still expose symbols with TORCH_STABLE_ONLY.
-    The torch/csrc/stable/c/shim.h header is tested in check_stable_c_shim_symbols
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    stable_dir = include_dir / "torch" / "csrc" / "stable"
-    assert stable_dir.exists(), f"Expected {stable_dir} to be present"
-
-    stable_headers = list(stable_dir.rglob("*.h"))
-    if not stable_headers:
-        raise RuntimeError("Could not find any stable headers")
-
-    includes = []
-    for header in stable_headers:
-        rel_path = header.relative_to(include_dir)
-        includes.append(f"#include <{rel_path.as_posix()}>")
-
-    includes_str = "\n".join(includes)
-    test_stable_content = f"""
-{includes_str}
-int main() {{ return 0; }}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_stable = _compile_and_extract_symbols(
-        cpp_content=test_stable_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_stable = len(symbols_stable)
-    print(f"Found {num_symbols_stable} symbols in torch/csrc/stable")
-    assert num_symbols_stable > 0, (
-        f"Expected stable headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_stable} symbols"
-    )
-
-
-def check_headeronly_symbols(install_root: Path) -> None:
-    """
-    Test that header-only utility headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # Find all headers in torch/headeronly
-    headeronly_dir = include_dir / "torch" / "headeronly"
-    assert headeronly_dir.exists(), f"Expected {headeronly_dir} to be present"
-    headeronly_headers = list(headeronly_dir.rglob("*.h"))
-    if not headeronly_headers:
-        raise RuntimeError("Could not find any headeronly headers")
-
-    # Filter out platform-specific headers that may not compile everywhere
-    platform_specific_keywords = [
-        "cpu/vec",
-    ]
-
-    filtered_headers = []
-    for header in headeronly_headers:
-        rel_path = header.relative_to(include_dir).as_posix()
-        if not any(
-            keyword in rel_path.lower() for keyword in platform_specific_keywords
-        ):
-            filtered_headers.append(header)
-
-    includes = []
-    for header in filtered_headers:
-        rel_path = header.relative_to(include_dir)
-        includes.append(f"#include <{rel_path.as_posix()}>")
-
-    includes_str = "\n".join(includes)
-    test_headeronly_content = f"""
-{includes_str}
-int main() {{ return 0; }}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_headeronly = _compile_and_extract_symbols(
-        cpp_content=test_headeronly_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_headeronly = len(symbols_headeronly)
-    print(f"Found {num_symbols_headeronly} symbols in torch/headeronly")
-    assert num_symbols_headeronly > 0, (
-        f"Expected headeronly headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_headeronly} symbols"
-    )
-
-
-def check_aoti_shim_symbols(install_root: Path) -> None:
-    """
-    Test that AOTI shim headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # There are no constexpr symbols etc., so we need to actually use functions
-    # so that some symbols are found.
-    test_shim_content = """
-#include <torch/csrc/inductor/aoti_torch/c/shim.h>
-int main() {
-    int32_t (*fp1)() = &aoti_torch_device_type_cpu;
-    int32_t (*fp2)() = &aoti_torch_dtype_float32;
-    (void)fp1; (void)fp2;
-    return 0;
-}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_shim = _compile_and_extract_symbols(
-        cpp_content=test_shim_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_shim = len(symbols_shim)
-    assert num_symbols_shim > 0, (
-        f"Expected shim headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_shim} symbols"
-    )
-
-
-def check_stable_c_shim_symbols(install_root: Path) -> None:
-    """
-    Test that stable C shim headers still expose symbols with TORCH_STABLE_ONLY.
-    """
-    include_dir = install_root / "include"
-    assert include_dir.exists(), f"Expected {include_dir} to be present"
-
-    # Check if the stable C shim exists
-    stable_shim = include_dir / "torch" / "csrc" / "stable" / "c" / "shim.h"
-    if not stable_shim.exists():
-        raise RuntimeError("Could not find stable c shim")
-
-    # There are no constexpr symbols etc., so we need to actually use functions
-    # so that some symbols are found.
-    test_stable_shim_content = """
-#include <torch/csrc/stable/c/shim.h>
-int main() {
-    // Reference stable C API functions to create undefined symbols
-    AOTITorchError (*fp1)(const char*, uint32_t*, int32_t*) = &torch_parse_device_string;
-    AOTITorchError (*fp2)(uint32_t*) = &torch_get_num_threads;
-    (void)fp1; (void)fp2;
-    return 0;
-}
-"""
-
-    compile_flags = [
-        "g++",
-        "-std=c++17",
-        f"-I{include_dir}",
-        f"-I{include_dir}/torch/csrc/api/include",
-        "-c",
-        "-DTORCH_STABLE_ONLY",
-    ]
-
-    symbols_stable_shim = _compile_and_extract_symbols(
-        cpp_content=test_stable_shim_content,
-        compile_flags=compile_flags,
-    )
-    num_symbols_stable_shim = len(symbols_stable_shim)
-    assert num_symbols_stable_shim > 0, (
-        f"Expected stable C shim headers to expose symbols with TORCH_STABLE_ONLY, "
-        f"but found {num_symbols_stable_shim} symbols"
-    )
-
-
 def check_lib_symbols_for_abi_correctness(lib: str) -> None:
     print(f"lib: {lib}")
     cxx11_symbols = grep_symbols(lib, LIBTORCH_CXX11_PATTERNS)
@@ -460,13 +129,6 @@ def main() -> None:
     check_lib_symbols_for_abi_correctness(libtorch_cpu_path)
     check_lib_statically_linked_libstdc_cxx_abi_symbols(libtorch_cpu_path)
 
-    # Check symbols when TORCH_STABLE_ONLY is defined
-    check_stable_only_symbols(install_root)
-    check_stable_api_symbols(install_root)
-    check_headeronly_symbols(install_root)
-    check_aoti_shim_symbols(install_root)
-    check_stable_c_shim_symbols(install_root)
-
 
 if __name__ == "__main__":
     main()

.ci/pytorch/smoke_test/smoke_test.py

@@ -353,17 +353,6 @@ def test_linalg(device="cpu") -> None:
             torch.linalg.svd(A)
 
 
-def test_sdpa(device="cpu", dtype=torch.float16) -> None:
-    """Regression test for https://github.com/pytorch/pytorch/issues/167602
-    Without nvrtc_builtins on CuDNN-9.13 on CUDA-13 fails with ` No valid execution plans built.`
-    """
-    print(f"Testing SDPA on {device} using type {dtype}")
-    k, q, v = torch.rand(3, 1, 16, 77, 64, dtype=dtype, device=device).unbind(0)
-    attn = torch.rand(1, 1, 77, 77, dtype=dtype, device=device)
-    rc = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn)
-    assert rc.isnan().any().item() is False
-
-
 def smoke_test_compile(device: str = "cpu") -> None:
     supported_dtypes = [torch.float16, torch.float32, torch.float64]
 
@@ -500,12 +489,10 @@ def main() -> None:
     smoke_test_conv2d()
     test_linalg()
     test_numpy()
-    test_sdpa()
 
     if is_cuda_system:
         test_linalg("cuda")
         test_cuda_gds_errors_captured()
-        test_sdpa("cuda")
 
     if options.package == "all":
         smoke_test_modules()

.ci/pytorch/test.sh

@@ -344,18 +344,8 @@ test_python_smoke() {
 }
 
 test_python_smoke_b200() {
-  # Targeted smoke tests for B200 including FlashAttention CuTe coverage
-  install_flash_attn_cute
-  time python test/run_test.py \
-    --include \
-      test_matmul_cuda \
-      test_scaled_matmul_cuda \
-      inductor/test_fp8 \
-      nn/attention/test_fa4 \
-      nn/attention/test_open_registry \
-      inductor/test_flex_flash \
-    $PYTHON_TEST_EXTRA_OPTION \
-    --upload-artifacts-while-running
+  # Targeted smoke tests for B200 - staged approach to avoid too many failures
+  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 

.github/actionlint.yaml

@@ -63,7 +63,7 @@ self-hosted-runner:
     - linux.rocm.gpu.gfx942.1
     - linux.rocm.gpu.gfx942.2
     - linux.rocm.gpu.gfx942.4
-    - linux.rocm.gfx942.docker-cache
+    - rocm-docker
     # Org wise AWS `mac2.metal` runners (2020 Mac mini hardware powered by Apple silicon M1 processors)
     - macos-m1-stable
     - macos-m1-14

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-07b6cbde121417a70e4dc871adb6d27030e0ce3f
+ad5816f0eee1c873df1b7d371c69f1f811a89387

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-acccf86477759b2d3500f1ae1be065f7b1e409ec
+ccb801b88af136454798b945175c4c87e636ac33

.github/labeler.yml

@@ -165,16 +165,3 @@
 - torch/_inductor/kernel/mm.py
 - test/inductor/test_max_autotune.py
 - third_party/fbgemm
-
-"ciflow/mps":
-- aten/src/ATen/mps/**
-- aten/src/ATen/native/mps/**
-- torch/_inductor/codegen/mps.py
-- test/test_mps.py
-- test/inductor/test_mps_basic.py
-
-"ciflow/h100-symm-mem":
-- torch/csrc/distributed/c10d/symm_mem/**
-- torch/distributed/_symmetric_memory/**
-- test/distributed/**/*mem*
-- test/distributed/**/*mem*/**

.github/scripts/lintrunner.sh

@@ -34,9 +34,6 @@ python3 torch/utils/data/datapipes/gen_pyi.py
 # Also check generated pyi files
 find torch -name '*.pyi' -exec git add --force -- "{}" +
 
-# Print current environment
-python3 -m pip freeze
-
 RC=0
 # Run lintrunner on all files
 if ! lintrunner --force-color --tee-json=lint.json ${ADDITIONAL_LINTRUNNER_ARGS} 2> /dev/null; then

.github/workflows/docker-builds.yml

@@ -119,22 +119,6 @@ jobs:
         with:
           docker-image: ${{ steps.build-docker-image.outputs.docker-image }}
 
-      - name: Generate output
-        if: contains(matrix.docker-image-name, 'rocm')
-        id: generate_output
-        run: |
-          docker_image_name="${{ matrix.docker-image-name }}"
-          docker_image_tag="${{ steps.build-docker-image.outputs.docker-image }}"
-          echo "${docker_image_name}=${docker_image_tag}" >> docker-builds-output-${docker_image_name}.txt
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4.4.0
-        if: contains(matrix.docker-image-name, 'rocm')
-        with:
-          name: docker-builds-artifacts-${{ matrix.docker-image-name }}
-          retention-days: 14
-          path: ./docker-builds-output-${{ matrix.docker-image-name }}.txt
-
       - uses: nick-fields/retry@7152eba30c6575329ac0576536151aca5a72780e # v3.0.0
         name: Push to https://ghcr.io/
         id: push-to-ghcr-io

.github/workflows/docker-cache-mi300.yml

@@ -0,0 +1,55 @@
+name: docker-cache-mi300
+
+on:
+  # run every 6 hours
+  schedule:
+    - cron: 0 0,6,12,18 * * *
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  docker-cache:
+    if: github.repository_owner == 'pytorch'
+    runs-on: rocm-docker
+    steps:
+      - name: Checkout PyTorch
+        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
+        with:
+          no-sudo: true
+
+      - name: configure aws credentials
+        id: aws_creds
+        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+        with:
+          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+          aws-region: us-east-1
+          role-duration-seconds: 18000
+
+      - name: Login to Amazon ECR
+        id: login-ecr
+        continue-on-error: false
+        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
+
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+          push: false
+
+      - name: Pull docker image
+        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
+        with:
+          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
+
+      - name: Tar and upload to S3 bucket
+        run: |
+          sudo docker save -o ~/docker-data/pytorch/pytorch_docker_image.tar ${{ steps.calculate-docker-image.outputs.docker-image }}
+          sudo rclone copy -P --s3-upload-concurrency 64 --s3-chunk-size 200M --s3-upload-cutoff 300M ~/docker-data/pytorch/pytorch_docker_image.tar oci:pytorchbucket0002/pytorch_docker_image --progress

.github/workflows/docker-cache-rocm.yml

@@ -1,105 +0,0 @@
-name: docker-cache-rocm
-
-on:
-  workflow_run:
-    workflows: [docker-builds]
-    branches: [main, release]
-    types:
-      - completed
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-  actions: read
-
-jobs:
-  download-docker-builds-artifacts:
-    if: github.repository_owner == 'pytorch'
-    name: download-docker-builds-artifacts
-    runs-on: ubuntu-latest
-    outputs:
-      pytorch-linux-jammy-rocm-n-py3: ${{ steps.process-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3 }}
-      pytorch-linux-noble-rocm-n-py3: ${{ steps.process-artifacts.outputs.pytorch-linux-noble-rocm-n-py3 }}
-      pytorch-linux-jammy-rocm-n-py3-benchmarks: ${{ steps.process-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3-benchmarks }}
-    steps:
-      - name: Download artifacts
-        uses: actions/download-artifact@v4.1.7
-        with:
-          run-id: ${{ github.event.workflow_run.id }}
-          path: ./docker-builds-artifacts
-          merge-multiple: true
-          github-token: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Process artifacts
-        id: process-artifacts
-        run: |
-          ls -R ./docker-builds-artifacts
-          cat ./docker-builds-artifacts/*txt >> "${GITHUB_OUTPUT}"
-          cat "${GITHUB_OUTPUT}"
-
-  docker-cache:
-    if: github.repository_owner == 'pytorch'
-    needs: download-docker-builds-artifacts
-    strategy:
-      fail-fast: false
-      matrix:
-        runner: [linux.rocm.gfx942.docker-cache]
-        docker-image: [
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3 }}",
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-noble-rocm-n-py3 }}",
-          "${{ needs.download-docker-builds-artifacts.outputs.pytorch-linux-jammy-rocm-n-py3-benchmarks }}"
-        ]
-    runs-on: "${{ matrix.runner }}"
-    steps:
-      - name: debug
-        run: |
-          JSON_STRINGIFIED="${{ toJSON(needs.download-docker-builds-artifacts.outputs) }}"
-          echo "Outputs of download-docker-builds-artifacts job: ${JSON_STRINGIFIED}"
-
-      - name: configure aws credentials
-        id: aws_creds
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-          aws-region: us-east-1
-          role-duration-seconds: 18000
-
-      - name: Login to Amazon ECR
-        id: login-ecr
-        continue-on-error: false
-        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
-
-      - name: Generate ghrc.io tag
-        id: ghcr-io-tag
-        run: |
-            ecr_image="${{ matrix.docker-image }}"
-            ghcr_image="ghcr.io/pytorch/ci-image:${ecr_image##*:}"
-            echo "ghcr_image=${ghcr_image}" >> "$GITHUB_OUTPUT"
-
-      - name: Pull docker image
-        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-        with:
-          docker-image: ${{ steps.ghcr-io-tag.outputs.ghcr_image }}
-
-      - name: Save as tarball
-        run: |
-          docker_image_tag=${{ matrix.docker-image }}
-          docker_image_tag="${docker_image_tag#*:}" # Remove everything before and including first ":"
-          docker_image_tag="${docker_image_tag%-*}" # Remove everything after and including last "-"
-          ref_name=${{ github.event.workflow_run.head_branch }}
-          if [[ $ref_name =~ "release/" ]]; then
-            ref_suffix="release"
-          elif [[ $ref_name == "main" ]]; then
-            ref_suffix="main"
-          else
-            echo "Unexpected branch in ref_name: ${ref_name}" && exit 1
-          fi
-          docker tag ${{ steps.ghcr-io-tag.outputs.ghcr_image }} ${{ matrix.docker-image }}
-          # mv is atomic operation, so we use intermediate tar.tmp file to prevent read-write contention
-          docker save -o ~/pytorch-data/docker/${docker_image_tag}.tar.tmp ${{ matrix.docker-image }}
-          mv ~/pytorch-data/docker/${docker_image_tag}.tar.tmp ~/pytorch-data/docker/${docker_image_tag}_${ref_suffix}.tar

.github/workflows/h100-distributed.yml

@@ -37,6 +37,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      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-rocm-mi200.yml

@@ -1,4 +1,4 @@
-name: inductor-rocm-mi200
+name: inductor-rocm
 
 on:
   schedule:

.github/workflows/nightly.yml

@@ -5,11 +5,9 @@ on:
     - cron: 0 0 * * *
   push:
     tags:
-      # NOTE: Doc build pipelines should only get triggered on:
-      # Major or minor release candidates builds
-      - v[0-9]+.[0-9]+.0+-rc[0-9]+
-      # Final RC for major, minor and patch releases
-      - v[0-9]+.[0-9]+.[0-9]+
+      # NOTE: Doc build pipelines should only get triggered on release candidate builds
+      # Release candidate tags look like: v1.11.0-rc1
+      - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
       - ciflow/nightly/*
   workflow_dispatch:
 

.github/workflows/rocm-mi200.yml

@@ -1,4 +1,4 @@
-name: rocm-mi200
+name: rocm
 
 on:
   push:

.github/workflows/test-b200.yml

@@ -5,9 +5,7 @@
 # Flow:
 # 1. Builds PyTorch with CUDA 12.8+ and sm100 architecture for B200
 # 2. Runs smoke tests on linux.dgx.b200 runner
-# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke_b200() function
-#    - Includes matmul, scaled_matmul, FP8, and FlashAttention CuTe tests
-#    - FlashAttention CuTe DSL is installed as part of test execution
+# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke() function
 #
 # Triggered by:
 # - Pull requests modifying this workflow file

.github/workflows/test-h100.yml

@@ -41,6 +41,7 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.12xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm90
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       cuda-arch-list: '9.0'

.github/workflows/trunk-rocm-mi300.yml

@@ -1,83 +0,0 @@
-name: trunk-rocm-mi300
-
-on:
-  push:
-    branches:
-      - main
-      - release/*
-  workflow_dispatch:
-  schedule:
-    - cron: 29 8 * * *  # about 1:29am PDT
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  llm-td:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/llm_td_retrieval.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  target-determination:
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    needs: llm-td
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1.b" },
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.gfx942.4.b" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    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: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/upload-test-stats.yml

@@ -5,7 +5,6 @@ on:
     workflows:
       - pull
       - trunk
-      - trunk-rocm-mi300
       - periodic
       - periodic-rocm-mi200
       - periodic-rocm-mi300

.lintrunner.toml

@@ -186,8 +186,6 @@ include_patterns = [
     'aten/src/ATen/native/nested/cuda/*.h',
     'aten/src/ATen/native/nested/*.cpp',
     'aten/src/ATen/native/nested/*.h',
-    'aten/src/ATen/xpu/**/*.h',
-    'aten/src/ATen/xpu/**/*.cpp',
     'c10/**/*.cpp',
     'c10/**/*.h',
     'torch/*.h',

BUILD.bazel

@@ -195,6 +195,13 @@ filegroup(
     ]),
 )
 
+filegroup(
+    name = "aten_mklrng_cpp",
+    srcs = glob([
+        "aten/src/ATen/mklrng/*.cpp",
+    ]),
+)
+
 filegroup(
     name = "aten_native_mkldnn_cpp",
     srcs = glob(["aten/src/ATen/native/mkldnn/*.cpp"]),
@@ -357,6 +364,7 @@ cc_library(
         ":aten_base_cpp",
         ":aten_base_metal",
         ":aten_base_vulkan",
+        ":aten_mklrng_cpp",
         ":aten_native_cpp",
         ":aten_native_mkl_cpp",
         ":aten_native_mkldnn_cpp",

CMakeLists.txt

@@ -736,44 +736,6 @@ if(NOT DEFINED USE_BLAS)
   set(USE_BLAS ON)
 endif()
 
-# Prioritized Text Linker Optimization
-if(USE_PRIORITIZED_TEXT_FOR_LD)
-
-  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
-  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
-
-  execute_process(
-    COMMAND ${Python_EXECUTABLE}
-            ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py
-            --filein "${LINKER_SCRIPT_FILE_IN}"
-            --fout  "${LINKER_SCRIPT_FILE_OUT}"
-    RESULT_VARIABLE _gen_result
-    OUTPUT_VARIABLE _gen_output
-    ERROR_VARIABLE  _gen_error
-  )
-
-  if(NOT _gen_result EQUAL 0)
-    message(FATAL_ERROR
-      "Failed to generate linker script:\n${_gen_output}\n${_gen_error}")
-  endif()
-
-  append_cxx_flag_if_supported("-ffunction-sections" CMAKE_CXX_FLAGS)
-  append_cxx_flag_if_supported("-fdata-sections" CMAKE_CXX_FLAGS)
-  append_c_flag_if_supported("-ffunction-sections" CMAKE_C_FLAGS)
-  append_c_flag_if_supported("-fdata-sections" CMAKE_C_FLAGS)
-
-  set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -T${LINKER_SCRIPT_FILE_OUT}")
-  set(CMAKE_MODULE_LINKER_FLAGS "${CMAKE_MODULE_LINKER_FLAGS} -T${LINKER_SCRIPT_FILE_OUT}")
-
-else()
-  if(LINUX AND CPU_AARCH64)
-    message(WARNING [[
-    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
-    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
-    ]])
-  endif()
-endif()
-
 # Build libtorch mobile library, which contains ATen/TH ops and native support
 # for TorchScript model, but doesn't contain not-yet-unified caffe2 ops;
 if(INTERN_BUILD_MOBILE)
@@ -1440,6 +1402,9 @@ if(BUILD_JNI)
   add_subdirectory(android/pytorch_android)
 endif()
 
+include(cmake/Summary.cmake)
+caffe2_print_configuration_summary()
+
 # Parse custom debug info
 if(DEFINED USE_CUSTOM_DEBINFO)
   string(REPLACE ";" " " SOURCE_FILES "${USE_CUSTOM_DEBINFO}")
@@ -1479,5 +1444,56 @@ if(BUILD_BUNDLE_PTXAS AND USE_CUDA)
           DESTINATION "${CMAKE_INSTALL_BINDIR}")
 endif()
 
-include(cmake/Summary.cmake)
-caffe2_print_configuration_summary()
+if(USE_PRIORITIZED_TEXT_FOR_LD)
+  add_compile_options(
+    $<$<COMPILE_LANGUAGE:C,CXX>:-ffunction-sections>
+    $<$<COMPILE_LANGUAGE:C,CXX>:-fdata-sections>
+  )
+  set(LINKER_SCRIPT_FILE_OUT "${CMAKE_SOURCE_DIR}/cmake/linker_script.ld")
+  set(LINKER_SCRIPT_FILE_IN "${CMAKE_SOURCE_DIR}/cmake/prioritized_text.txt")
+
+  add_custom_command(
+    OUTPUT "${LINKER_SCRIPT_FILE_OUT}"
+    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py --filein "${LINKER_SCRIPT_FILE_IN}" --fout "${LINKER_SCRIPT_FILE_OUT}"
+    DEPENDS ${CMAKE_SOURCE_DIR}/tools/setup_helpers/generate_linker_script.py "${LINKER_SCRIPT_FILE_IN}"
+    COMMENT "Generating prioritized text linker files"
+    VERBATIM
+  )
+
+  add_custom_target(generate_linker_script DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
+
+  if(BUILD_PYTHON)
+    set(LINKER_OPT_TARGETS torch_python)
+  endif()
+
+  if(NOT BUILD_LIBTORCHLESS)
+    list(APPEND LINKER_OPT_TARGETS torch_cpu c10)
+    if(USE_CUDA)
+      list(APPEND LINKER_OPT_TARGETS torch_cuda c10_cuda)
+    endif()
+    if(USE_XPU)
+      list(APPEND LINKER_OPT_TARGETS torch_xpu c10_xpu)
+    endif()
+    if(USE_ROCM)
+      list(APPEND LINKER_OPT_TARGETS torch_hip c10_hip)
+    endif()
+  endif()
+
+  foreach(tgt IN LISTS LINKER_OPT_TARGETS)
+    if(TARGET ${tgt})
+      add_dependencies("${tgt}" generate_linker_script)
+      target_link_options_if_supported(${tgt} "-T,${LINKER_SCRIPT_FILE_OUT}")
+      set_property(TARGET ${tgt} APPEND PROPERTY LINK_DEPENDS "${LINKER_SCRIPT_FILE_OUT}")
+    else()
+       message(WARNING "Requested target '${tgt}' for linker script optimization was not found.")
+    endif()
+  endforeach()
+
+else()
+  if(LINUX AND CPU_AARCH64)
+    message(WARNING [[
+    It is strongly recommend to enable linker script optimization for all AArch64 Linux builds.
+    To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
+    ]])
+  endif()
+endif()

LICENSE

@@ -37,7 +37,7 @@ Copyright (c) 2024 Tri Dao.
 All rights reserved.
 
 All contributions by Arm:
-Copyright (c) 2021, 2023-2025 Arm Limited and/or its affiliates
+Copyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates
 
 All contributions from Caffe:
 Copyright(c) 2013, 2014, 2015, the respective contributors

SECURITY.md

@@ -18,8 +18,6 @@ Please report security issues using https://github.com/pytorch/pytorch/security/
 
 All reports submitted through the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
 
-**Note on crashes and out of bounds access**: PyTorch is a computational framework that performs operations on behalf of the caller. Like many low-level libraries, PyTorch generally does not validate all inputs to every function—the responsibility for providing valid arguments lies with the calling code. While crashes and out of bounds memory access should be reported as bugs, they are generally not considered security vulnerabilities in PyTorch's threat model.
-
 Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:
 
 https://www.facebook.com/whitehat

aten/src/ATen/CMakeLists.txt

@@ -85,6 +85,7 @@ file(GLOB miopen_h "miopen/*.h")
 file(GLOB miopen_cpp "miopen/*.cpp")
 
 file(GLOB mkl_cpp "mkl/*.cpp")
+file(GLOB mklrng_cpp "mklrng/*.cpp")
 file(GLOB mkldnn_cpp "mkldnn/*.cpp")
 
 file(GLOB mkldnn_xpu_h "native/mkldnn/xpu/*.h" "native/mkldnn/xpu/detail/*.h")
@@ -392,6 +393,7 @@ if(USE_LIGHTWEIGHT_DISPATCH)
 endif()
 if(AT_MKL_ENABLED)
   set(all_cpu_cpp ${all_cpu_cpp} ${mkl_cpp})
+  set(all_cpu_cpp ${all_cpu_cpp} ${mklrng_cpp})
 endif()
 if(AT_KLEIDIAI_ENABLED)
   set(all_cpu_cpp ${all_cpu_cpp} ${native_kleidiai})

aten/src/ATen/CPUGeneratorImpl.cpp

@@ -1,9 +1,14 @@
 #include <ATen/CPUGeneratorImpl.h>
+#include <ATen/Config.h>
 #include <ATen/Utils.h>
 #include <ATen/core/MT19937RNGEngine.h>
 #include <c10/util/MathConstants.h>
 #include <algorithm>
 
+#if AT_MKL_ENABLED()
+#include <ATen/mklrng/MKLGeneratorImpl.h>
+#endif
+
 namespace at {
 
 namespace detail {
@@ -43,6 +48,10 @@ struct CPUGeneratorImplState {
   CPUGeneratorImplStateLegacy legacy_pod;
   float next_float_normal_sample;
   bool is_next_float_normal_sample_valid;
+#if AT_MKL_ENABLED()
+  uint64_t mkl_seed;
+  uint64_t mkl_offset;
+#endif
 };
 
 /**
@@ -82,7 +91,15 @@ CPUGeneratorImpl::CPUGeneratorImpl(uint64_t seed_in)
   : c10::GeneratorImpl{Device(DeviceType::CPU), DispatchKeySet(c10::DispatchKey::CPU)},
     engine_{seed_in},
     next_float_normal_sample_{std::optional<float>()},
-    next_double_normal_sample_{std::optional<double>()} { }
+    next_double_normal_sample_{std::optional<double>()} {
+#if AT_MKL_ENABLED()
+      {
+        auto mkl_gen = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+        std::scoped_lock lock(mkl_gen->mutex_);
+        mkl_gen->set_current_seed(seed_in);
+      }
+#endif
+    }
 
 /**
  * Manually seeds the engine with the seed input
@@ -92,6 +109,13 @@ void CPUGeneratorImpl::set_current_seed(uint64_t seed) {
   next_float_normal_sample_.reset();
   next_double_normal_sample_.reset();
   engine_ = mt19937(seed);
+#if AT_MKL_ENABLED()
+  {
+    auto mkl_gen = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+    std::scoped_lock lock(mkl_gen->mutex_);
+    mkl_gen->set_current_seed(seed);
+  }
+#endif
 }
 
 /**
@@ -126,6 +150,13 @@ uint64_t CPUGeneratorImpl::current_seed() const {
 uint64_t CPUGeneratorImpl::seed() {
   auto random = c10::detail::getNonDeterministicRandom();
   this->set_current_seed(random);
+#if AT_MKL_ENABLED()
+  {
+    auto mkl_gen = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+    std::scoped_lock lock(mkl_gen->mutex_);
+    mkl_gen->set_current_seed(random);
+  }
+#endif
   return random;
 }
 
@@ -169,6 +200,14 @@ void CPUGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
   this->next_double_normal_sample_ = legacy_pod->normal_is_valid
       ? std::optional<double>(legacy_pod->normal_y)
       : std::optional<double>();
+#if AT_MKL_ENABLED()
+  {
+    auto mkl_gen = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+    std::scoped_lock lock(mkl_gen->mutex_);
+    mkl_gen->set_current_seed(rng_state->mkl_seed);
+    mkl_gen->skip_ahead(rng_state->mkl_offset);
+  }
+#endif
 }
 
 /**
@@ -207,6 +246,15 @@ c10::intrusive_ptr<c10::TensorImpl> CPUGeneratorImpl::get_state() const {
     accum_state->next_float_normal_sample = *(this->next_float_normal_sample_);
   }
 
+#if AT_MKL_ENABLED()
+  {
+    auto mkl_gen = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+    std::scoped_lock lock(mkl_gen->mutex_);
+    accum_state->mkl_seed = mkl_gen->current_seed();
+    accum_state->mkl_offset = mkl_gen->get_offset();
+  }
+#endif
+
   memcpy(rng_state, accum_state.get(), size);
   return state_tensor.getIntrusivePtr();
 }

aten/src/ATen/DeviceAccelerator.h

@@ -94,6 +94,11 @@ TORCH_API inline void resetPeakStats(c10::DeviceIndex device_index) {
   at::getDeviceAllocator(device_type)->resetPeakStats(device_index);
 }
 
+TORCH_API inline std::pair<size_t, size_t> getMemoryInfo(
+    c10::DeviceIndex device_index) {
+  const auto device_type = getAccelerator(true).value();
+  return at::getDeviceAllocator(device_type)->getMemoryInfo(device_index);
+}
 } // namespace at::accelerator
 
 namespace at {

aten/src/ATen/core/ivalue.h

@@ -18,8 +18,6 @@
 #include <unordered_set>
 #include <utility>
 
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
-
 namespace torch {
 class TORCH_API CustomClassHolder : public c10::intrusive_ptr_target {};
 namespace jit {
@@ -1632,6 +1630,4 @@ struct TORCH_API WeakOrStrongTypePtr {
 
 } // namespace c10
 
-C10_DIAGNOSTIC_POP()
-
 #include <ATen/core/ivalue_inl.h> // IWYU pragma: keep

aten/src/ATen/core/ivalue_inl.h

@@ -29,8 +29,6 @@
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/irange.h>
 
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
-
 namespace torch {
 namespace jit {
 struct Function;
@@ -2569,5 +2567,3 @@ TypePtr IValue::type() const {
 }
 
 } // namespace c10
-
-C10_DIAGNOSTIC_POP()

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

@@ -11,8 +11,6 @@
 #include <sleef.h>
 #endif
 
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
-
 // Sleef offers vectorized versions of some transcedentals
 // such as sin, cos, tan etc..
 // However for now opting for STL, since we are not building
@@ -652,5 +650,3 @@ inline Vectorized<float> Vectorized<float>::erf() const {
 
 } // namespace CPU_CAPABILITY
 } // namespace at::vec
-
-C10_DIAGNOSTIC_POP()

aten/src/ATen/cuda/CUDAContextLight.h

@@ -3,7 +3,6 @@
 
 #include <cstdint>
 #include <map>
-#include <shared_mutex>
 
 #include <cuda_runtime_api.h>
 #include <cusparse.h>
@@ -89,13 +88,8 @@ TORCH_CUDA_CPP_API cublasHandle_t getCurrentCUDABlasHandle();
 TORCH_CUDA_CPP_API cublasLtHandle_t getCurrentCUDABlasLtHandle();
 
 TORCH_CUDA_CPP_API void clearCublasWorkspaces();
-struct WorkspaceMapWithMutex {
-  std::map<std::tuple<void*, void*>, at::DataPtr> map;
-  std::shared_mutex mutex;
-};
-
-TORCH_CUDA_CPP_API WorkspaceMapWithMutex& cublas_handle_stream_to_workspace();
-TORCH_CUDA_CPP_API WorkspaceMapWithMutex& cublaslt_handle_stream_to_workspace();
+TORCH_CUDA_CPP_API std::map<std::tuple<void *, void *>, at::DataPtr>& cublas_handle_stream_to_workspace();
+TORCH_CUDA_CPP_API std::map<std::tuple<void *, void *>, at::DataPtr>& cublaslt_handle_stream_to_workspace();
 TORCH_CUDA_CPP_API size_t getChosenWorkspaceSize();
 TORCH_CUDA_CPP_API size_t getCUDABlasLtWorkspaceSize();
 TORCH_CUDA_CPP_API void* getCUDABlasLtWorkspace();

aten/src/ATen/cuda/CublasHandlePool.cpp

@@ -107,27 +107,19 @@ using CuBlasPoolType = DeviceThreadHandlePool<cublasHandle_t, createCublasHandle
 
 } // namespace
 
-WorkspaceMapWithMutex& cublas_handle_stream_to_workspace() {
-  static auto& instance = *new WorkspaceMapWithMutex;
+std::map<std::tuple<void *, void *>, at::DataPtr>& cublas_handle_stream_to_workspace() {
+  static auto& instance = *new std::map<std::tuple<void *, void *>, at::DataPtr>;
   return instance;
 }
 
-WorkspaceMapWithMutex& cublaslt_handle_stream_to_workspace() {
-  static auto& instance = *new WorkspaceMapWithMutex;
+std::map<std::tuple<void *, void *>, at::DataPtr>& cublaslt_handle_stream_to_workspace() {
+  static auto& instance = *new std::map<std::tuple<void *, void *>, at::DataPtr>;
   return instance;
 }
 
 void clearCublasWorkspaces() {
-  {
-    auto& workspace = cublas_handle_stream_to_workspace();
-    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
-    workspace.map.clear();
-  }
-  {
-    auto& workspace = cublaslt_handle_stream_to_workspace();
-    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
-    workspace.map.clear();
-  }
+  cublas_handle_stream_to_workspace().clear();
+  cublaslt_handle_stream_to_workspace().clear();
 }
 
 size_t parseChosenWorkspaceSize() {
@@ -249,10 +241,8 @@ void* getCUDABlasLtWorkspace() {
     auto stream = c10::cuda::getCurrentCUDAStream();
     cudaStream_t _stream = stream;
     auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-    auto& workspace = at::cuda::cublas_handle_stream_to_workspace();
-    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
-    auto workspace_it = workspace.map.find(key);
-    TORCH_INTERNAL_ASSERT(workspace_it != workspace.map.end());
+    auto workspace_it = at::cuda::cublas_handle_stream_to_workspace().find(key);
+    TORCH_INTERNAL_ASSERT(workspace_it != at::cuda::cublas_handle_stream_to_workspace().end());
     return workspace_it->second.mutable_get();
   }
 #endif
@@ -260,35 +250,12 @@ void* getCUDABlasLtWorkspace() {
   auto stream = c10::cuda::getCurrentCUDAStream();
   cudaStream_t _stream = stream;
   auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-
-  auto& workspace = cublaslt_handle_stream_to_workspace();
-
-  // Fast path: check if workspace already exists
-  {
-    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
-    auto workspace_it = workspace.map.find(key);
-    if (workspace_it != workspace.map.end()) {
+  auto workspace_it = cublaslt_handle_stream_to_workspace().find(key);
+  if (workspace_it == cublaslt_handle_stream_to_workspace().end()) {
+    workspace_it = cublaslt_handle_stream_to_workspace().insert(workspace_it, {key, getNewCUDABlasLtWorkspace()});
+  }
   return workspace_it->second.mutable_get();
 }
-  }
-
-  // Slow path: allocate workspace outside the lock
-  auto new_workspace = getNewCUDABlasLtWorkspace();
-
-  // Insert with lock (double-check in case another thread inserted while we
-  // were allocating)
-  {
-    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
-    auto workspace_it = workspace.map.find(key);
-    if (workspace_it == workspace.map.end()) {
-      workspace_it =
-          workspace.map.emplace(key, std::move(new_workspace)).first;
-    }
-    // else: another thread inserted it, our new_workspace will be automatically
-    // freed
-    return workspace_it->second.mutable_get();
-  }
-}
 
 cublasHandle_t getCurrentCUDABlasHandle() {
   c10::DeviceIndex device = 0;
@@ -333,39 +300,11 @@ cublasHandle_t getCurrentCUDABlasHandle() {
   // all the memory and cublas's cudaMallocAsync will return OOM
   cudaStream_t _stream = stream;
   auto key = std::make_tuple(static_cast<void *>(handle), static_cast<void *>(_stream));
-
-  auto& workspace = cublas_handle_stream_to_workspace();
-
-  size_t workspace_size = getChosenWorkspaceSize();
-
-  // Fast path: check if workspace already exists
-  {
-    std::shared_lock<std::shared_mutex> lock(workspace.mutex);
-    auto workspace_it = workspace.map.find(key);
-    if (workspace_it != workspace.map.end()) {
-      TORCH_CUDABLAS_CHECK(cublasSetWorkspace(
-          handle, workspace_it->second.get(), workspace_size));
-      return handle;
-    }
-  }
-
-  // Slow path: allocate workspace outside the lock
-  auto new_workspace = getNewWorkspace();
-
-  // Insert with lock (double-check in case another thread inserted while we
-  // were allocating)
-  {
-    std::unique_lock<std::shared_mutex> lock(workspace.mutex);
-    auto workspace_it = workspace.map.find(key);
-    if (workspace_it == workspace.map.end()) {
-      workspace_it =
-          workspace.map.emplace(key, std::move(new_workspace)).first;
-    }
-    // else: another thread inserted it, our new_workspace will be automatically
-    // freed
-    TORCH_CUDABLAS_CHECK(
-        cublasSetWorkspace(handle, workspace_it->second.get(), workspace_size));
+  auto workspace_it = cublas_handle_stream_to_workspace().find(key);
+  if (workspace_it == cublas_handle_stream_to_workspace().end()) {
+    workspace_it = cublas_handle_stream_to_workspace().insert(workspace_it, {key, getNewWorkspace()});
   }
+  TORCH_CUDABLAS_CHECK(cublasSetWorkspace(handle, workspace_it->second.get(), getChosenWorkspaceSize()));
 #if !defined(USE_ROCM)
   // On CUDA >= 11, and architecture >= Ampere, cuBLAS can use TF32 to speedup
   // FP32 data type calculations based on the value of the allow_tf32 flag.

aten/src/ATen/cuda/NumericLimits.cuh

@@ -55,6 +55,14 @@ struct numeric_limits<int8_t> {
   static inline __host__ __device__ int8_t upper_bound() { return INT8_MAX; }
 };
 
+template <>
+struct numeric_limits<uint16_t> {
+  static inline __host__ __device__ uint16_t lowest() { return 0; }
+  static inline __host__ __device__ uint16_t max() { return UINT16_MAX; }
+  static inline __host__ __device__ uint16_t lower_bound() { return 0; }
+  static inline __host__ __device__ uint16_t upper_bound() { return UINT16_MAX; }
+};
+
 template <>
 struct numeric_limits<int16_t> {
   static inline __host__ __device__ int16_t lowest() { return INT16_MIN; }
@@ -63,6 +71,14 @@ struct numeric_limits<int16_t> {
   static inline __host__ __device__ int16_t upper_bound() { return INT16_MAX; }
 };
 
+template <>
+struct numeric_limits<uint32_t> {
+  static inline __host__ __device__ uint32_t lowest() { return 0; }
+  static inline __host__ __device__ uint32_t max() { return UINT32_MAX; }
+  static inline __host__ __device__ uint32_t lower_bound() { return 0; }
+  static inline __host__ __device__ uint32_t upper_bound() { return UINT32_MAX; }
+};
+
 template <>
 struct numeric_limits<int32_t> {
   static inline __host__ __device__ int32_t lowest() { return INT32_MIN; }
@@ -71,6 +87,21 @@ struct numeric_limits<int32_t> {
   static inline __host__ __device__ int32_t upper_bound() { return INT32_MAX; }
 };
 
+template <>
+struct numeric_limits<uint64_t> {
+#ifdef _MSC_VER
+  static inline __host__ __device__ uint64_t lowest() { return 0; }
+  static inline __host__ __device__ uint64_t max() { return _UI64_MAX; }
+  static inline __host__ __device__ uint64_t lower_bound() { return 0; }
+  static inline __host__ __device__ uint64_t upper_bound() { return _UI64_MAX; }
+#else
+  static inline __host__ __device__ uint64_t lowest() { return 0; }
+  static inline __host__ __device__ uint64_t max() { return UINT64_MAX; }
+  static inline __host__ __device__ uint64_t lower_bound() { return 0; }
+  static inline __host__ __device__ uint64_t upper_bound() { return UINT64_MAX; }
+#endif
+};
+
 template <>
 struct numeric_limits<int64_t> {
 #ifdef _MSC_VER

aten/src/ATen/mklrng/MKLGeneratorImpl.cpp

@@ -0,0 +1,155 @@
+#include <ATen/mklrng/MKLGeneratorImpl.h>
+#include <ATen/Utils.h>
+#include <cstdint>
+
+namespace at {
+namespace detail {
+
+/**
+ * PyTorch maintains a collection of default generators that get
+ * initialized once. The purpose of these default generators is to
+ * maintain a global running state of the pseudo random number generation,
+ * when a user does not explicitly mention any generator.
+ * getDefaultMKLGenerator gets the default generator for a particular
+ * device.
+ */
+const Generator& getDefaultMKLGenerator() {
+  static auto gen = createMKLGenerator(c10::detail::getNonDeterministicRandom());
+  return gen;
+}
+
+/**
+ * Utility to create an MKLGeneratorImpl. Returns a shared_ptr
+ */
+Generator createMKLGenerator(uint64_t seed_val) {
+  return make_generator<MKLGeneratorImpl>(seed_val);
+}
+
+} // namespace detail
+
+/**
+ * MKLGeneratorImpl class implementation
+ */
+MKLGeneratorImpl::MKLGeneratorImpl(uint64_t seed_in)
+  : c10::GeneratorImpl{Device(DeviceType::CPU), DispatchKeySet(c10::DispatchKey::CPU)},
+    seed_(seed_in),
+    offset_(0) {
+    vslNewStream(&stream_, VSL_BRNG_PHILOX4X32X10, seed_);
+  }
+
+/**
+ * Manually seeds the engine with the seed input
+ * See Note [Acquire lock when using random generators]
+ */
+void MKLGeneratorImpl::set_current_seed(uint64_t seed) {
+  this->seed_ = seed;
+  vslDeleteStream(&stream_);
+  vslNewStream(&stream_, VSL_BRNG_PHILOX4X32X10, seed_);
+  this->offset_ = 0;
+}
+
+/**
+ * Gets a nondeterministic random number from /dev/urandom or time,
+ * seeds the MKLGeneratorImpl with it and then returns that number.
+ * See Note [Acquire lock when using random generators]
+ */
+uint64_t MKLGeneratorImpl::seed() {
+  auto random = c10::detail::getNonDeterministicRandom();
+  this->set_current_seed(static_cast<uint64_t>(random));
+  return random;
+}
+
+/**
+ * Gets the current seed of CPUGeneratorImpl.
+ */
+uint64_t MKLGeneratorImpl::current_seed() const {
+  return this->seed_;
+}
+
+/**
+ * Gets the DeviceType of MKLGeneratorImpl.
+ * Used for type checking during run time.
+ */
+DeviceType MKLGeneratorImpl::device_type() {
+  return DeviceType::CPU;
+}
+
+/**
+ * Gets the copy of VSLStreamStatePtr in MKLGenerator
+ * to be used for variate generation in a thread-safe way
+ * (each thread should receive its own stream copy).
+ * See Note [Acquire lock when using random generators]
+ */
+void MKLGeneratorImpl::get_stream_copy(VSLStreamStatePtr &streamCopy) {
+  vslCopyStream(&streamCopy, stream_);
+}
+
+/**
+ * Progresses the internal PRNG state n steps ahead --
+ * used to account for variates generated by the copies
+ * of the stream in MKLGenerator.
+ * See Note [Acquire lock when using random generators]
+ */
+void MKLGeneratorImpl::skip_ahead(uint64_t n) {
+  vslSkipAheadStream(stream_, n);
+  this->advance_offset(n);
+}
+
+/**
+ * Private clone method implementation
+ * See Note [Acquire lock when using random generators]
+ */
+MKLGeneratorImpl* MKLGeneratorImpl::clone_impl() const {
+  auto gen = new MKLGeneratorImpl();
+  return gen;
+}
+
+/**
+ * Public clone method implementation
+ * See Note [Acquire lock when using random generators]
+ */
+std::shared_ptr<MKLGeneratorImpl> MKLGeneratorImpl::clone() const {
+  return std::shared_ptr<MKLGeneratorImpl>(this->clone_impl());
+}
+
+/**
+ * Sets the offset of RNG state.
+ * See Note [Acquire lock when using random generators]
+ */
+void MKLGeneratorImpl::set_offset(uint64_t offset) {
+  TORCH_CHECK(false, "MKL Generator does not allow to set offset");
+}
+
+/**
+ * Gets the offset of RNG state.
+ * See Note [Acquire lock when using random generators]
+ */
+uint64_t MKLGeneratorImpl::get_offset() const {
+  return this->offset_;
+}
+
+/**
+ * Private method to advance the offset of RNG state.
+ * See Note [Acquire lock when using random generators]
+ */
+void MKLGeneratorImpl::advance_offset(uint64_t n) {
+  this->offset_ += n;
+}
+
+/**
+ * Gets the current internal state of MKLGeneratorImpl. The internal
+ * state is returned as a CPU byte tensor.
+ */
+c10::intrusive_ptr<c10::TensorImpl> MKLGeneratorImpl::get_state() const {
+  TORCH_CHECK(false, "MKL Generator does not use get_state");
+}
+
+/**
+ * Sets the internal state of MKLGeneratorImpl. The new internal state
+ * must be a strided CPU byte tensor.
+ */
+ void MKLGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
+  TORCH_CHECK(false, "MKL Generator does not use set_state");
+}
+
+} // namespace at

aten/src/ATen/mklrng/MKLGeneratorImpl.h

@@ -0,0 +1,46 @@
+#pragma once
+
+#include <ATen/Config.h>
+#include <ATen/core/Generator.h>
+#include <c10/core/GeneratorImpl.h>
+#include <cstdint>
+
+#include <mkl.h>
+
+namespace at {
+
+struct TORCH_API MKLGeneratorImpl : public c10::GeneratorImpl {
+  // Constructors
+  MKLGeneratorImpl(uint64_t seed_in = default_rng_seed_val);
+  ~MKLGeneratorImpl() override = default;
+
+  // MKLGeneratorImpl methods
+  std::shared_ptr<MKLGeneratorImpl> clone() const;
+  void set_current_seed(uint64_t seed) override;
+  uint64_t seed() override;
+  uint64_t current_seed() const override;
+  void set_offset(uint64_t offset) override;
+  uint64_t get_offset() const override;
+  static c10::DeviceType device_type();
+  void get_stream_copy(VSLStreamStatePtr &streamCopy);
+  void skip_ahead(uint64_t n);
+  void set_state(const c10::TensorImpl& new_state) override;
+  c10::intrusive_ptr<c10::TensorImpl> get_state() const override;
+
+ private:
+  MKLGeneratorImpl* clone_impl() const override;
+  void advance_offset(uint64_t n);
+  VSLStreamStatePtr stream_;
+  uint64_t seed_;
+  uint64_t offset_;
+};
+
+namespace detail {
+
+TORCH_API const Generator& getDefaultMKLGenerator();
+TORCH_API Generator
+createMKLGenerator(uint64_t seed_val = default_rng_seed_val);
+
+} // namespace detail
+
+} // namespace at

aten/src/ATen/mps/MPSAllocator.mm

@@ -440,7 +440,7 @@ bool MPSHeapAllocatorImpl::release_cached_buffers() {
   // we need to release the lock temporarily as synchronizing may cause deadlock with completion handlers.
   m_mutex.unlock();
   auto stream = getDefaultMPSStream();
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
+  dispatch_sync(stream->queue(), ^() {
     stream->synchronize(SyncType::COMMIT_AND_WAIT);
   });
   m_mutex.lock();

aten/src/ATen/mps/MPSStream.h

@@ -110,9 +110,6 @@ class TORCH_API MPSStream {
     return _stream;
   }
 
-  MTLBuffer_t getErrorBuffer();
-  void checkLastError();
-
  private:
   Stream _stream;
   MTLCommandQueue_t _commandQueue = nil;
@@ -124,8 +121,6 @@ class TORCH_API MPSStream {
   dispatch_queue_t _serialQueue = nullptr;
   // CommitAndContinue is enabled by default
   bool _enableCommitAndContinue = true;
-  // Buffer that contains last raised error
-  MTLBuffer_t _errorBuffer = nil;
 
   // use synchronize() to access any of these commit functions outside MPSStream
   void commit();
@@ -160,7 +155,4 @@ class TORCH_API MPSStreamImpl {
   MPSStreamImpl();
 };
 
-#ifdef __OBJC__
-void dispatch_sync_with_rethrow(dispatch_queue_t queue, void (^block)());
-#endif
 } // namespace at::mps

aten/src/ATen/mps/MPSStream.mm

@@ -3,13 +3,13 @@
 #include <ATen/mps/MPSAllocatorInterface.h>
 #include <ATen/mps/MPSProfiler.h>
 #include <ATen/mps/MPSStream.h>
-#include <c10/metal/error.h>
 
 @interface MPSGraphExecutionDescriptor ()
 @property(readwrite, atomic) BOOL enableCommitAndContinue;
 @end
 
 namespace at::mps {
+
 //-----------------------------------------------------------------
 //  MPSStream
 //-----------------------------------------------------------------
@@ -30,10 +30,6 @@ MPSStream::MPSStream(Stream stream) : _stream(stream) {
   // Choose level which optimizes for GPU
   _compilationDescriptor.optimizationLevel = MPSGraphOptimizationLevel0;
   _executionDescriptor.compilationDescriptor = _compilationDescriptor;
-
-  _errorBuffer = [MPSDevice::getInstance()->device() newBufferWithLength:sizeof(c10::metal::ErrorMessages)
-                                                                 options:MTLResourceStorageModeShared];
-  std::memset([_errorBuffer contents], 0, 1024);
 }
 
 MPSStream::~MPSStream() {
@@ -42,8 +38,6 @@ MPSStream::~MPSStream() {
   [_executionDescriptor release];
   [_compilationDescriptor release];
   _executionDescriptor = nil;
-  [_errorBuffer release];
-  _errorBuffer = nil;
   _compilationDescriptor = nil;
 
   assert(_commandBuffer == nil);
@@ -110,7 +104,6 @@ void MPSStream::commitAndWait() {
     [_prevCommandBuffer waitUntilCompleted];
     [_prevCommandBuffer release];
     _prevCommandBuffer = nil;
-    checkLastError();
   }
 
   if (_commandBuffer) {
@@ -118,7 +111,6 @@ void MPSStream::commitAndWait() {
     [_commandBuffer waitUntilCompleted];
     [_commandBuffer release];
     _commandBuffer = nil;
-    checkLastError();
   }
 }
 
@@ -161,7 +153,7 @@ void MPSStream::fill(id<MTLBuffer> buffer, uint8_t value, size_t length, size_t
   if (length == 0) {
     return;
   }
-  dispatch_sync_with_rethrow(_serialQueue, ^() {
+  dispatch_sync(_serialQueue, ^() {
     @autoreleasepool {
       endKernelCoalescing();
       id<MTLBlitCommandEncoder> blitEncoder = [commandBuffer() blitCommandEncoder];
@@ -191,7 +183,7 @@ void MPSStream::copy(id<MTLBuffer> srcBuffer,
                      size_t dstOffset,
                      uint64_t profileId,
                      SyncType syncType) {
-  dispatch_sync_with_rethrow(_serialQueue, ^() {
+  dispatch_sync(_serialQueue, ^() {
     @autoreleasepool {
       endKernelCoalescing();
       id<MTLBlitCommandEncoder> blitEncoder = [commandBuffer() blitCommandEncoder];
@@ -244,7 +236,7 @@ void MPSStream::executeMPSGraph(MPSGraph* mpsGraph, NSDictionary* feeds, NSDicti
   auto& profiler = getMPSProfiler();
   const bool isGraphProfilingEnabled = profiler.isOperationProfilingEnabled();
 
-  dispatch_sync_with_rethrow(_serialQueue, ^() {
+  dispatch_sync(_serialQueue, ^() {
     endKernelCoalescing();
     if (isGraphProfilingEnabled) {
       // this function call is only relevant for interval-based Signposts
@@ -274,24 +266,6 @@ void MPSStream::executeMPSGraph(MPSGraph* mpsGraph, NSDictionary* feeds, NSDicti
   });
 }
 
-id<MTLBuffer> MPSStream::getErrorBuffer() {
-  return _errorBuffer;
-}
-
-void MPSStream::checkLastError() {
-  auto msgs = reinterpret_cast<c10::metal::ErrorMessages*>([_errorBuffer contents]);
-  const auto& msg = msgs->msg[0];
-  if (!msgs) {
-    return;
-  }
-  unsigned int count = 0;
-  std::swap(count, msgs->count);
-  if (!count) {
-    return;
-  }
-  throw c10::AcceleratorError({msg.func, msg.file, msg.line}, 1, msg.message);
-}
-
 //-----------------------------------------------------------------
 //  MPSStreamImpl
 //-----------------------------------------------------------------
@@ -315,19 +289,4 @@ MPSStream* getDefaultMPSStream() {
   return MPSStreamImpl::getInstance();
 }
 
-// Helper methods
-void dispatch_sync_with_rethrow(dispatch_queue_t queue, void (^block)()) {
-  __block std::optional<std::exception_ptr> block_exception;
-  dispatch_sync(queue, ^() {
-    try {
-      block();
-    } catch (...) {
-      block_exception = std::current_exception();
-    }
-  });
-  if (block_exception) {
-    std::rethrow_exception(*block_exception);
-  }
-}
-
 } // namespace at::mps

aten/src/ATen/native/LinearAlgebra.cpp

@@ -1936,7 +1936,7 @@ static bool should_fold(const Tensor& tensor1, const Tensor& tensor2, bool has_o
 
   // We order the tensors. t1 will be the larger tensor
   // We can always transpose tensor2 as the dimensions are always >= 1 (precondition from matmul)
-  // and tensor1_larger iff tensor2.dim() > tensor1.dim()
+  // and tensor1_larger iff tensor2.dim() > tensor1.dim(9
   const auto t1 = tensor1_larger ? MaybeOwned<Tensor>::borrowed(tensor1)
                                  : MaybeOwned<Tensor>::owned(tensor2.mT());
   const int64_t dim_t1 = t1->dim();
@@ -1948,11 +1948,20 @@ static bool should_fold(const Tensor& tensor1, const Tensor& tensor2, bool has_o
     return false;
   }
 
-  // If we require a gradient, we should fold to minimize backward memory usage - even if this
-  // leads to a copy in forward because is needed in backward,
-  // only time we avoid this strict pre-allocated memory usage (has_out = True)
-  bool requires_grad = tensor1.requires_grad() || tensor2.requires_grad();
-  if (requires_grad && !has_out) {
+  // In this case we *do* incur in an extra copy to avoid creating an unnecessary large tensor in the backward
+  // Suppose we don't fold here. Let t1.shape = [b, m, n] t2.shape = [n, k] like in a transformer
+  // t2 will be expanded to a tensor of shape [b, n, k] and then we do t1.bmm(t2_expanded)
+  // The issue appears in the backward.
+  // The output gradient g of this operation would have shape [b, m, k]
+  // The backward wrt. t2 of bmm would be given by t1.mH @ g, which has shape [b, n, k]
+  // Then, the backward of expand is simply `sum(0)`. As such, we are instantiating a tensor
+  // of shape [b, n, k] unnecessarily, which may cause a large memory footprint, and in the
+  // worst case, an OOM
+  bool t2_requires_grad = tensor1_larger ? tensor2.requires_grad() : tensor1.requires_grad();
+  if (t2_requires_grad && !has_out) {
+    // We should be checking !at::GradMode::is_enabled(), but apparently
+    // this regresses performance in some cases:
+    // https://github.com/pytorch/pytorch/issues/118548#issuecomment-1916022394
     return true;
   }
 

aten/src/ATen/native/PackedSequence.cpp

@@ -142,7 +142,6 @@ Tensor _pack_padded_sequence_backward_symint(const Tensor& grad, c10::SymIntArra
 std::tuple<Tensor, Tensor> _pad_packed_sequence(const Tensor& data, const Tensor& _batch_sizes, bool batch_first, const Scalar& padding_value, int64_t total_length) {
   auto batch_sizes_t = _batch_sizes.contiguous();
   checkLongTensor(batch_sizes_t);
-  TORCH_CHECK(batch_sizes_t.numel() > 0, "batch_sizes can not be empty");
 
   int64_t * batch_sizes = batch_sizes_t.data_ptr<int64_t>();
   int64_t max_batch_size = batch_sizes[0];

aten/src/ATen/native/TensorCompare.cpp

@@ -23,7 +23,6 @@
 #include <ATen/ops/_aminmax_native.h>
 #include <ATen/ops/_assert_async_native.h>
 #include <ATen/ops/_assert_scalar_native.h>
-#include <ATen/ops/_async_error_native.h>
 #include <ATen/ops/_functional_assert_async_native.h>
 #include <ATen/ops/_functional_assert_scalar_native.h>
 #include <ATen/ops/_make_per_tensor_quantized_tensor.h>
@@ -480,14 +479,6 @@ Tensor isfinite(const Tensor& self) {
   });
 }
 
-void _async_error(std::string_view msg) {
-  TORCH_CHECK(0, msg);
-}
-
-void _async_error_meta(std::string_view msg) {
-  // Do NOT error, it's an async error!
-}
-
 void _assert_async_cpu(const Tensor& self) {
   TORCH_CHECK(
       native::is_nonzero(self),

aten/src/ATen/native/TransposeType.h

@@ -1,8 +1,6 @@
 #pragma once
 #include <c10/util/Exception.h>
 
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
-
 namespace at::native {
 
 // Used as an interface between the different BLAS-like libraries
@@ -23,5 +21,3 @@ static inline char to_blas(TransposeType trans) {
 }
 
 }  // namespace at::native
-
-C10_DIAGNOSTIC_POP()

aten/src/ATen/native/cpu/DistributionKernels.cpp

@@ -18,9 +18,9 @@
 #include <limits>
 #include <type_traits>
 
-// Disable MKL rng until https://github.com/pytorch/pytorch/issues/132395 is addressed
-#if AT_MKL_ENABLED() && defined(FBCODE_CAFFE2)
+#if AT_MKL_ENABLED()
 #include <mkl.h>
+#include <ATen/mklrng/MKLGeneratorImpl.h>
 #include <cpuinfo.h>
 #endif
 
@@ -37,8 +37,7 @@ void bernoulli_tensor_kernel(const TensorBase &self, const TensorBase &p_, std::
   templates::cpu::bernoulli_kernel(self, p_, generator);
 }
 
-// Disable MKL rng until https://github.com/pytorch/pytorch/issues/132395 is addressed
-#if !AT_MKL_ENABLED() || (AT_MKL_ENABLED() && !defined(FBCODE_CAFFE2))
+#if !AT_MKL_ENABLED()
 void bernoulli_scalar_kernel_default(const TensorBase &self, double p, std::optional<Generator> gen) {
   CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
   templates::cpu::bernoulli_kernel(self, p, generator);
@@ -49,13 +48,6 @@ void bernoulli_scalar_kernel(const TensorBase &self, double p, std::optional<Gen
 }
 #else
 void bernoulli_scalar_kernel(const TensorBase &self, double p, std::optional<Generator> gen) {
-  CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
-  int64_t seed;
-  {
-    // See Note [Acquire lock when using random generators]
-    std::lock_guard<std::mutex> lock(generator->mutex_);
-    seed = generator->random();
-  }
   int64_t n = self.numel();
   bool contig = self.is_contiguous();
 
@@ -71,15 +63,28 @@ void bernoulli_scalar_kernel(const TensorBase &self, double p, std::optional<Gen
     scalar_t *self_ptr = self.data_ptr<scalar_t>();
     int *sample_int_ptr = tmp_int_tensor.data_ptr<int>();
 
+    auto mklGenerator = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+    VSLStreamStatePtr main_stream;
+
+    // Get a local copy of the global stream and immediately advance the global
+    // state before the generation step to avoid multiple threads using the same state.
+    {
+      // See Note [Acquire lock when using random generators]
+      std::lock_guard<std::mutex> lock(mklGenerator->mutex_);
+      mklGenerator->get_stream_copy(main_stream);
+      mklGenerator->skip_ahead(n);
+    }
+
     auto sample = [&](int64_t begin, int64_t end) {
       int64_t len = end - begin;
       if (len > 0) {
-        VSLStreamStatePtr stream;
-        vslNewStream(&stream, VSL_BRNG_MCG31, seed);
-        vslSkipAheadStream(stream, begin);
-        viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, stream, len,
+        VSLStreamStatePtr sample_stream;
+        vslCopyStream(&sample_stream, main_stream);
+        vslSkipAheadStream(sample_stream, begin);
+
+        viRngBernoulli(VSL_RNG_METHOD_BERNOULLI_ICDF, sample_stream, len,
           sample_int_ptr + begin, p);
-        vslDeleteStream(&stream);
+        vslDeleteStream(&sample_stream);
 
         // vectorized copy if using buffer and contiguous, i.e., being non-int
         // type and contiguous
@@ -92,6 +97,7 @@ void bernoulli_scalar_kernel(const TensorBase &self, double p, std::optional<Gen
     };
 
     parallel_for(0, n, /* grain_size= */ 800, sample);
+    vslDeleteStream(&main_stream);
 
     // copy_ if using buffer and non contiguous
     if (!contig) {
@@ -106,27 +112,15 @@ void exponential_kernel_default(TensorIteratorBase& iter, double lambda, std::op
   templates::cpu::exponential_kernel(iter, lambda, generator);
 }
 
-// Disable MKL rng until https://github.com/pytorch/pytorch/issues/132395 is addressed
-#if (!AT_MKL_ENABLED() || defined(FBCODE_CAFFE2) || 1)
+#if (!AT_MKL_ENABLED() || defined(FBCODE_CAFFE2))
 void exponential_kernel(TensorIteratorBase& iter, double lambda, std::optional<Generator> gen) {
   exponential_kernel_default(iter, lambda, gen);
 }
 #else
 void exponential_kernel(TensorIteratorBase &iter, double lambda, std::optional<Generator> gen) {
   TORCH_CHECK(isFloatingType(iter.dtype()), "Exponential distribution is a continuous probability distribution. dtype must be a floating point but you specified ", iter.dtype());
-
   Tensor self = iter.tensor(0);
   if (lambda > 0 && !std::isinf(lambda) && !std::isnan(lambda)) {
-    CPUGeneratorImpl* generator = get_generator_or_default<CPUGeneratorImpl>(gen, detail::getDefaultCPUGenerator());
-    int64_t seed;
-    {
-      // See Note [Acquire lock when using random generators]
-      std::lock_guard<std::mutex> lock(generator->mutex_);
-      if (self.scalar_type() == at::kDouble)
-        seed = generator->random64();
-      else
-        seed = generator->random();
-    }
     int64_t n = self.numel();
     bool contig = self.is_contiguous();
 
@@ -158,23 +152,35 @@ void exponential_kernel(TensorIteratorBase &iter, double lambda, std::optional<G
       // Variance:    V[X+eps] = 1/lambda**2
       auto eps = std::numeric_limits<tmp_scalar_t>::min();
 
+      auto mklGenerator = check_generator<MKLGeneratorImpl>(detail::getDefaultMKLGenerator());
+      VSLStreamStatePtr main_stream;
+
+      // Get a local copy of the global stream and immediately advance the global
+      // state before the generation step to avoid multiple threads using the same state.
+      {
+        // See Note [Acquire lock when using random generators]
+        std::lock_guard<std::mutex> lock(mklGenerator->mutex_);
+        mklGenerator->get_stream_copy(main_stream);
+        mklGenerator->skip_ahead(n);
+      }
+
       auto sample = [&](int64_t begin, int64_t end) {
         int64_t len = end - begin;
         if (len > 0) {
-          VSLStreamStatePtr stream;
+          VSLStreamStatePtr sample_stream;
+          vslCopyStream(&sample_stream, main_stream);
+          vslSkipAheadStream(sample_stream, begin);
+
           if constexpr (std::is_same_v<scalar_t, double>) {
-            vslNewStream(&stream, VSL_BRNG_MCG31, seed);
-            vslSkipAheadStream(stream, begin);
-            vdRngExponential(VSL_RNG_METHOD_EXPONENTIAL_ICDF, stream, len,
+            vdRngExponential(VSL_RNG_METHOD_EXPONENTIAL_ICDF, sample_stream, len,
               (double *)(sample_ptr + begin), eps, 1./lambda);
-            vslDeleteStream(&stream);
+            vslDeleteStream(&sample_stream);
           } else {
-            vslNewStream(&stream, VSL_BRNG_MCG31, seed);
-            vslSkipAheadStream(stream, begin);
-            vsRngExponential(VSL_RNG_METHOD_EXPONENTIAL_ICDF, stream, len,
+            vsRngExponential(VSL_RNG_METHOD_EXPONENTIAL_ICDF, sample_stream, len,
               (float *) (sample_ptr + begin), eps, 1./lambda);
-            vslDeleteStream(&stream);
+            vslDeleteStream(&sample_stream);
           }
+
           // vectorized copy if using buffer and contiguous
           if (!is_df && contig) {
             scalar_t *self_seg = self_ptr + begin;
@@ -185,6 +191,7 @@ void exponential_kernel(TensorIteratorBase &iter, double lambda, std::optional<G
       };
 
       parallel_for(0, n, /* grain_size= */ 800, sample);
+      vslDeleteStream(&main_stream);
 
       // copy_ if using buffer and non contiguous
       if (!contig) {

aten/src/ATen/native/cpu/ReduceAllOpsKernel.cpp

@@ -5,6 +5,7 @@
 #include <ATen/native/ReduceOpsUtils.h>
 
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/Parallel.h>
 #include <ATen/TensorIterator.h>
 #include <ATen/OpMathType.h>
@@ -78,12 +79,12 @@ void min_all_kernel_impl(Tensor& result, const Tensor& input) {
     reduce_all_impl<int64_t>(result, input, upper_bound<int64_t>(),
       [=](int64_t a, int64_t b) -> int64_t { return min_impl(a, b); });
   } else {
-    AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, input.scalar_type(), "min_all", [&] {
+    AT_DISPATCH_V2(input.scalar_type(), "min_all", AT_WRAP([&] {
       using Vec = Vectorized<opmath_type<scalar_t>>;
       reduce_all_impl_vec<scalar_t>(result, input, upper_bound<scalar_t>(),
         [=] (scalar_t a , scalar_t b) -> scalar_t { return min_impl(a, b); },
         [=](Vec a, Vec b) -> Vec { return minimum(a, b); });
-    });
+    }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
   }
 }
 
@@ -103,12 +104,12 @@ void max_all_kernel_impl(Tensor& result, const Tensor& input) {
     reduce_all_impl<int64_t>(result, input, lower_bound<int64_t>(),
       [=](int64_t a, int64_t b) -> int64_t { return max_impl(a, b); });
   } else {
-    AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, input.scalar_type(), "max_all", [&] {
+    AT_DISPATCH_V2(input.scalar_type(), "max_all", AT_WRAP([&] {
       using Vec = Vectorized<opmath_type<scalar_t>>;
       reduce_all_impl_vec<scalar_t>(result, input, lower_bound<scalar_t>(),
         [=] (scalar_t a , scalar_t b) -> scalar_t { return max_impl(a, b); },
         [=](Vec a, Vec b) -> Vec { return maximum(a, b); });
-    });
+    }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
   }
 }
 
@@ -199,7 +200,7 @@ void aminmax_allreduce_kernel(
       }
     );
   } else {
-    AT_DISPATCH_ALL_TYPES_AND2(kBFloat16, kHalf, input.scalar_type(), "aminmax_cpu", [&] {
+    AT_DISPATCH_V2(input.scalar_type(), "aminmax_cpu", AT_WRAP([&] {
       using Vec = Vectorized<opmath_type<scalar_t>>;
       using scalar_t_pair = std::pair<scalar_t, scalar_t>;
       reduce_all_impl_vec_two_outputs<scalar_t>(
@@ -214,7 +215,7 @@ void aminmax_allreduce_kernel(
         [=](Vec a, Vec b) -> Vec { return minimum(a, b); },
         [=](Vec a, Vec b) -> Vec { return maximum(a, b); }
       );
-    });
+    }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf);
   }
 }
 

aten/src/ATen/native/cpu/ReduceOpsKernel.cpp

@@ -3,6 +3,7 @@
 
 #include <ATen/core/Tensor.h>
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/OpMathType.h>
 #include <ATen/cpu/vec/vec.h>
 #include <ATen/cpu/vec/functional.h>
@@ -347,34 +348,35 @@ struct MinValuesOps: public at::native::MinOps<scalar_t> {
 };
 
 void min_values_kernel_impl(TensorIterator& iter) {
-  if (iter.dtype() == kLong) {
   // This case is special because of Vectorized<int64_t> does not
   // handle upper_bound<int64_t>().
   // See: https://github.com/pytorch/pytorch/issues/43254
-    using scalar_t = int64_t;
+  if (iter.dtype() == kLong || iter.dtype() == kUInt64) {
+    AT_DISPATCH_V2(iter.dtype(), "min_values_cpu", AT_WRAP([&iter] {
       binary_kernel_reduce(
         iter,
         MinValuesOps<scalar_t>{},
         std::pair<scalar_t, int64_t>(upper_bound<scalar_t>(), -1));
+    }), kLong, kUInt64);
     return;
   }
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.dtype(), "min_values_cpu", [&iter] {
+  AT_DISPATCH_V2(iter.dtype(), "min_values_cpu", AT_WRAP([&iter] {
     binary_kernel_reduce_vec(
       iter,
       [](scalar_t a, scalar_t b) -> scalar_t { return min_impl(a, b); },
       [](Vectorized<scalar_t> a, Vectorized<scalar_t> b) { return minimum(a, b); },
       static_cast<double>(upper_bound<scalar_t>()));
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void max_values_kernel_impl(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.dtype(), "max_values_cpu", [&iter] {
+  AT_DISPATCH_V2(iter.dtype(), "max_values_cpu", AT_WRAP([&iter] {
     binary_kernel_reduce_vec(
       iter,
       [](scalar_t a, scalar_t b) -> scalar_t { return max_impl(a, b); },
       [](Vectorized<scalar_t> a, Vectorized<scalar_t> b) { return maximum(a, b); },
       lower_bound<scalar_t>());
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void argmax_kernel_impl(TensorIterator &iter) {

aten/src/ATen/native/cpu/TensorCompareKernel.cpp

@@ -11,6 +11,7 @@
 #include <vector>
 
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/Parallel.h>
 #include <ATen/NumericUtils.h>
 #include <ATen/TensorIterator.h>
@@ -106,7 +107,7 @@ void min_kernel_impl(
     bool keepdim) {
   int64_t self_dim_size = ensure_nonempty_size(self, dim);
 
-  AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool, self.scalar_type(), "min_cpu", [&] {
+  AT_DISPATCH_V2(self.scalar_type(), "min_cpu", AT_WRAP([&] {
     compare_base_kernel<scalar_t>(result, indice, self, dim, keepdim, [&] (
       scalar_t* result_data, int64_t* indice_data,
       const scalar_t* self_data, auto self_dim_stride) {
@@ -128,7 +129,7 @@ void min_kernel_impl(
         *indice_data = index;
       }
     );
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool);
 }
 
 void max_kernel_impl(
@@ -139,7 +140,7 @@ void max_kernel_impl(
     bool keepdim) {
   int64_t self_dim_size = ensure_nonempty_size(self, dim);
 
-  AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool, self.scalar_type(), "max_cpu", [&] {
+  AT_DISPATCH_V2(self.scalar_type(), "max_cpu", AT_WRAP([&] {
     compare_base_kernel<scalar_t>(result, indice, self, dim, keepdim, [&] (
       scalar_t* result_data, int64_t* indice_data,
       const scalar_t* self_data, auto self_dim_stride) {
@@ -161,7 +162,7 @@ void max_kernel_impl(
         *indice_data = index;
       }
     );
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool);
 }
 
 void aminmax_kernel(
@@ -186,7 +187,7 @@ void aminmax_kernel(
     return;
   }
 
-  AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Bool, ScalarType::BFloat16, ScalarType::Half, self.scalar_type(), "aminmax_cpu", [&] {
+  AT_DISPATCH_V2(self.scalar_type(), "aminmax_cpu", AT_WRAP([&] {
     compare_base_kernel<scalar_t, scalar_t>(min_result, max_result, self, wrap_dim, keepdim, [&] (
       scalar_t* min_result_data, scalar_t* max_result_data,
       const scalar_t* self_data, auto self_dim_stride) {
@@ -209,7 +210,7 @@ void aminmax_kernel(
         *max_result_data = max_number;
       }
     );
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), ScalarType::Bool, ScalarType::BFloat16, ScalarType::Half);
 }
 
 void where_kernel_impl(TensorIterator &iter) {

aten/src/ATen/native/cuda/GroupedBlas.cpp

@@ -669,12 +669,9 @@ std::optional<c10::ScalarType> out_dtype) {
   // _scaled_mm_allowed_device is used here within _grouped_mm_cuda which seems incorrect since scale is not used.
   // the _grouped_mm_fallback should be safe for any ROCm GPU since it's just calling typical mm/bmm
   bool use_fast_path = false;
-  // On non CK system(w/ ROCm), make sure use_fast_path is false
-#if defined(USE_ROCM_CK_GEMM)
   if (at::detail::getCUDAHooks().isGPUArch({"gfx942", "gfx950"})) {
     use_fast_path = true;
   }
-#endif //USE_ROCM_CK_GEMM
 #endif
   const auto out_dtype_ = _resolve_grouped_mm_out_dtype(mat_a, mat_b, out_dtype);
   Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
@@ -683,11 +680,7 @@ std::optional<c10::ScalarType> out_dtype) {
 #ifndef USE_ROCM
     at::cuda::detail::bf16bf16_grouped_mm(mat_a, mat_b, offs, bias, out);
 #else
-#if defined(USE_ROCM_CK_GEMM)
     at::hip::detail::group_gemm_ck(mat_a, mat_b, offs, bias, out);
-#else
-    TORCH_WARN("ROCm: Group Gemm through CK not selected.");
-#endif //USE_ROCM_CK_GEMM
 #endif
   } else {
     _grouped_mm_fallback(mat_a, mat_b, offs, bias, out_dtype, out);

aten/src/ATen/native/cuda/ReduceAMinMaxKernel.cu

@@ -1,5 +1,6 @@
 #define TORCH_ASSERT_NO_OPERATORS
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/NumericUtils.h>
 #include <ATen/native/DispatchStub.h>
 #include <ATen/native/ReduceAllOps.h>
@@ -28,22 +29,22 @@ void _min_max_values_kernel_cuda_impl(TensorIterator& iter) {
 }
 
 void aminmax_allreduce_launch_kernel(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(
-      kBFloat16, kHalf, kBool, iter.input_dtype(), "aminmax_all_cuda", [&] {
+  AT_DISPATCH_V2(
+      iter.input_dtype(), "aminmax_all_cuda", AT_WRAP([&] {
         _min_max_values_kernel_cuda_impl<scalar_t>(iter);
-      });
+      }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void aminmax_launch_kernel(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(
-      kBFloat16, kHalf, kBool, iter.input_dtype(), "aminmax_cuda", [&]() {
+  AT_DISPATCH_V2(
+      iter.input_dtype(), "aminmax_cuda", AT_WRAP([&]() {
         gpu_reduce_kernel<scalar_t, scalar_t>(
             iter,
             MinMaxOps<scalar_t, scalar_t, int32_t>{},
             thrust::pair<scalar_t, scalar_t>(
                 at::numeric_limits<scalar_t>::upper_bound(),
                 at::numeric_limits<scalar_t>::lower_bound()));
-      });
+      }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 } // namespace at::native

aten/src/ATen/native/cuda/ReduceMaxValuesKernel.cu

@@ -1,5 +1,6 @@
 #define TORCH_ASSERT_NO_OPERATORS
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/NumericUtils.h>
 #include <ATen/native/DispatchStub.h>
 #include <ATen/native/ReduceAllOps.h>
@@ -33,27 +34,27 @@ void max_values_kernel_cuda_impl(TensorIterator& iter) {
 }
 
 void max_values_kernel_cuda(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(
-      kBFloat16, kHalf, kBool, iter.dtype(), "max_values_cuda", [&]() {
+  AT_DISPATCH_V2(
+      iter.dtype(), "max_values_cuda", AT_WRAP([&]() {
         max_values_kernel_cuda_impl<scalar_t>(iter);
-      });
+      }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void max_launch_kernel(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(
-      kBFloat16, kHalf, kBool, iter.input_dtype(), "max_cuda", [&]() {
+  AT_DISPATCH_V2(
+      iter.input_dtype(), "max_cuda", AT_WRAP([&]() {
         gpu_reduce_kernel<scalar_t, scalar_t>(
             iter,
             MaxOps<scalar_t>{},
             thrust::pair<scalar_t, int64_t>(
                 at::numeric_limits<scalar_t>::lower_bound(), 0));
-      });
+      }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void max_all_launch_kernel(TensorIterator &iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.input_dtype(), "max_all_cuda", [&] {
+  AT_DISPATCH_V2(iter.input_dtype(), "max_all_cuda", AT_WRAP([&] {
     max_values_kernel_cuda_impl<scalar_t>(iter);
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 REGISTER_DISPATCH(max_values_stub, &max_values_kernel_cuda)

aten/src/ATen/native/cuda/ReduceMinValuesKernel.cu

@@ -12,6 +12,7 @@
 #include <ATen/NumericUtils.h>
 
 #include <ATen/Dispatch.h>
+#include <ATen/Dispatch_v2.h>
 #include <ATen/NumericUtils.h>
 #include <ATen/cuda/NumericLimits.cuh>
 
@@ -33,24 +34,24 @@ void min_values_kernel_cuda_impl(TensorIterator& iter) {
 }
 
 void min_values_kernel_cuda(TensorIterator& iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.dtype(), "min_values_cuda", [&]() {
+  AT_DISPATCH_V2(iter.dtype(), "min_values_cuda", AT_WRAP([&]() {
     min_values_kernel_cuda_impl<scalar_t>(iter);
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void min_launch_kernel(TensorIterator &iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.input_dtype(), "min_cuda", [&]() {
+  AT_DISPATCH_V2(iter.input_dtype(), "min_cuda", AT_WRAP([&]() {
     gpu_reduce_kernel<scalar_t, scalar_t>(
       iter,
       MinOps<scalar_t>{},
       thrust::pair<scalar_t, int64_t>(at::numeric_limits<scalar_t>::upper_bound(), 0));
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 void min_all_launch_kernel(TensorIterator &iter) {
-  AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, iter.input_dtype(), "min_all_cuda", [&] {
+  AT_DISPATCH_V2(iter.input_dtype(), "min_all_cuda", AT_WRAP([&] {
     min_values_kernel_cuda_impl<scalar_t>(iter);
-  });
+  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf, kBool);
 }
 
 REGISTER_DISPATCH(min_values_stub, &min_values_kernel_cuda)

aten/src/ATen/native/mps/OperationUtils.h

@@ -40,6 +40,8 @@ using namespace at::mps;
 
 namespace at::native::mps {
 
+void dispatch_sync_with_rethrow(dispatch_queue_t queue, void (^block)());
+
 struct MPSScalar {
   id<MTLBuffer> getMTLBuffer() const {
     return __builtin_bit_cast(id<MTLBuffer>, buffer.get());

aten/src/ATen/native/mps/OperationUtils.mm

@@ -53,6 +53,21 @@
 @end
 
 namespace at::native::mps {
+
+void dispatch_sync_with_rethrow(dispatch_queue_t queue, void (^block)()) {
+  __block std::optional<std::exception_ptr> block_exception;
+  dispatch_sync(queue, ^() {
+    try {
+      block();
+    } catch (...) {
+      block_exception = std::current_exception();
+    }
+  });
+  if (block_exception) {
+    std::rethrow_exception(*block_exception);
+  }
+}
+
 /**
  * Computes distance from lowest to highest element offset in given tensor.
  */

aten/src/ATen/native/mps/kernels/Indexing.metal

@@ -1,5 +1,4 @@
 #include <c10/metal/atomic.h>
-#include <c10/metal/error.h>
 #include <c10/metal/indexing.h>
 #include <metal_stdlib>
 
@@ -32,24 +31,10 @@ OffsetT index_apply_indices(
     constant IndexAB* indices,
     constant int64_t* sizes,
     constant int64_t* strides,
-    uint num_indices,
-    thread bool& error,
-    device ErrorMessages* error_buf) {
+    uint num_indices) {
   OffsetT rc = offs.x;
   for (uint i = 0; i < num_indices; i++) {
     auto idx = indices[i].indexArray[offs.y];
-    if (idx < -sizes[i] || idx >= sizes[i]) {
-      TORCH_REPORT_ERROR(
-          error_buf,
-          "index ",
-          idx,
-          " is out of bounds for dimension ",
-          i,
-          " with size ",
-          sizes[i]);
-      error = true;
-      break;
-    }
     if (idx < 0) {
       idx += sizes[i];
     }
@@ -70,7 +55,6 @@ kernel void index_select(
     constant int64_t* index_sizes,
     constant int64_t* index_strides,
     constant uint4& ndim_nindices_numel,
-    device ErrorMessages* error_buffer,
     uint thread_index [[thread_position_in_grid]]) {
   const auto ndim = ndim_nindices_numel.x;
   const auto num_indices = ndim_nindices_numel.y;
@@ -81,19 +65,8 @@ kernel void index_select(
       indices_strides,
       ndim,
       thread_index);
-  bool error = false;
   auto input_offs = index_apply_indices<OffsetT>(
-      offs.yz,
-      indices,
-      index_sizes,
-      index_strides,
-      num_indices,
-      error,
-      error_buffer);
-  if (error) {
-    output[offs.x / sizeof(T)] = 0;
-    return;
-  }
+      offs.yz, indices, index_sizes, index_strides, num_indices);
   output[offs.x / sizeof(T)] = input[input_offs / sizeof(T)];
 }
 
@@ -109,9 +82,7 @@ inline void index_put_impl(
     constant int64_t* index_sizes,
     constant int64_t* index_strides,
     constant uint4& ndim_nindices_numel,
-    device ErrorMessages* error_buffer,
     uint thread_index) {
-  bool error = false;
   const auto ndim = ndim_nindices_numel.x;
   const auto num_indices = ndim_nindices_numel.y;
   const auto offs = index_get_offsets(
@@ -122,16 +93,7 @@ inline void index_put_impl(
       ndim,
       thread_index);
   auto output_offs = index_apply_indices<OffsetT>(
-      offs.xz,
-      indices,
-      index_sizes,
-      index_strides,
-      num_indices,
-      error,
-      error_buffer);
-  if (error) {
-    return;
-  }
+      offs.xz, indices, index_sizes, index_strides, num_indices);
   output[output_offs / sizeof(T)] = input[offs.y / sizeof(T)];
 }
 
@@ -147,7 +109,6 @@ kernel void index_put(
     constant int64_t* index_sizes,
     constant int64_t* index_strides,
     constant uint4& ndim_nindices_numel,
-    device ErrorMessages* error_buffer,
     uint thread_index [[thread_position_in_grid]]) {
   index_put_impl(
       output,
@@ -160,7 +121,6 @@ kernel void index_put(
       index_sizes,
       index_strides,
       ndim_nindices_numel,
-      error_buffer,
       thread_index);
 }
 
@@ -176,7 +136,6 @@ kernel void index_put_serial(
     constant int64_t* index_sizes,
     constant int64_t* index_strides,
     constant uint4& ndim_nindices_numel,
-    device ErrorMessages* error_buffer,
     uint thread_index [[thread_position_in_grid]]) {
   (void)thread_index; // Suppress unused vairable varning
   for (uint idx = 0; idx < ndim_nindices_numel.z; ++idx) {
@@ -191,7 +150,6 @@ kernel void index_put_serial(
         index_sizes,
         index_strides,
         ndim_nindices_numel,
-        error_buffer,
         idx);
   }
 }
@@ -208,7 +166,6 @@ kernel void index_put_accumulate(
     constant int64_t* index_sizes,
     constant int64_t* index_strides,
     constant uint4& ndim_nindices_numel,
-    device ErrorMessages* error_buffer,
     uint thread_index [[thread_position_in_grid]]) {
   const auto ndim = ndim_nindices_numel.x;
   const auto num_indices = ndim_nindices_numel.y;
@@ -219,18 +176,8 @@ kernel void index_put_accumulate(
       indices_strides,
       ndim,
       thread_index);
-  bool error = false;
   auto output_offs = index_apply_indices<OffsetT>(
-      offs.xz,
-      indices,
-      index_sizes,
-      index_strides,
-      num_indices,
-      error,
-      error_buffer);
-  if (error) {
-    return;
-  }
+      offs.xz, indices, index_sizes, index_strides, num_indices);
   AtomicType<T>::atomic_add(
       reinterpret_cast<device AtomicType_t<T>*>(output),
       output_offs / sizeof(T),
@@ -250,7 +197,6 @@ kernel void index_put_accumulate(
           constant int64_t* index_sizes,                            \
           constant int64_t* index_strides,                          \
           constant uint4& ndim_nindices_numel,                      \
-          device ErrorMessages* error_buffer,                       \
           uint thread_index [[thread_position_in_grid]])
 
 #define REGISTER_INDEX_OP_ALL_DTYPES(OP_NAME) \

aten/src/ATen/native/mps/operations/EmbeddingBag.mm

@@ -220,7 +220,7 @@ Tensor _embedding_bag_dense_backward_mps(const Tensor& output_grad,
   auto num_threads = (params.mode == EmbeddingBagMode::MAX) ? output_grad.numel() : num_indices * params.feature_size;
   MPSStream* stream = getCurrentMPSStream();
 
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
+  mps::dispatch_sync_with_rethrow(stream->queue(), ^() {
     @autoreleasepool {
       id<MTLComputeCommandEncoder> computeEncoder = stream->commandEncoder();
       auto pipeline_state = lib.getPipelineStateForFunc(fmt::format("embedding_bag_backward_{}_{}",
@@ -273,7 +273,7 @@ Tensor _embedding_bag_per_sample_weights_backward_mps(const Tensor& output_grad,
   auto num_threads = num_indices * feature_size;
   MPSStream* stream = getCurrentMPSStream();
 
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
+  mps::dispatch_sync_with_rethrow(stream->queue(), ^() {
     @autoreleasepool {
       id<MTLComputeCommandEncoder> computeEncoder = stream->commandEncoder();
       auto pipeline_state = lib.getPipelineStateForFunc(fmt::format("embedding_bag_per_sample_weights_backward_{}_{}",

aten/src/ATen/native/mps/operations/Indexing.mm

@@ -179,8 +179,7 @@ static void dispatch_index_kernel(TensorIteratorBase& iter,
                    iter.strides(2),
                    index_size,
                    index_stride,
-                   ndim_nindiees,
-                   mpsStream->getErrorBuffer());
+                   ndim_nindiees);
     mtl_dispatch1DJob(computeEncoder, indexSelectPSO, serial ? 1 : iter.numel());
   });
 }
@@ -300,7 +299,7 @@ static Tensor& nonzero_out_native_mps(const Tensor& self, Tensor& out_) {
   MPSStream* stream = getCurrentMPSStream();
   using CachedGraph = MPSUnaryCachedGraph;
 
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
+  dispatch_sync(stream->queue(), ^() {
     stream->synchronize(SyncType::COMMIT_AND_WAIT);
   });
   int64_t total_nonzero = at::count_nonzero(self).item<int64_t>();
@@ -385,7 +384,7 @@ Tensor& nonzero_out_mps(const Tensor& self, Tensor& out_) {
   MPSStream* stream = getCurrentMPSStream();
   using CachedGraph = MPSUnaryCachedGraph;
 
-  dispatch_sync_with_rethrow(stream->queue(), ^() {
+  dispatch_sync(stream->queue(), ^() {
     stream->synchronize(SyncType::COMMIT_AND_WAIT);
   });
   int64_t total_nonzero = at::count_nonzero(self).item<int64_t>();

aten/src/ATen/native/mps/operations/Normalization.mm

@@ -923,7 +923,7 @@ std::tuple<Tensor, Tensor, Tensor> layer_norm_mps(const Tensor& input,
   MPSStream* stream = getCurrentMPSStream();
   TORCH_CHECK_NOT_IMPLEMENTED(input.scalar_type() != kLong, "Not implemented for long on MPS");
   @autoreleasepool {
-    dispatch_sync_with_rethrow(stream->queue(), ^() {
+    mps::dispatch_sync_with_rethrow(stream->queue(), ^() {
       // which kernel variant to use based on the normalized axis N size
       const int N_READS = 4;
       auto metalType = mps::scalarToMetalTypeString(input);

aten/src/ATen/native/native_functions.yaml

@@ -192,11 +192,6 @@
     CompositeExplicitAutograd: _assert_tensor_metadata
     Meta: _assert_tensor_metadata_meta_symint
 
-- func: _async_error(str msg) -> ()
-  dispatch:
-    CompositeExplicitAutograd: _async_error
-    Meta: _async_error_meta
-
 - func: _print(str s) -> ()
   dispatch:
     CompositeExplicitAutograd: _print

aten/src/ATen/native/sparse/cuda/SoftMax.cu

@@ -47,7 +47,6 @@
 #include <c10/macros/Macros.h>
 #include <thrust/copy.h>
 #include <thrust/device_ptr.h>
-#include <thrust/distance.h>
 #include <thrust/for_each.h>
 #include <thrust/functional.h>
 #include <thrust/gather.h>

aten/src/ATen/test/CMakeLists.txt

@@ -61,7 +61,6 @@ list(APPEND ATen_CUDA_TEST_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_complex_math_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_complex_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_cub_test.cu
-  ${CMAKE_CURRENT_SOURCE_DIR}/cuda_cublas_handle_pool_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_device_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_distributions_test.cu
   ${CMAKE_CURRENT_SOURCE_DIR}/cuda_dlconvertor_test.cpp

aten/src/ATen/test/cuda_cublas_handle_pool_test.cpp

@@ -1,77 +0,0 @@
-#include <gtest/gtest.h>
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDACachingAllocator.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include <atomic>
-#include <thread>
-#include <vector>
-
-// Test concurrent access to getCurrentCUDABlasHandle and getCUDABlasLtWorkspace
-// to verify that the data race fix is working correctly
-
-TEST(CUDABlasHandlePoolTest, ConcurrentGetAndClearWorkspaces) {
-  if (!at::cuda::is_available()) {
-    return;
-  }
-
-  constexpr int num_accessor_threads = 15;
-  constexpr int num_clear_threads = 5;
-  constexpr int iterations_per_thread = 50;
-
-  std::atomic<bool> stop{false};
-  std::atomic<int> error_count{0};
-  std::vector<std::thread> threads;
-  threads.reserve(num_accessor_threads + num_clear_threads);
-
-  // Launch accessor threads
-  for (int i = 0; i < num_accessor_threads; ++i) {
-    threads.emplace_back([&stop, &error_count]() {
-      try {
-        at::cuda::CUDAGuard device_guard(0);
-
-        while (!stop.load(std::memory_order_relaxed)) {
-          const auto handle = at::cuda::getCurrentCUDABlasHandle();
-          const auto workspace = at::cuda::getCUDABlasLtWorkspace();
-
-          if (handle == nullptr || workspace == nullptr) {
-            error_count++;
-          }
-        }
-      } catch (const std::exception& e) {
-        error_count++;
-      }
-    });
-  }
-
-  // Launch threads that clear workspaces
-  for (int i = 0; i < num_clear_threads; ++i) {
-    threads.emplace_back([&error_count]() {
-      try {
-        for (int j = 0; j < iterations_per_thread; ++j) {
-          at::cuda::clearCublasWorkspaces();
-          std::this_thread::yield();
-        }
-      } catch (const std::exception& e) {
-        error_count++;
-      }
-    });
-  }
-
-  // Let them run for a bit
-  std::this_thread::sleep_for(std::chrono::milliseconds(100));
-  stop.store(true, std::memory_order_relaxed);
-
-  for (auto& thread : threads) {
-    thread.join();
-  }
-
-  EXPECT_EQ(error_count.load(), 0);
-}
-
-int main(int argc, char* argv[]) {
-  ::testing::InitGoogleTest(&argc, argv);
-  c10::cuda::CUDACachingAllocator::init(1);
-  return RUN_ALL_TESTS();
-}

aten/src/ATen/xpu/XPUEvent.h

@@ -1,3 +1,191 @@
 #pragma once
 #include <ATen/xpu/XPUContext.h>
-#include <c10/xpu/XPUEvent.h>
+
+#include <optional>
+
+namespace at::xpu {
+
+/*
+ * XPUEvent are movable not copyable wrappers around SYCL event. XPUEvent are
+ * constructed lazily when first recorded. It has a device, and this device is
+ * acquired from the first recording stream. Later streams that record the event
+ * must match the same device.
+ *
+ * Currently, XPUEvent does NOT support to export an inter-process event from
+ * another process via inter-process communication(IPC). So it means that
+ * inter-process communication for event handles between different processes is
+ * not available. This could impact some applications that rely on cross-process
+ * synchronization and communication.
+ */
+struct TORCH_XPU_API XPUEvent {
+  // Constructors
+  XPUEvent(bool enable_timing = false) noexcept
+      : enable_timing_{enable_timing} {}
+
+  ~XPUEvent() {
+    if (isCreated()) {
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_deletion(
+            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+    }
+  }
+
+  XPUEvent(const XPUEvent&) = delete;
+  XPUEvent& operator=(const XPUEvent&) = delete;
+
+  XPUEvent(XPUEvent&& other) = default;
+  XPUEvent& operator=(XPUEvent&& other) = default;
+
+  operator sycl::event&() const {
+    return event();
+  }
+
+  std::optional<at::Device> device() const {
+    if (isCreated()) {
+      return at::Device(at::kXPU, device_index_);
+    } else {
+      return std::nullopt;
+    }
+  }
+
+  inline bool isCreated() const {
+    return (event_.get() != nullptr);
+  }
+
+  DeviceIndex device_index() const {
+    return device_index_;
+  }
+
+  sycl::event& event() const {
+    return *event_;
+  }
+
+  bool query() const {
+    using namespace sycl::info;
+    if (!isCreated()) {
+      return true;
+    }
+
+    return event().get_info<event::command_execution_status>() ==
+        event_command_status::complete;
+  }
+
+  void record() {
+    record(getCurrentXPUStream());
+  }
+
+  void recordOnce(const XPUStream& stream) {
+    if (!isCreated()) {
+      record(stream);
+    }
+  }
+
+  void record(const XPUStream& stream) {
+    if (!isCreated()) {
+      device_index_ = stream.device_index();
+      assignEvent(stream.queue());
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_creation(
+            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+    } else {
+      TORCH_CHECK(
+          device_index_ == stream.device_index(),
+          "Event device ",
+          device_index_,
+          " does not match recording stream's device ",
+          stream.device_index(),
+          ".");
+      reassignEvent(stream.queue());
+    }
+    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+    if (C10_UNLIKELY(interp)) {
+      (*interp)->trace_gpu_event_record(
+          at::kXPU,
+          reinterpret_cast<uintptr_t>(event_.get()),
+          reinterpret_cast<uintptr_t>(&stream.queue()));
+    }
+  }
+
+  void block(const XPUStream& stream) {
+    if (isCreated()) {
+      std::vector<sycl::event> event_list{event()};
+      // Make this stream wait until event_ is completed.
+      stream.queue().ext_oneapi_submit_barrier(event_list);
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_wait(
+            at::kXPU,
+            reinterpret_cast<uintptr_t>(event_.get()),
+            reinterpret_cast<uintptr_t>(&stream.queue()));
+      }
+    }
+  }
+
+  double elapsed_time(const XPUEvent& other) const {
+    TORCH_CHECK(
+        isCreated() && other.isCreated(),
+        "Both events must be recorded before calculating elapsed time.");
+    TORCH_CHECK(
+        query() && other.query(),
+        "Both events must be completed before calculating elapsed time.");
+    TORCH_CHECK(
+        enable_timing_ && other.enable_timing_,
+        "Both events must be created with argument 'enable_timing=True'.");
+
+#if SYCL_COMPILER_VERSION < 20250000
+    TORCH_CHECK_NOT_IMPLEMENTED(
+        false,
+        "elapsed_time of XPUEvent requires PyTorch to be built with SYCL compiler version 2025.0.0 or newer.");
+#endif
+
+    using namespace sycl::info::event_profiling;
+    // Block until both of the recorded events are completed.
+    uint64_t end_time_ns = other.event().get_profiling_info<command_end>();
+    uint64_t start_time_ns = event().get_profiling_info<command_end>();
+    // Return the eplased time in milliseconds.
+    return 1e-6 *
+        (static_cast<double>(end_time_ns) - static_cast<double>(start_time_ns));
+  }
+
+  void synchronize() const {
+    if (isCreated()) {
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_synchronization(
+            at::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
+      }
+      event().wait_and_throw();
+    }
+  }
+
+ private:
+  void assignEvent(sycl::queue& queue) {
+#if SYCL_COMPILER_VERSION >= 20250000
+    if (enable_timing_) {
+      event_ = std::make_unique<sycl::event>(
+          sycl::ext::oneapi::experimental::submit_profiling_tag(queue));
+    } else {
+      event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
+    }
+#else
+    event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
+#endif
+  }
+
+  void reassignEvent(sycl::queue& queue) {
+    event_.reset();
+    assignEvent(queue);
+  }
+
+  bool enable_timing_ = false;
+  DeviceIndex device_index_ = -1;
+  // Only need to track the last event, as events in an in-order queue are
+  // executed sequentially.
+  std::unique_ptr<sycl::event> event_;
+};
+
+} // namespace at::xpu

benchmarks/dynamo/check_accuracy.py

@@ -50,7 +50,6 @@ def check_accuracy(actual_csv, expected_csv, expected_filename):
                 "mobilenet_v2",
                 "pytorch_CycleGAN_and_pix2pix",
                 "pytorch_stargan",
-                "repvgg_a2",
                 "resnet152",
                 "resnet18",
                 "resnet50",

benchmarks/dynamo/ci_expected_accuracy/dynamic_inductor_timm_training.csv

@@ -10,7 +10,7 @@ beit_base_patch16_224,pass,7
 
 
 
-convnextv2_nano.fcmae_ft_in22k_in1k,fail_accuracy,7
+convnextv2_nano.fcmae_ft_in22k_in1k,pass,7
 
 
 
@@ -66,7 +66,7 @@ visformer_small,pass,7
 
 
 
-vit_base_patch14_dinov2.lvd142m,fail_accuracy,7
+vit_base_patch14_dinov2.lvd142m,pass,7
 
 
 

benchmarks/dynamo/ci_expected_accuracy/rocm/dynamic_inductor_timm_training.csv

@@ -50,7 +50,7 @@ nfnet_l0,pass,7
 
 
 
-repvgg_a2,pass,7
+repvgg_a2,fail_accuracy,7
 
 
 

benchmarks/dynamo/common.py

@@ -952,7 +952,7 @@ def latency_experiment_summary(suite_name, args, model, timings, **kwargs):
         first_fields.append(kwargs["tag"])
     headers = first_headers + ["speedup", "abs_latency"]
     row = first_fields + [float(speedup), median[1] * 1000]
-    msg = f"{median[0] * 1000} ms, {median[1] * 1000} ms, {speedup:.3f}x"
+    msg = f"{speedup:.3f}x"
     if args.baseline:
         headers.extend(
             [
@@ -1010,7 +1010,7 @@ def latency_experiment_summary(suite_name, args, model, timings, **kwargs):
     # Hypothetically you can use this from other places, but it's currently
     # inaccessible, and when this assert fails you need to update the
     # event_name here to account for the other cases you are using this
-    assert any([args.quantization, args.optimus])
+    assert args.quantization is not None
     output_signpost(
         dict(zip(headers, row)),
         args,
@@ -2288,9 +2288,11 @@ class BenchmarkRunner:
                     )
                 ):
                     is_same = False
-            except Exception:
+            except Exception as e:
                 # Sometimes torch.allclose may throw RuntimeError
-                is_same = False
+                exception_string = str(e)
+                accuracy_status = f"fail_exception: {exception_string}"
+                return record_status(accuracy_status, dynamo_start_stats=start_stats)
 
             if not is_same:
                 accuracy_status = "eager_two_runs_differ"
@@ -2407,9 +2409,11 @@ class BenchmarkRunner:
                     force_max_multiplier=force_max_multiplier,
                 ):
                     is_same = False
-            except Exception:
+            except Exception as e:
                 # Sometimes torch.allclose may throw RuntimeError
-                is_same = False
+                exception_string = str(e)
+                accuracy_status = f"fail_exception: {exception_string}"
+                return record_status(accuracy_status, dynamo_start_stats=start_stats)
 
             if not is_same:
                 if self.args.skip_accuracy_check:
@@ -2583,9 +2587,6 @@ class BenchmarkRunner:
                 **experiment_kwargs,
             )
 
-            # reset dynamo
-            torch._dynamo.reset()
-
             if self.args.export_aot_inductor:
                 optimized_model_iter_fn = optimize_ctx
             else:
@@ -2949,7 +2950,7 @@ class BenchmarkRunner:
             status = self.check_tolerance(name, model, example_inputs, optimize_ctx)
             print(status)
         elif self.args.performance:
-            if self.args.backend in ["torchao", "optimus"]:
+            if self.args.backend == "torchao":
                 status = self.run_performance_test_non_alternate(
                     name, model, example_inputs, optimize_ctx, experiment, tag
                 )
@@ -3525,12 +3526,6 @@ def parse_args(args=None):
         action="store_true",
         help="Measure speedup with TorchInductor",
     )
-    group.add_argument(
-        "--optimus",
-        choices=["vertical_opt", "horizontal_opt", "all"],
-        default=None,
-        help="Measure speedup of Optimus with TorchInductor baseline",
-    )
     group.add_argument(
         "--quantization",
         choices=[
@@ -3788,9 +3783,6 @@ def run(runner, args, original_dir=None):
     if args.inductor:
         assert args.backend is None
         args.backend = "inductor"
-    if args.optimus:
-        assert args.backend is None
-        args.backend = "optimus"
     if args.quantization:
         assert args.backend is None
         args.backend = "torchao"
@@ -4075,22 +4067,10 @@ def run(runner, args, original_dir=None):
 
             runner.model_iter_fn = model_iter_fn_and_mark_step
             optimize_ctx = torchao_optimize_ctx(args.quantization)
-        elif args.backend == "optimus":
-            from .optimus import get_baseline_ctx, get_optimus_optimize_ctx
-
-            baseline_ctx = get_baseline_ctx(
-                nopython=args.nopython, inductor_compile_mode=args.inductor_compile_mode
-            )
-            runner.model_iter_fn = baseline_ctx(runner.model_iter_fn)
-            optimize_ctx = get_optimus_optimize_ctx(
-                args.optimus, args.nopython, args.inductor_compile_mode
-            )
         else:
             optimize_ctx = torch._dynamo.optimize(args.backend, nopython=args.nopython)
         experiment = (
-            speedup_experiment
-            if args.backend not in ["torchao", "optimus"]
-            else latency_experiment
+            speedup_experiment if args.backend != "torchao" else latency_experiment
         )
         if args.accuracy:
             output_filename = f"accuracy_{args.backend}.csv"
@@ -4111,12 +4091,7 @@ def run(runner, args, original_dir=None):
     if args.only in runner.disable_cudagraph_models:
         args.disable_cudagraphs = True
 
-    if (
-        args.inductor
-        or args.backend == "inductor"
-        or args.export_aot_inductor
-        or args.backend == "optimus"
-    ):
+    if args.inductor or args.backend == "inductor" or args.export_aot_inductor:
         inductor_config.triton.cudagraphs = not args.disable_cudagraphs
         inductor_config.triton.persistent_reductions = (
             not args.disable_persistent_reductions

benchmarks/dynamo/optimus.py

@@ -1,62 +0,0 @@
-import functools
-
-import torch
-
-
-def get_baseline_ctx(nopython, inductor_compile_mode):
-    return functools.partial(
-        torch.compile,
-        backend="inductor",
-        fullgraph=nopython,
-        mode=inductor_compile_mode,
-    )
-
-
-def get_optimus_optimize_ctx(config, nopython, inductor_compile_mode):
-    if config == "vertical_opt":
-        optimus_inductor_config = {
-            "pre_grad_fusion_options": {
-                "normalization_pass": {},
-                "merge_splits_pass": {},
-                "split_cat_pass": {},
-                "unbind_stack_pass": {},
-                "unbind_cat_to_view_pass": {},
-            }
-        }
-    elif config == "horizontal_opt":
-        optimus_inductor_config = {
-            "pre_grad_fusion_options": {
-                "normalization_pass": {},
-                "batch_linear": {},
-                "batch_layernorm": {},
-            },
-        }
-    elif config == "all":
-        optimus_inductor_config = {
-            "pre_grad_fusion_options": {
-                "normalization_pass": {},
-                "batch_linear": {},
-                "batch_layernorm": {},
-                "merge_splits_pass": {},
-                "split_cat_pass": {},
-                "unbind_stack_pass": {},
-                "unbind_cat_to_view_pass": {},
-            },
-        }
-    else:
-        raise RuntimeError(f"Unknown optimus config: {config}")
-
-    def _inner(fn):
-        if "pre_grad_fusion_options" in optimus_inductor_config:
-            torch._inductor.config.pre_grad_fusion_options = optimus_inductor_config[
-                "pre_grad_fusion_options"
-            ]
-        if "post_grad_fusion_options" in optimus_inductor_config:
-            torch._inductor.config.post_grad_fusion_options = optimus_inductor_config[
-                "post_grad_fusion_options"
-            ]
-        return torch.compile(
-            fn, backend="inductor", fullgraph=nopython, mode=inductor_compile_mode
-        )
-
-    return _inner

benchmarks/dynamo/parse_logs.py

@@ -2,7 +2,6 @@ import csv
 import os
 import re
 import sys
-from pathlib import Path
 
 
 # This script takes the logs produced by the benchmark scripts (e.g.,
@@ -16,7 +15,8 @@ from pathlib import Path
 # This script is not very well written, feel free to rewrite it as necessary
 
 assert len(sys.argv) == 2
-full_log = Path(sys.argv[1]).read_text()
+
+full_log = open(sys.argv[1]).read()
 
 # If the log contains a gist URL, extract it so we can include it in the CSV
 gist_url = ""

benchmarks/operator_benchmark/aarch64_expected_ci_operator_benchmark_eager_float32_cpu.csv

@@ -484,106 +484,24 @@ PyTorch,sum,sum_R256_V512_dim0_contiguousTrue_cpu,short,False,50.954394,0.000000
 PyTorch,sum,sum_R256_V512_dim0_contiguousFalse_cpu,short,False,57.957757,0.000000
 PyTorch,sum,sum_R256_V512_dim1_contiguousTrue_cpu,short,False,53.592068,0.000000
 PyTorch,sum,sum_R256_V512_dim1_contiguousFalse_cpu,short,False,51.339726,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.bool,short,False,0.927,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.uint8,short,False,6.261,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int8,short,False,6.351,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int16,short,False,6.177,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int32,short,False,6.333,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int64,short,False,6.588,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float16,short,False,8.117,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.bfloat16,short,False,9.358,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float32,short,False,7.844,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float64,short,False,8.097,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.bool,short,False,6.159,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.uint8,short,False,0.926,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int8,short,False,6.192,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int16,short,False,6.276,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int32,short,False,6.461,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int64,short,False,6.524,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float16,short,False,8.136,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.bfloat16,short,False,6.854,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float32,short,False,6.446,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float64,short,False,6.829,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.bool,short,False,6.088,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.uint8,short,False,6.059,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int8,short,False,0.922,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int16,short,False,6.263,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int32,short,False,6.330,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int64,short,False,6.688,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float16,short,False,8.176,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.bfloat16,short,False,6.959,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float32,short,False,6.430,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float64,short,False,6.818,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.bool,short,False,6.350,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.uint8,short,False,6.221,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int8,short,False,6.193,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int16,short,False,0.922,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int32,short,False,6.263,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int64,short,False,6.525,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float16,short,False,7.960,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.bfloat16,short,False,6.801,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float32,short,False,6.594,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float64,short,False,7.089,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.bool,short,False,6.498,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.uint8,short,False,6.358,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int8,short,False,6.390,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int16,short,False,6.415,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int32,short,False,0.925,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int64,short,False,6.657,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float16,short,False,7.954,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.bfloat16,short,False,6.930,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float32,short,False,6.737,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float64,short,False,6.948,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.bool,short,False,6.757,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.uint8,short,False,6.402,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int8,short,False,6.550,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int16,short,False,6.518,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int32,short,False,6.766,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int64,short,False,0.929,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float16,short,False,8.557,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.bfloat16,short,False,9.045,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float32,short,False,7.672,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float64,short,False,7.276,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.bool,short,False,6.414,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.uint8,short,False,7.736,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int8,short,False,7.889,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int16,short,False,8.170,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int32,short,False,7.783,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int64,short,False,7.743,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float16,short,False,0.927,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.bfloat16,short,False,7.018,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float32,short,False,8.428,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float64,short,False,6.767,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.bool,short,False,6.479,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.uint8,short,False,7.827,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int8,short,False,6.450,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int16,short,False,6.320,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int32,short,False,6.385,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int64,short,False,8.119,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float16,short,False,8.063,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.bfloat16,short,False,0.925,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float32,short,False,8.629,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float64,short,False,6.638,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.bool,short,False,6.425,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.uint8,short,False,7.803,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int8,short,False,6.502,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int16,short,False,6.429,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int32,short,False,6.549,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int64,short,False,7.749,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float16,short,False,7.301,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.bfloat16,short,False,7.682,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float32,short,False,0.930,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float64,short,False,6.738,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.bool,short,False,6.798,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.uint8,short,False,6.506,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int8,short,False,6.494,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int16,short,False,6.668,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int32,short,False,6.696,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int64,short,False,7.115,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float16,short,False,7.910,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.bfloat16,short,False,7.410,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float32,short,False,6.868,0.000000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float64,short,False,0.924,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N16_cpu,short,False,7.040985,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N64_cpu,short,False,7.168604,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N128_cpu,short,False,7.434442,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N16_cpu,short,False,7.078318,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N64_cpu,short,False,7.426670,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N128_cpu,short,False,7.679027,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N16_cpu,short,False,7.281365,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N64_cpu,short,False,7.682783,0.000000
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N128_cpu,short,False,8.381938,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N16_cpu,short,False,7.039854,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N64_cpu,short,False,7.399855,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N128_cpu,short,False,7.715193,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N16_cpu,short,False,7.255140,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N64_cpu,short,False,7.753522,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N128_cpu,short,False,8.364281,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N16_cpu,short,False,7.476377,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N64_cpu,short,False,8.458564,0.000000
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N128_cpu,short,False,9.391939,0.000000
 PyTorch,addcmul,addcmul_M1_N2_cpu_dtypetorch.float32,short,False,4.461410,0.000000
 PyTorch,addcmul,addcmul_M1_N2_cpu_dtypetorch.bfloat16,short,False,4.560082,0.000000
 PyTorch,addcmul,addcmul_M32_N64_cpu_dtypetorch.float32,short,False,5.141248,0.000000

benchmarks/operator_benchmark/pt/tensor_to_test.py

@@ -4,84 +4,74 @@ import torch
 
 
 tensor_conversion_short_configs = op_bench.cross_product_configs(
-    M=[32],
-    N=[128],
+    M=(
+        8,
+        16,
+        32,
+    ),
+    N=(
+        16,
+        64,
+        128,
+    ),
     device=["cpu", "cuda"],
-    dtype_one=[
-        torch.bool,
-        torch.uint8,
-        torch.int8,
-        torch.int16,
-        torch.int32,
-        torch.int64,
-        torch.half,
-        torch.bfloat16,
-        torch.float,
-        torch.double,
-    ],
-    dtype_two=[
-        torch.bool,
-        torch.uint8,
-        torch.int8,
-        torch.int16,
-        torch.int32,
-        torch.int64,
-        torch.half,
-        torch.bfloat16,
-        torch.float,
-        torch.double,
-    ],
     tags=["short"],
 )
 
 tensor_conversion_long_configs = op_bench.cross_product_configs(
-    M=[1024],
-    N=[1024],
+    M=(
+        64,
+        128,
+        256,
+        512,
+    ),
+    N=(
+        256,
+        512,
+        1024,
+        2048,
+    ),
     device=["cpu", "cuda"],
-    dtype_one=[
-        torch.bool,
-        torch.uint8,
-        torch.int8,
-        torch.int16,
-        torch.int32,
-        torch.int64,
-        torch.half,
-        torch.bfloat16,
-        torch.float,
-        torch.double,
-    ],
-    dtype_two=[
-        torch.bool,
-        torch.uint8,
-        torch.int8,
-        torch.int16,
-        torch.int32,
-        torch.int64,
-        torch.half,
-        torch.bfloat16,
-        torch.float,
-        torch.double,
-    ],
     tags=["long"],
 )
 
 
-class TensorConversionBenchmark(op_bench.TorchBenchmarkBase):
-    def init(self, M, N, dtype_one, dtype_two, device):
+class FloatToHalfTensorConversionBenchmark(op_bench.TorchBenchmarkBase):
+    def init(self, M, N, device):
         self.inputs = {
             "input": torch.rand(
                 M, N, device=device, requires_grad=False, dtype=torch.float
-            ).to(dtype=dtype_one)
+            )
         }
-        self.dtype_one = dtype_one
-        self.dtype_two = dtype_two
 
     def forward(self, input):
-        return input.to(dtype=self.dtype_two)
+        return input.to(torch.half)
 
 
-op_bench.generate_pt_test(tensor_conversion_short_configs, TensorConversionBenchmark)
-op_bench.generate_pt_test(tensor_conversion_long_configs, TensorConversionBenchmark)
+class HalfToFloatTensorConversionBenchmark(op_bench.TorchBenchmarkBase):
+    def init(self, M, N, device):
+        self.inputs = {
+            "input": torch.rand(
+                M, N, device=device, requires_grad=False, dtype=torch.half
+            )
+        }
+
+    def forward(self, input):
+        return input.to(torch.float)
+
+
+op_bench.generate_pt_test(
+    tensor_conversion_short_configs, FloatToHalfTensorConversionBenchmark
+)
+op_bench.generate_pt_test(
+    tensor_conversion_long_configs, FloatToHalfTensorConversionBenchmark
+)
+op_bench.generate_pt_test(
+    tensor_conversion_short_configs, HalfToFloatTensorConversionBenchmark
+)
+op_bench.generate_pt_test(
+    tensor_conversion_long_configs, HalfToFloatTensorConversionBenchmark
+)
 
 if __name__ == "__main__":
     op_bench.benchmark_runner.main()

benchmarks/operator_benchmark/x86_64_expected_ci_operator_benchmark_eager_float32_cpu.csv

@@ -349,106 +349,24 @@ PyTorch,sum,sum_R256_V512_dim0_contiguousTrue_cpu,short,FALSE,12.5841
 PyTorch,sum,sum_R256_V512_dim0_contiguousFALSE_cpu,short,FALSE,20.8765
 PyTorch,sum,sum_R256_V512_dim1_contiguousTrue_cpu,short,FALSE,15.4414
 PyTorch,sum,sum_R256_V512_dim1_contiguousFALSE_cpu,short,FALSE,15.3287
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.bool,short,False,0.797
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.uint8,short,False,6.071
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int8,short,False,6.031
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int16,short,False,6.243
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int32,short,False,7.231
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.int64,short,False,7.791
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float16,short,False,12.661
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.bfloat16,short,False,11.225
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float32,short,False,9.772
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bool_dtype_twotorch.float64,short,False,9.872
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.bool,short,False,6.033
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.uint8,short,False,0.781
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int8,short,False,6.060
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int16,short,False,6.180
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int32,short,False,7.258
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.int64,short,False,7.758
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float16,short,False,10.504
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.bfloat16,short,False,6.749
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float32,short,False,7.679
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.uint8_dtype_twotorch.float64,short,False,7.797
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.bool,short,False,6.019
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.uint8,short,False,6.079
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int8,short,False,0.785
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int16,short,False,6.188
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int32,short,False,7.288
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.int64,short,False,7.770
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float16,short,False,10.466
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.bfloat16,short,False,6.676
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float32,short,False,7.736
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int8_dtype_twotorch.float64,short,False,7.780
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.bool,short,False,6.130
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.uint8,short,False,6.221
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int8,short,False,6.101
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int16,short,False,0.791
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int32,short,False,6.254
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.int64,short,False,7.733
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float16,short,False,10.562
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.bfloat16,short,False,6.704
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float32,short,False,7.819
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int16_dtype_twotorch.float64,short,False,8.276
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.bool,short,False,6.361
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.uint8,short,False,6.364
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int8,short,False,6.309
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int16,short,False,6.362
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int32,short,False,0.791
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.int64,short,False,7.746
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float16,short,False,9.462
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.bfloat16,short,False,6.678
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float32,short,False,7.827
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int32_dtype_twotorch.float64,short,False,8.200
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.bool,short,False,6.925
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.uint8,short,False,6.947
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int8,short,False,6.962
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int16,short,False,6.906
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int32,short,False,7.664
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.int64,short,False,0.782
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float16,short,False,10.528
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.bfloat16,short,False,10.123
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float32,short,False,9.234
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.int64_dtype_twotorch.float64,short,False,8.694
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.bool,short,False,12.653
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.uint8,short,False,9.348
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int8,short,False,8.774
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int16,short,False,9.063
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int32,short,False,10.012
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.int64,short,False,13.641
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float16,short,False,0.788
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.bfloat16,short,False,13.757
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float32,short,False,7.170
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float16_dtype_twotorch.float64,short,False,12.511
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.bool,short,False,6.516
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.uint8,short,False,8.539
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int8,short,False,6.483
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int16,short,False,6.468
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int32,short,False,7.752
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.int64,short,False,9.868
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float16,short,False,10.556
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.bfloat16,short,False,0.792
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float32,short,False,7.577
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.bfloat16_dtype_twotorch.float64,short,False,8.267
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.bool,short,False,6.819
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.uint8,short,False,7.715
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int8,short,False,6.754
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int16,short,False,6.825
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int32,short,False,7.790
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.int64,short,False,9.219
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float16,short,False,5.977
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.bfloat16,short,False,7.069
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float32,short,False,0.794
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float32_dtype_twotorch.float64,short,False,8.301
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.bool,short,False,7.401
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.uint8,short,False,7.843
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int8,short,False,7.117
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int16,short,False,7.170
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int32,short,False,8.000
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.int64,short,False,9.284
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float16,short,False,7.179
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.bfloat16,short,False,7.645
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float32,short,False,7.988
-PyTorch,TensorConversionBenchmark,TensorConversionBenchmark_M32_N128_cpu_dtype_onetorch.float64_dtype_twotorch.float64,short,False,0.792
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N16_cpu,short,FALSE,5.0499
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N64_cpu,short,FALSE,5.3229
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M8_N128_cpu,short,FALSE,5.4418
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N16_cpu,short,FALSE,5.0868
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N64_cpu,short,FALSE,5.4495
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M16_N128_cpu,short,FALSE,5.5578
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N16_cpu,short,FALSE,5.2631
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N64_cpu,short,FALSE,5.5646
+PyTorch,FloatToHalfTensorConversionBenchmark,FloatToHalfTensorConversionBenchmark_M32_N128_cpu,short,FALSE,5.7898
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N16_cpu,short,FALSE,5.0228
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N64_cpu,short,FALSE,5.3692
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M8_N128_cpu,short,FALSE,5.4006
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N16_cpu,short,FALSE,5.1107
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N64_cpu,short,FALSE,5.4119
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M16_N128_cpu,short,FALSE,5.5583
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N16_cpu,short,FALSE,5.3818
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N64_cpu,short,FALSE,5.5742
+PyTorch,HalfToFloatTensorConversionBenchmark,HalfToFloatTensorConversionBenchmark_M32_N128_cpu,short,FALSE,6.8414
 PyTorch,relu,"relu_dims(3,4,5)_contigFALSE_inplaceFALSE_dtypetorch.quint8",short,FALSE,9.4657
 PyTorch,relu,"relu_dims(3,4,5)_contigFALSE_inplaceFALSE_dtypetorch.qint8",short,FALSE,9.4625
 PyTorch,relu,"relu_dims(3,4,5)_contigFALSE_inplaceFALSE_dtypetorch.qint32",short,FALSE,9.4165

benchmarks/sparse/spmm.py

@@ -83,13 +83,10 @@ if __name__ == "__main__":
 
     if args.outfile == "stdout":
         outfile = sys.stdout
-        need_close = False
     elif args.outfile == "stderr":
         outfile = sys.stderr
-        need_close = False
     else:
         outfile = open(args.outfile, "a")
-        need_close = True
 
     test_count = args.test_count
     m = args.m
@@ -150,5 +147,3 @@ if __name__ == "__main__":
             time,
             file=outfile,
         )
-    if need_close:
-        outfile.close()

benchmarks/sparse/spmv.py

@@ -82,13 +82,10 @@ if __name__ == "__main__":
 
     if args.outfile == "stdout":
         outfile = sys.stdout
-        need_close = False
     elif args.outfile == "stderr":
         outfile = sys.stderr
-        need_close = False
     else:
         outfile = open(args.outfile, "a")
-        need_close = True
 
     test_count = args.test_count
     m = args.m
@@ -135,5 +132,3 @@ if __name__ == "__main__":
             time_csr,
             file=outfile,
         )
-    if need_close:
-        outfile.close()

benchmarks/sparse/triton_ops.py

@@ -179,13 +179,10 @@ if __name__ == "__main__":
 
     if args.outfile == "stdout":
         outfile = sys.stdout
-        need_close = False
     elif args.outfile == "stderr":
         outfile = sys.stderr
-        need_close = False
     else:
         outfile = open(args.outfile, "a")
-        need_close = True
 
     ops = args.ops.split(",")
 
@@ -437,5 +434,3 @@ if __name__ == "__main__":
                 if op not in {"bsr_scatter_mm6", "bsr_dense_mm_with_meta"}:
                     # Break on operations that do not consume parameters
                     break
-    if need_close:
-        outfile.close()

c10/core/CachingDeviceAllocator.h

@@ -96,6 +96,10 @@ struct C10_API DeviceAllocator : public c10::Allocator {
 
   // Resets peak memory usage statistics for the specified device
   virtual void resetPeakStats(c10::DeviceIndex device) = 0;
+
+  // Return the free memory size and total memory size in bytes for the
+  // specified device.
+  virtual std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) = 0;
 };
 
 // This function is used to get the DeviceAllocator for a specific device type

c10/core/ScalarType.h

@@ -27,7 +27,6 @@
 #include <torch/headeronly/core/ScalarType.h>
 
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
 
 namespace c10 {
 
@@ -206,12 +205,6 @@ inline bool isSignedType(ScalarType t) {
       break;
       // Do not add default here, but rather define behavior of every new entry
       // here.  `-Wswitch-enum` would raise a warning in those cases.
-      // TODO: get PyTorch to adopt exhaustive switches by default with a way to
-      // opt specific switches to being non-exhaustive.
-      // Exhaustive:
-      // `-Wswitch-enum`, `-Wswitch-default`, `-Wno-covered-switch-default`
-      // Non-Exhaustive:
-      // `-Wno-switch-enum`, `-Wswitch-default`, `-Wcovered-switch-default`
   }
   TORCH_CHECK(false, "Unknown ScalarType ", t);
 #undef CASE_ISSIGNED

c10/core/TensorImpl.h

@@ -57,8 +57,6 @@ C10_DECLARE_bool(caffe2_keep_on_shrink);
 // respect caffe2_keep_on_shrink.
 C10_DECLARE_int64(caffe2_max_keep_on_shrink_memory);
 
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default")
-
 namespace at {
 class Tensor;
 class TensorBase;
@@ -3305,5 +3303,3 @@ static_assert(
 #undef C10_GCC_VERSION_MINOR
 
 } // namespace c10
-
-C10_DIAGNOSTIC_POP()

c10/cuda/CUDACachingAllocator.h

@@ -345,6 +345,13 @@ class CUDAAllocator : public DeviceAllocator {
       c10::DeviceIndex device,
       std::shared_ptr<AllocatorState> pps) = 0;
   virtual std::string name() = 0;
+  std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) override {
+    c10::DeviceGuard device_guard({at::kCUDA, device});
+    size_t free = 0;
+    size_t total = 0;
+    C10_CUDA_CHECK(cudaMemGetInfo(&free, &total));
+    return {free, total};
+  }
 };
 
 // Allocator object, statically initialized

c10/cuda/CUDADeviceAssertionHost.cpp

@@ -295,19 +295,11 @@ DeviceAssertionsData* CUDAKernelLaunchRegistry::
   C10_CUDA_CHECK_WO_DSA(
       cudaMallocManaged(&uvm_assertions_ptr, sizeof(DeviceAssertionsData)));
 
-#if CUDART_VERSION >= 13000
-  cudaMemLocation cpuDevice;
-  cpuDevice.type = cudaMemLocationTypeDevice;
-  cpuDevice.id = cudaCpuDeviceId;
-#else
-  const auto cpuDevice = cudaCpuDeviceId;
-#endif
-
   C10_CUDA_CHECK_WO_DSA(cudaMemAdvise(
       uvm_assertions_ptr,
       sizeof(DeviceAssertionsData),
       cudaMemAdviseSetPreferredLocation,
-      cpuDevice));
+      cudaCpuDeviceId));
 
   // GPU will establish direct mapping of data in CPU memory, no page faults
   // will be generated
@@ -315,7 +307,7 @@ DeviceAssertionsData* CUDAKernelLaunchRegistry::
       uvm_assertions_ptr,
       sizeof(DeviceAssertionsData),
       cudaMemAdviseSetAccessedBy,
-      cpuDevice));
+      cudaCpuDeviceId));
 
   // Initialize the memory from the CPU; otherwise, pages may have to be created
   // on demand. We think that UVM documentation indicates that first access may

c10/metal/error.h

@@ -1,111 +0,0 @@
-#pragma once
-#include <c10/metal/common.h>
-
-namespace c10 {
-namespace metal {
-C10_METAL_CONSTEXPR unsigned error_message_count = 30;
-struct ErrorMessage {
-  char file[128];
-  char func[128];
-  char message[250];
-  unsigned int line;
-};
-
-struct ErrorMessages {
-#ifdef __METAL__
-  ::metal::atomic<unsigned int> count;
-#else
-  unsigned int count;
-#endif
-  ErrorMessage msg[error_message_count];
-};
-
-#ifdef __METAL__
-namespace detail {
-static uint strncpy(device char* dst, constant const char* src, unsigned len) {
-  uint i = 0;
-  while (src[i] != 0 && i < len - 1) {
-    dst[i] = src[i];
-    i++;
-  }
-  dst[i] = 0;
-  return i;
-}
-
-inline uint print_arg(
-    device char* ptr,
-    unsigned len,
-    constant const char* arg) {
-  return strncpy(ptr, arg, len);
-}
-
-// Returns number length as string in base10
-static inline uint base10_length(long num) {
-  uint rc = 1;
-  if (num < 0) {
-    num = -num;
-    rc += 1;
-  }
-  while (num > 9) {
-    num /= 10;
-    rc++;
-  }
-  return rc;
-}
-
-// Converts signed integer to string
-inline uint print_arg(device char* ptr, unsigned len, long arg) {
-  const auto arg_len = base10_length(arg);
-  if (arg_len >= len)
-    return 0;
-  if (arg < 0) {
-    ptr[0] = '-';
-    arg = -arg;
-  }
-  uint idx = 1;
-  do {
-    ptr[arg_len - idx] = '0' + (arg % 10);
-    arg /= 10;
-    idx++;
-  } while (arg > 0);
-  ptr[arg_len] = 0;
-  return arg_len;
-}
-
-template <typename T>
-inline void print_args(device char* ptr, unsigned len, T arg) {
-  print_arg(ptr, len, arg);
-}
-
-template <typename T, typename... Args>
-inline void print_args(device char* ptr, unsigned len, T arg, Args... args) {
-  const auto rc = print_arg(ptr, len, arg);
-  print_args(ptr + rc, len - rc, args...);
-}
-
-} // namespace detail
-
-template <typename... Args>
-static void report_error(
-    device ErrorMessages* msgs,
-    constant const char* file,
-    int line,
-    constant const char* func,
-    Args... args) {
-  const auto idx =
-      atomic_fetch_add_explicit(&msgs->count, 1, ::metal::memory_order_relaxed);
-  if (idx >= error_message_count) {
-    return;
-  }
-  device auto* msg = &msgs->msg[idx];
-  detail::strncpy(msg->file, file, 128);
-  detail::strncpy(msg->func, func, 128);
-  detail::print_args(msg->message, 250, args...);
-  msg->line = line;
-}
-
-#define TORCH_REPORT_ERROR(buf, ...) \
-  ::c10::metal::report_error(buf, __FILE__, __LINE__, __func__, __VA_ARGS__)
-#endif
-} // namespace metal
-} // namespace c10

c10/test/util/Metaprogramming_test.cpp

@@ -1,8 +1,9 @@
+#include <c10/test/util/Macros.h>
+#include <c10/util/Metaprogramming.h>
 #include <gtest/gtest.h>
-#include <torch/headeronly/util/Metaprogramming.h>
 #include <cstdlib>
 
-using namespace torch::headeronly::guts;
+using namespace c10::guts;
 
 // NOLINTBEGIN(modernize*, cppcoreguidelines-special-member-functions)
 namespace {
@@ -64,15 +65,6 @@ static_assert(
         typename make_function_traits_t<void, typelist::typelist<int, float>>::
             func_type>::value,
     "");
-
-struct Functor final {
-  std::string operator()(int64_t a, float b) const;
-};
-static_assert(
-    std::is_same<
-        std::string(int64_t, float),
-        typename infer_function_traits_t<Functor>::func_type>::value,
-    "");
 } // namespace test_function_traits
 
 struct MovableOnly {

c10/test/util/TypeList_test.cpp

@@ -1,8 +1,8 @@
+#include <c10/util/TypeList.h>
 #include <gtest/gtest.h>
-#include <torch/headeronly/util/TypeList.h>
 #include <memory>
 
-using namespace torch::headeronly::guts::typelist;
+using namespace c10::guts::typelist;
 // NOLINTBEGIN(modernize-unary-static-assert)
 namespace test_size {
 class MyClass {};

c10/test/util/TypeTraits_test.cpp

@@ -1,7 +1,7 @@
+#include <c10/util/TypeTraits.h>
 #include <gtest/gtest.h>
-#include <torch/headeronly/util/TypeTraits.h>
 
-using namespace torch::headeronly::guts;
+using namespace c10::guts;
 
 // NOLINTBEGIN(modernize-unary-static-assert)
 namespace {

c10/util/Metaprogramming.cpp

@@ -0,0 +1 @@
+#include <c10/util/Metaprogramming.h>

c10/util/Metaprogramming.h

@@ -1 +1,224 @@
-#include <torch/headeronly/util/Metaprogramming.h>
+#pragma once
+
+#include <c10/util/TypeList.h>
+#include <type_traits>
+
+namespace c10::guts {
+
+/**
+ * Access information about result type or arguments from a function type.
+ * Example:
+ * using A = function_traits<int (float, double)>::return_type // A == int
+ * using A = function_traits<int (float, double)>::parameter_types::tuple_type
+ * // A == tuple<float, double>
+ */
+template <class Func>
+struct function_traits {
+  static_assert(
+      !std::is_same_v<Func, Func>,
+      "In function_traits<Func>, Func must be a plain function type.");
+};
+template <class Result, class... Args>
+struct function_traits<Result(Args...)> {
+  using func_type = Result(Args...);
+  using return_type = Result;
+  using parameter_types = typelist::typelist<Args...>;
+  static constexpr auto number_of_parameters = sizeof...(Args);
+};
+
+/**
+ * infer_function_traits: creates a `function_traits` type for a simple
+ * function (pointer) or functor (lambda/struct). Currently does not support
+ * class methods.
+ */
+
+template <typename Functor>
+struct infer_function_traits {
+  using type = function_traits<
+      c10::guts::detail::strip_class_t<decltype(&Functor::operator())>>;
+};
+
+template <typename Result, typename... Args>
+struct infer_function_traits<Result (*)(Args...)> {
+  using type = function_traits<Result(Args...)>;
+};
+
+template <typename Result, typename... Args>
+struct infer_function_traits<Result(Args...)> {
+  using type = function_traits<Result(Args...)>;
+};
+
+template <typename T>
+using infer_function_traits_t = typename infer_function_traits<T>::type;
+
+/**
+ * make_function_traits: creates a `function_traits` type given a Return type
+ * and a typelist of Argument types
+ *
+ * Example:
+ * bool f(int, int);
+ *
+ * infer_function_traits_t<f> == make_function_traits_t<bool,
+ * typelist::typelist<int, int>>
+ */
+template <typename Result, typename ArgList>
+struct make_function_traits {
+  static_assert(
+      false_t<ArgList>::value,
+      "In guts::make_function_traits<Result, TypeList>, the ArgList argument must be typelist<...>.");
+};
+
+template <typename Result, typename... Args>
+struct make_function_traits<Result, typelist::typelist<Args...>> {
+  using type = function_traits<Result(Args...)>;
+};
+
+template <typename Result, typename ArgList>
+using make_function_traits_t =
+    typename make_function_traits<Result, ArgList>::type;
+
+/**
+ * make_offset_index_sequence<Start, N>
+ * Like make_index_sequence<N>, but starting from Start instead of 0.
+ *
+ * Example:
+ *  make_offset_index_sequence<10, 3> == std::index_sequence<10, 11, 12>
+ */
+template <size_t Start, size_t N, size_t... Is>
+struct make_offset_index_sequence_impl
+    : make_offset_index_sequence_impl<Start, N - 1, Start + N - 1, Is...> {
+  static_assert(
+      static_cast<int>(Start) >= 0,
+      "make_offset_index_sequence: Start < 0");
+  static_assert(static_cast<int>(N) >= 0, "make_offset_index_sequence: N < 0");
+};
+
+template <size_t Start, size_t... Is>
+struct make_offset_index_sequence_impl<Start, 0, Is...> {
+  typedef std::index_sequence<Is...> type;
+};
+
+template <size_t Start, size_t N>
+using make_offset_index_sequence =
+    typename make_offset_index_sequence_impl<Start, N>::type;
+
+/**
+ * Use tuple_elements to extract a position-indexed subset of elements
+ * from the argument tuple into a result tuple.
+ *
+ * Example:
+ *  std::tuple<int, const char*, double> t = std::make_tuple(0, "HEY", 2.0);
+ *  std::tuple<int, double> result = tuple_elements(t, std::index_sequence<0,
+ * 2>());
+ */
+template <class Tuple, size_t... Is>
+constexpr auto tuple_elements(Tuple t, std::index_sequence<Is...> /*unused*/) {
+  return std::tuple<std::tuple_element_t<Is, Tuple>...>(std::get<Is>(t)...);
+}
+
+/**
+ * Use tuple_take to extract the first or last n elements from the argument
+ * tuple into a result tuple.
+ *
+ * Example:
+ *  std::tuple<int, const char*, double> t = std::make_tuple(0, "HEY", 2.0);
+ *  std::tuple<int, const char*> first_two = tuple_take<decltype(t), 2>(t);
+ *  std::tuple<const char*, double> last_two = tuple_take<decltype(t), -2>(t);
+ */
+template <class Tuple, int N, class Enable = void>
+struct TupleTake {};
+
+template <class Tuple, int N>
+struct TupleTake<Tuple, N, std::enable_if_t<N >= 0, void>> {
+  static auto call(Tuple t) {
+    constexpr size_t size = std::tuple_size<Tuple>();
+    static_assert(N <= size, "tuple_take: N > size");
+    return tuple_elements(t, std::make_index_sequence<N>{});
+  }
+};
+
+template <class Tuple, int N>
+    struct TupleTake < Tuple,
+    N, std::enable_if_t<N<0, void>> {
+  static auto call(Tuple t) {
+    constexpr size_t size = std::tuple_size<Tuple>();
+    static_assert(-N <= size, "tuple_take: -N > size");
+    return tuple_elements(t, make_offset_index_sequence<size + N, -N>{});
+  }
+};
+
+template <class Tuple, int N>
+auto tuple_take(Tuple t) {
+  return TupleTake<Tuple, N>::call(t);
+}
+
+/**
+ * Use tuple_slice to extract a contiguous subtuple from the argument.
+ *
+ * Example:
+ *  std::tuple<int, const char*, double, bool> t = std::make_tuple(0,
+ * "HEY", 2.0, false); std::tuple<int, const char*> middle_two =
+ * tuple_slice<decltype(t), 1, 2>(t);
+ */
+template <class Tuple, size_t Start, size_t N>
+constexpr auto tuple_slice(Tuple t) {
+  constexpr size_t size = std::tuple_size<Tuple>();
+  static_assert(Start + N <= size, "tuple_slice: Start + N > size");
+  return tuple_elements(t, make_offset_index_sequence<Start, N>{});
+}
+
+/**
+ * Use tuple_map to run a mapping function over a tuple to get a new tuple.
+ *
+ * Example 1:
+ *   auto result = tuple_map(std::tuple<int32_t, int32_t, int32_t>(3, 4, 5), []
+ * (int32_t a) -> int16_t {return a+1;});
+ *   // result == std::tuple<int16_t, int16_t, int16_t>(4, 5, 6)
+ *
+ * Example 2:
+ *   struct Mapper {
+ *     std::string operator()(int32_t a) const {
+ *       return std::to_string(a);
+ *     }
+ *     int64_t operator()(const std::string& a) const {
+ *        return atoi(a.c_str());
+ *     }
+ *   };
+ *   auto result = tuple_map(std::tuple<int32_t, std::string>(3, "4"),
+ * Mapper());
+ *   // result == std::tuple<std::string, int64_t>("3", 4)
+ *
+ * Example 3:
+ *   struct A final {
+ *    int32_t func() {
+ *      return 5;
+ *    }
+ *  };
+ *  struct B final {
+ *    std::string func() {
+ *      return "5";
+ *    }
+ *  };
+ *  auto result = tuple_map(std::make_tuple(A(), B()), [] (auto a) { return
+ * a.func(); });
+ *  // result == std::tuple<int32_t, std::string>(5, "5");
+ */
+namespace detail {
+template <class Mapper, class... Args, size_t... Indices>
+auto tuple_map(
+    // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved)
+    std::tuple<Args...>&& tuple,
+    const Mapper& mapper,
+    std::index_sequence<Indices...> /*unused*/) {
+  return std::tuple<decltype(mapper(std::forward<Args>(std::get<Indices>(
+      tuple))))...>(mapper(std::forward<Args>(std::get<Indices>(tuple)))...);
+}
+} // namespace detail
+
+template <class Mapper, class... Args>
+auto tuple_map(std::tuple<Args...>&& tuple, const Mapper& mapper) {
+  return detail::tuple_map(
+      std::move(tuple), mapper, std::index_sequence_for<Args...>());
+}
+
+} // namespace c10::guts

c10/util/TypeList.h

@@ -1 +1,515 @@
-#include <torch/headeronly/util/TypeList.h>
+#pragma once
+
+#include <c10/util/TypeTraits.h>
+#include <algorithm>
+#include <cstddef>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+
+namespace c10::guts {
+
+template <class... T>
+struct false_t : std::false_type {};
+template <template <class> class... T>
+struct false_higher_t : std::false_type {};
+
+namespace typelist {
+
+/**
+ * Type holding a list of types for compile time type computations
+ */
+template <class... Items>
+struct typelist final {
+ public:
+  typelist() = delete; // not for instantiation
+};
+
+/**
+ * Returns the number of types in a typelist
+ * Example:
+ *   3  ==  size<typelist<int, int, double>>::value
+ */
+template <class TypeList>
+struct size final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::size<T>, T must be typelist<...>.");
+};
+template <class... Types>
+struct size<typelist<Types...>> final {
+  static constexpr size_t value = sizeof...(Types);
+};
+
+/**
+ * Transforms a list of types into a tuple holding these types.
+ * Example:
+ *   std::tuple<int, string>  ==  to_tuple_t<typelist<int, string>>
+ */
+template <class TypeList>
+struct to_tuple final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::to_tuple<T>, T must be typelist<...>.");
+};
+template <class... Types>
+struct to_tuple<typelist<Types...>> final {
+  using type = std::tuple<Types...>;
+};
+template <class TypeList>
+using to_tuple_t = typename to_tuple<TypeList>::type;
+
+/**
+ * Creates a typelist containing the types of a given tuple.
+ * Example:
+ *   typelist<int, string>  ==  from_tuple_t<std::tuple<int, string>>
+ */
+template <class Tuple>
+struct from_tuple final {
+  static_assert(
+      false_t<Tuple>::value,
+      "In typelist::from_tuple<T>, T must be std::tuple<...>.");
+};
+template <class... Types>
+struct from_tuple<std::tuple<Types...>> final {
+  using type = typelist<Types...>;
+};
+template <class Tuple>
+using from_tuple_t = typename from_tuple<Tuple>::type;
+
+/**
+ * Concatenates multiple type lists.
+ * Example:
+ *   typelist<int, string, int>  ==  concat_t<typelist<int, string>,
+ * typelist<int>>
+ */
+template <class... TypeLists>
+struct concat final {
+  static_assert(
+      false_t<TypeLists...>::value,
+      "In typelist::concat<T1, ...>, the T arguments each must be typelist<...>.");
+};
+template <class... Head1Types, class... Head2Types, class... TailLists>
+struct concat<typelist<Head1Types...>, typelist<Head2Types...>, TailLists...>
+    final {
+  using type =
+      typename concat<typelist<Head1Types..., Head2Types...>, TailLists...>::
+          type;
+};
+template <class... HeadTypes>
+struct concat<typelist<HeadTypes...>> final {
+  using type = typelist<HeadTypes...>;
+};
+template <>
+struct concat<> final {
+  using type = typelist<>;
+};
+template <class... TypeLists>
+using concat_t = typename concat<TypeLists...>::type;
+
+/**
+ * Filters the types in a type list by a type trait.
+ * Examples:
+ *   typelist<int&, const string&&>  ==  filter_t<std::is_reference,
+ * typelist<void, string, int&, bool, const string&&, int>>
+ */
+template <template <class> class Condition, class TypeList>
+struct filter final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::filter<Condition, TypeList>, the TypeList argument must be typelist<...>.");
+};
+template <template <class> class Condition, class Head, class... Tail>
+struct filter<Condition, typelist<Head, Tail...>> final {
+  static_assert(
+      is_type_condition<Condition>::value,
+      "In typelist::filter<Condition, TypeList>, the Condition argument must be a condition type trait, i.e. have a static constexpr bool ::value member.");
+  using type = std::conditional_t<
+      Condition<Head>::value,
+      concat_t<
+          typelist<Head>,
+          typename filter<Condition, typelist<Tail...>>::type>,
+      typename filter<Condition, typelist<Tail...>>::type>;
+};
+template <template <class> class Condition>
+struct filter<Condition, typelist<>> final {
+  static_assert(
+      is_type_condition<Condition>::value,
+      "In typelist::filter<Condition, TypeList>, the Condition argument must be a condition type trait, i.e. have a static constexpr bool ::value member.");
+  using type = typelist<>;
+};
+template <template <class> class Condition, class TypeList>
+using filter_t = typename filter<Condition, TypeList>::type;
+
+/**
+ * Counts how many types in the list fulfill a type trait
+ * Examples:
+ *   2  ==  count_if<std::is_reference, typelist<void, string, int&, bool, const
+ * string&&, int>>
+ */
+template <template <class> class Condition, class TypeList>
+struct count_if final {
+  static_assert(
+      is_type_condition<Condition>::value,
+      "In typelist::count_if<Condition, TypeList>, the Condition argument must be a condition type trait, i.e. have a static constexpr bool ::value member.");
+  static_assert(
+      is_instantiation_of<typelist, TypeList>::value,
+      "In typelist::count_if<Condition, TypeList>, the TypeList argument must be typelist<...>.");
+  // TODO Direct implementation might be faster
+  static constexpr size_t value = size<filter_t<Condition, TypeList>>::value;
+};
+
+/**
+ * Checks if a typelist contains a certain type.
+ * Examples:
+ *  contains<typelist<int, string>, string> == true_type
+ *  contains<typelist<int, string>, double> == false_type
+ */
+namespace detail {
+template <class TypeList, class Type, class Enable = void>
+struct contains {};
+template <class Type>
+struct contains<typelist<>, Type, void> : std::false_type {};
+template <class Type, class Head, class... Tail>
+struct contains<
+    typelist<Head, Tail...>,
+    Type,
+    std::enable_if_t<std::is_same_v<Head, Type>>> : std::true_type {};
+template <class Type, class Head, class... Tail>
+struct contains<
+    typelist<Head, Tail...>,
+    Type,
+    std::enable_if_t<!std::is_same_v<Head, Type>>>
+    : contains<typelist<Tail...>, Type> {};
+} // namespace detail
+template <class TypeList, class Type>
+using contains = typename detail::contains<TypeList, Type>::type;
+
+/**
+ * Returns true iff the type trait is true for all types in the type list
+ * Examples:
+ *   true   ==  all<std::is_reference, typelist<int&, const float&&, const
+ * MyClass&>>::value false  ==  all<std::is_reference, typelist<int&, const
+ * float&&, MyClass>>::value
+ */
+template <template <class> class Condition, class TypeList>
+struct all {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::all<Condition, TypeList>, the TypeList argument must be typelist<...>.");
+};
+template <template <class> class Condition, class... Types>
+struct all<Condition, typelist<Types...>>
+    : std::conjunction<Condition<Types>...> {
+  static_assert(
+      is_type_condition<Condition>::value,
+      "In typelist::all<Condition, TypeList>, the Condition argument must be a condition type trait, i.e. have a static constexpr bool ::value member.");
+};
+
+/**
+ * Returns true iff the type trait is true for any type in the type list
+ * Examples:
+ *   true   ==  true_for_any_type<std::is_reference, typelist<int, const
+ * float&&, const MyClass>>::value false  ==
+ * true_for_any_type<std::is_reference, typelist<int, const float,
+ * MyClass>>::value
+ */
+template <template <class> class Condition, class TypeList>
+struct true_for_any_type final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::true_for_any_type<Condition, TypeList>, the TypeList argument must be typelist<...>.");
+};
+template <template <class> class Condition, class... Types>
+struct true_for_any_type<Condition, typelist<Types...>> final
+    : std::disjunction<Condition<Types>...> {
+  static_assert(
+      is_type_condition<Condition>::value,
+      "In typelist::true_for_any_type<Condition, TypeList>, the Condition argument must be a condition type trait, i.e. have a static constexpr bool ::value member.");
+};
+
+/**
+ * Maps types of a type list using a type trait
+ * Example:
+ *  typelist<int&, double&, string&>  ==  map_t<std::add_lvalue_reference_t,
+ * typelist<int, double, string>>
+ */
+template <template <class> class Mapper, class TypeList>
+struct map final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::map<Mapper, TypeList>, the TypeList argument must be typelist<...>.");
+};
+template <template <class> class Mapper, class... Types>
+struct map<Mapper, typelist<Types...>> final {
+  using type = typelist<Mapper<Types>...>;
+};
+template <template <class> class Mapper, class TypeList>
+using map_t = typename map<Mapper, TypeList>::type;
+
+/**
+ * Returns the first element of a type list.
+ * Example:
+ *   int  ==  head_t<typelist<int, string>>
+ */
+template <class TypeList>
+struct head final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::head<T>, the T argument must be typelist<...>.");
+};
+template <class Head, class... Tail>
+struct head<typelist<Head, Tail...>> final {
+  using type = Head;
+};
+template <class TypeList>
+using head_t = typename head<TypeList>::type;
+
+/**
+ * Returns the first element of a type list, or the specified default if the
+ * type list is empty. Example: int  ==  head_t<bool, typelist<int, string>>
+ *   bool  ==  head_t<bool, typelist<>>
+ */
+template <class Default, class TypeList>
+struct head_with_default final {
+  using type = Default;
+};
+template <class Default, class Head, class... Tail>
+struct head_with_default<Default, typelist<Head, Tail...>> final {
+  using type = Head;
+};
+template <class Default, class TypeList>
+using head_with_default_t = typename head_with_default<Default, TypeList>::type;
+
+/**
+ * Returns the N-th element of a type list.
+ * Example:
+ * int == element_t<1, typelist<float, int, char>>
+ */
+
+/// Base template.
+template <size_t Index, class TypeList>
+struct element final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::element<T>, the T argument must be typelist<...>.");
+};
+
+/// Successful case, we have reached the zero index and can "return" the head
+/// type.
+template <class Head, class... Tail>
+struct element<0, typelist<Head, Tail...>> {
+  using type = Head;
+};
+
+/// Error case, we have an index but ran out of types! It will only be selected
+/// if `Ts...` is actually empty!
+template <size_t Index, class... Ts>
+struct element<Index, typelist<Ts...>> {
+  static_assert(
+      Index < sizeof...(Ts),
+      "Index is out of bounds in typelist::element");
+};
+
+/// Shave off types until we hit the <0, Head, Tail...> or <Index> case.
+template <size_t Index, class Head, class... Tail>
+struct element<Index, typelist<Head, Tail...>>
+    : element<Index - 1, typelist<Tail...>> {};
+
+/// Convenience alias.
+template <size_t Index, class TypeList>
+using element_t = typename element<Index, TypeList>::type;
+
+/**
+ * Returns the last element of a type list.
+ * Example:
+ *   int  ==  last_t<typelist<int, string>>
+ */
+template <class TypeList>
+struct last final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::last<T>, the T argument must be typelist<...>.");
+};
+template <class Head, class... Tail>
+struct last<typelist<Head, Tail...>> final {
+  using type = typename last<typelist<Tail...>>::type;
+};
+template <class Head>
+struct last<typelist<Head>> final {
+  using type = Head;
+};
+template <class TypeList>
+using last_t = typename last<TypeList>::type;
+static_assert(std::is_same_v<int, last_t<typelist<double, float, int>>>);
+
+/**
+ * Take/drop a number of arguments from a typelist.
+ * Example:
+ *   typelist<int, string> == take_t<typelist<int, string, bool>, 2>
+ *   typelist<bool> == drop_t<typelist<int, string, bool>, 2>
+ */
+namespace detail {
+template <class TypeList, size_t offset, class IndexSequence>
+struct take_elements final {};
+
+template <class TypeList, size_t offset, size_t... Indices>
+struct take_elements<TypeList, offset, std::index_sequence<Indices...>> final {
+  using type = typelist<typename element<offset + Indices, TypeList>::type...>;
+};
+} // namespace detail
+
+template <class TypeList, size_t num>
+struct take final {
+  static_assert(
+      is_instantiation_of<typelist, TypeList>::value,
+      "In typelist::take<T, num>, the T argument must be typelist<...>.");
+  static_assert(
+      num <= size<TypeList>::value,
+      "Tried to typelist::take more elements than there are in the list");
+  using type = typename detail::
+      take_elements<TypeList, 0, std::make_index_sequence<num>>::type;
+};
+template <class TypeList, size_t num>
+using take_t = typename take<TypeList, num>::type;
+
+template <class TypeList, size_t num>
+struct drop final {
+  static_assert(
+      is_instantiation_of<typelist, TypeList>::value,
+      "In typelist::drop<T, num>, the T argument must be typelist<...>.");
+  static_assert(
+      num <= size<TypeList>::value,
+      "Tried to typelist::drop more elements than there are in the list");
+  using type = typename detail::take_elements<
+      TypeList,
+      num,
+      std::make_index_sequence<size<TypeList>::value - num>>::type;
+};
+template <class TypeList, size_t num>
+using drop_t = typename drop<TypeList, num>::type;
+
+/**
+ * Like drop, but returns an empty list rather than an assertion error if `num`
+ * is larger than the size of the TypeList.
+ * Example:
+ *   typelist<> == drop_if_nonempty_t<typelist<string, bool>, 2>
+ *   typelist<> == drop_if_nonempty_t<typelist<int, string, bool>, 3>
+ */
+template <class TypeList, size_t num>
+struct drop_if_nonempty final {
+  static_assert(
+      is_instantiation_of<typelist, TypeList>::value,
+      "In typelist::drop<T, num>, the T argument must be typelist<...>.");
+  using type = typename detail::take_elements<
+      TypeList,
+      std::min(num, size<TypeList>::value),
+      std::make_index_sequence<
+          size<TypeList>::value - std::min(num, size<TypeList>::value)>>::type;
+};
+template <class TypeList, size_t num>
+using drop_if_nonempty_t = typename drop_if_nonempty<TypeList, num>::type;
+
+/**
+ * Reverses a typelist.
+ * Example:
+ *   typelist<int, string>  == reverse_t<typelist<string, int>>
+ */
+template <class TypeList>
+struct reverse final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::reverse<T>, the T argument must be typelist<...>.");
+};
+template <class Head, class... Tail>
+struct reverse<typelist<Head, Tail...>> final {
+  using type =
+      concat_t<typename reverse<typelist<Tail...>>::type, typelist<Head>>;
+};
+template <>
+struct reverse<typelist<>> final {
+  using type = typelist<>;
+};
+template <class TypeList>
+using reverse_t = typename reverse<TypeList>::type;
+
+/**
+ * Find the index of the first type in a typelist fulfilling a type trait
+ * condition. Example:
+ *
+ * 2 == find_if<typelist<char, int, char&, int&>, std::is_reference>::value
+ */
+template <class TypeList, template <class> class Condition, class Enable = void>
+struct find_if final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::find_if<TypeList, Condition>, the TypeList argument must be typelist<...>.");
+};
+template <template <class> class Condition>
+struct find_if<typelist<>, Condition, void> final {
+  static_assert(
+      false_higher_t<Condition>::value,
+      "In typelist::find_if<Type/List, Condition>, didn't find any type fulfilling the Condition.");
+};
+template <class Head, class... Tail, template <class> class Condition>
+struct find_if<
+    typelist<Head, Tail...>,
+    Condition,
+    std::enable_if_t<Condition<Head>::value>>
+    final {
+  static constexpr size_t value = 0;
+};
+template <class Head, class... Tail, template <class> class Condition>
+struct find_if<
+    typelist<Head, Tail...>,
+    Condition,
+    std::enable_if_t<!Condition<Head>::value>>
+    final {
+  static constexpr size_t value =
+      1 + find_if<typelist<Tail...>, Condition>::value;
+};
+
+/**
+ * Maps a list of types into a list of values.
+ * Examples:
+ *   // Example 1
+ *   auto sizes =
+ *     map_types_to_values<typelist<int64_t, bool, uint32_t>>(
+ *       [] (auto t) { return sizeof(decltype(t)::type); }
+ *     );
+ *   //  sizes  ==  std::tuple<size_t, size_t, size_t>{8, 1, 4}
+ *
+ *   // Example 2
+ *   auto shared_ptrs =
+ *     map_types_to_values<typelist<int, double>>(
+ *       [] (auto t) { return make_shared<typename decltype(t)::type>(); }
+ *     );
+ *   // shared_ptrs == std::tuple<shared_ptr<int>, shared_ptr<double>>()
+ */
+namespace detail {
+template <class T>
+struct type_ final {
+  using type = T;
+};
+template <class TypeList>
+struct map_types_to_values final {
+  static_assert(
+      false_t<TypeList>::value,
+      "In typelist::map_types_to_values<T>, the T argument must be typelist<...>.");
+};
+template <class... Types>
+struct map_types_to_values<typelist<Types...>> final {
+  template <class Func>
+  static auto call(Func&& func) {
+    return std::tuple{std::forward<Func>(func)(type_<Types>())...};
+  }
+};
+} // namespace detail
+
+template <class TypeList, class Func>
+auto map_types_to_values(Func&& func) {
+  return detail::map_types_to_values<TypeList>::call(std::forward<Func>(func));
+}
+
+} // namespace typelist
+} // namespace c10::guts

c10/util/TypeTraits.h

@@ -1 +1,151 @@
-#include <torch/headeronly/util/TypeTraits.h>
+#pragma once
+
+#include <functional>
+#include <type_traits>
+
+namespace c10::guts {
+
+/**
+ * is_equality_comparable<T> is true_type iff the equality operator is defined
+ * for T.
+ */
+template <class T, class Enable = void>
+struct is_equality_comparable : std::false_type {};
+template <class T>
+struct is_equality_comparable<
+    T,
+    std::void_t<decltype(std::declval<T&>() == std::declval<T&>())>>
+    : std::true_type {};
+template <class T>
+using is_equality_comparable_t = typename is_equality_comparable<T>::type;
+
+/**
+ * is_hashable<T> is true_type iff std::hash is defined for T
+ */
+template <class T, class Enable = void>
+struct is_hashable : std::false_type {};
+template <class T>
+struct is_hashable<T, std::void_t<decltype(std::hash<T>()(std::declval<T&>()))>>
+    : std::true_type {};
+template <class T>
+using is_hashable_t = typename is_hashable<T>::type;
+
+/**
+ * is_function_type<T> is true_type iff T is a plain function type (i.e.
+ * "Result(Args...)")
+ */
+template <class T>
+struct is_function_type : std::false_type {};
+template <class Result, class... Args>
+struct is_function_type<Result(Args...)> : std::true_type {};
+template <class T>
+using is_function_type_t = typename is_function_type<T>::type;
+
+/**
+ * is_instantiation_of<T, I> is true_type iff I is a template instantiation of T
+ * (e.g. vector<int> is an instantiation of vector) Example:
+ *    is_instantiation_of_t<vector, vector<int>> // true
+ *    is_instantiation_of_t<pair, pair<int, string>> // true
+ *    is_instantiation_of_t<vector, pair<int, string>> // false
+ */
+template <template <class...> class Template, class T>
+struct is_instantiation_of : std::false_type {};
+template <template <class...> class Template, class... Args>
+struct is_instantiation_of<Template, Template<Args...>> : std::true_type {};
+template <template <class...> class Template, class T>
+using is_instantiation_of_t = typename is_instantiation_of<Template, T>::type;
+
+namespace detail {
+/**
+ * strip_class: helper to remove the class type from pointers to `operator()`.
+ */
+
+template <typename T>
+struct strip_class {};
+template <typename Class, typename Result, typename... Args>
+struct strip_class<Result (Class::*)(Args...)> {
+  using type = Result(Args...);
+};
+template <typename Class, typename Result, typename... Args>
+struct strip_class<Result (Class::*)(Args...) const> {
+  using type = Result(Args...);
+};
+template <typename T>
+using strip_class_t = typename strip_class<T>::type;
+} // namespace detail
+
+/**
+ * Evaluates to true_type, iff the given class is a Functor
+ * (i.e. has a call operator with some set of arguments)
+ */
+
+template <class Functor, class Enable = void>
+struct is_functor : std::false_type {};
+template <class Functor>
+struct is_functor<
+    Functor,
+    std::enable_if_t<is_function_type<
+        detail::strip_class_t<decltype(&Functor::operator())>>::value>>
+    : std::true_type {};
+
+/**
+ * lambda_is_stateless<T> is true iff the lambda type T is stateless
+ * (i.e. does not have a closure).
+ * Example:
+ *  auto stateless_lambda = [] (int a) {return a;};
+ *  lambda_is_stateless<decltype(stateless_lambda)> // true
+ *  auto stateful_lambda = [&] (int a) {return a;};
+ *  lambda_is_stateless<decltype(stateful_lambda)> // false
+ */
+namespace detail {
+template <class LambdaType, class FuncType>
+struct is_stateless_lambda__ final {
+  static_assert(
+      !std::is_same_v<LambdaType, LambdaType>,
+      "Base case shouldn't be hit");
+};
+// implementation idea: According to the C++ standard, stateless lambdas are
+// convertible to function pointers
+template <class LambdaType, class C, class Result, class... Args>
+struct is_stateless_lambda__<LambdaType, Result (C::*)(Args...) const>
+    : std::is_convertible<LambdaType, Result (*)(Args...)> {};
+template <class LambdaType, class C, class Result, class... Args>
+struct is_stateless_lambda__<LambdaType, Result (C::*)(Args...)>
+    : std::is_convertible<LambdaType, Result (*)(Args...)> {};
+
+// case where LambdaType is not even a functor
+template <class LambdaType, class Enable = void>
+struct is_stateless_lambda_ final : std::false_type {};
+// case where LambdaType is a functor
+template <class LambdaType>
+struct is_stateless_lambda_<
+    LambdaType,
+    std::enable_if_t<is_functor<LambdaType>::value>>
+    : is_stateless_lambda__<LambdaType, decltype(&LambdaType::operator())> {};
+} // namespace detail
+template <class T>
+using is_stateless_lambda = detail::is_stateless_lambda_<std::decay_t<T>>;
+
+/**
+ * is_type_condition<C> is true_type iff C<...> is a type trait representing a
+ * condition (i.e. has a constexpr static bool ::value member) Example:
+ *   is_type_condition<std::is_reference>  // true
+ */
+template <template <class> class C, class Enable = void>
+struct is_type_condition : std::false_type {};
+template <template <class> class C>
+struct is_type_condition<
+    C,
+    std::enable_if_t<
+        std::is_same_v<bool, std::remove_cv_t<decltype(C<int>::value)>>>>
+    : std::true_type {};
+
+/**
+ * is_fundamental<T> is true_type iff the lambda type T is a fundamental type
+ * (that is, arithmetic type, void, or nullptr_t). Example: is_fundamental<int>
+ * // true We define it here to resolve a MSVC bug. See
+ * https://github.com/pytorch/pytorch/issues/30932 for details.
+ */
+template <class T>
+struct is_fundamental : std::is_fundamental<T> {};
+} // namespace c10::guts

c10/xpu/CMakeLists.txt

@@ -24,7 +24,6 @@ set(C10_XPU_HEADERS
     XPUCachingAllocator.h
     XPUDeviceProp.h
     XPUException.h
-    XPUEvent.h
     XPUFunctions.h
     XPUMacros.h
     XPUStream.h

c10/xpu/XPUCachingAllocator.cpp

@@ -926,15 +926,14 @@ class DeviceCachingAllocator {
           (release_cached_blocks() && alloc_block(params, true));
     }
     if (!block_found) {
-      c10::xpu::DeviceProp device_prop;
-      c10::xpu::get_device_properties(&device_prop, device);
-      auto device_total = device_prop.global_mem_size;
+      const auto& raw_device = c10::xpu::get_raw_device(device);
+      const auto device_total =
+          raw_device.get_info<sycl::info::device::global_mem_size>();
       // Estimate the available device memory when the SYCL runtime does not
       // support the corresponding aspect (ext_intel_free_memory).
-      size_t device_free = device_prop.global_mem_size -
+      size_t device_free = device_total -
           stats.reserved_bytes[static_cast<size_t>(StatType::AGGREGATE)]
               .current;
-      auto& raw_device = c10::xpu::get_raw_device(device);
       // TODO: Remove the aspect check once the SYCL runtime bug is fixed on
       // affected devices.
       if (raw_device.has(sycl::aspect::ext_intel_free_memory)) {
@@ -1052,21 +1051,37 @@ class DeviceCachingAllocator {
     }
   }
 
+  std::pair<size_t, size_t> getMemoryInfo() {
+    const auto& device = c10::xpu::get_raw_device(device_index);
+    const size_t total = device.get_info<sycl::info::device::global_mem_size>();
+    TORCH_CHECK(
+        device.has(sycl::aspect::ext_intel_free_memory),
+        "The device (",
+        device.get_info<sycl::info::device::name>(),
+        ") doesn't support querying the available free memory. ",
+        "You can file an issue at https://github.com/pytorch/pytorch/issues ",
+        "to help us prioritize its implementation.");
+    const size_t free =
+        device.get_info<sycl::ext::intel::info::device::free_memory>();
+    return {free, total};
+  }
+
   double getMemoryFraction() {
     if (!set_fraction) {
       return 1.0;
     }
 
-    c10::xpu::DeviceProp device_prop;
-    c10::xpu::get_device_properties(&device_prop, device_index);
+    const auto device_total =
+        xpu::get_raw_device(device_index)
+            .get_info<sycl::info::device::global_mem_size>();
     return static_cast<double>(allowed_memory_maximum) /
-        static_cast<double>(device_prop.global_mem_size);
+        static_cast<double>(device_total);
   }
 
   void setMemoryFraction(double fraction) {
-    c10::xpu::DeviceProp device_prop;
-    c10::xpu::get_device_properties(&device_prop, device_index);
-    auto device_total = device_prop.global_mem_size;
+    const auto device_total =
+        xpu::get_raw_device(device_index)
+            .get_info<sycl::info::device::global_mem_size>();
     allowed_memory_maximum = static_cast<size_t>(fraction * device_total);
     set_fraction = true;
   }
@@ -1240,6 +1255,11 @@ class XPUAllocator : public DeviceAllocator {
         c10::xpu::get_raw_device(dev_to_access));
   }
 
+  std::pair<size_t, size_t> getMemoryInfo(DeviceIndex device) override {
+    assertValidDevice(device);
+    return device_allocators[device]->getMemoryInfo();
+  }
+
   double getMemoryFraction(DeviceIndex device) {
     assertValidDevice(device);
     return device_allocators[device]->getMemoryFraction();

c10/xpu/XPUEvent.h

@@ -1,178 +0,0 @@
-#pragma once
-#include <c10/xpu/XPUStream.h>
-
-namespace c10::xpu {
-
-/*
- * XPUEvent are movable not copyable wrappers around SYCL event. XPUEvent are
- * constructed lazily when first recorded. It has a device, and this device is
- * acquired from the first recording stream. Later streams that record the event
- * must match the same device.
- *
- * Currently, XPUEvent does NOT support to export an inter-process event from
- * another process via inter-process communication(IPC). So it means that
- * inter-process communication for event handles between different processes is
- * not available. This could impact some applications that rely on cross-process
- * synchronization and communication.
- */
-struct XPUEvent {
-  // Constructors
-  XPUEvent(bool enable_timing = false) noexcept
-      : enable_timing_{enable_timing} {}
-
-  ~XPUEvent() {
-    if (isCreated()) {
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_deletion(
-            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-    }
-  }
-
-  C10_DISABLE_COPY_AND_ASSIGN(XPUEvent);
-
-  XPUEvent(XPUEvent&& other) = default;
-  XPUEvent& operator=(XPUEvent&& other) = default;
-
-  operator sycl::event&() const {
-    return event();
-  }
-
-  std::optional<c10::Device> device() const {
-    if (isCreated()) {
-      return c10::Device(c10::kXPU, device_index_);
-    } else {
-      return std::nullopt;
-    }
-  }
-
-  inline bool isCreated() const {
-    return (event_.get() != nullptr);
-  }
-
-  DeviceIndex device_index() const {
-    return device_index_;
-  }
-
-  sycl::event& event() const {
-    return *event_;
-  }
-
-  bool query() const {
-    using namespace sycl::info;
-    if (!isCreated()) {
-      return true;
-    }
-
-    return event().get_info<event::command_execution_status>() ==
-        event_command_status::complete;
-  }
-
-  void record() {
-    record(getCurrentXPUStream());
-  }
-
-  void recordOnce(const XPUStream& stream) {
-    if (!isCreated()) {
-      record(stream);
-    }
-  }
-
-  void record(const XPUStream& stream) {
-    if (!isCreated()) {
-      device_index_ = stream.device_index();
-      assignEvent(stream.queue());
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_creation(
-            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-    } else {
-      TORCH_CHECK(
-          device_index_ == stream.device_index(),
-          "Event device ",
-          device_index_,
-          " does not match recording stream's device ",
-          stream.device_index(),
-          ".");
-      reassignEvent(stream.queue());
-    }
-    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-    if (C10_UNLIKELY(interp)) {
-      (*interp)->trace_gpu_event_record(
-          c10::kXPU,
-          reinterpret_cast<uintptr_t>(event_.get()),
-          reinterpret_cast<uintptr_t>(&stream.queue()));
-    }
-  }
-
-  void block(const XPUStream& stream) {
-    if (isCreated()) {
-      std::vector<sycl::event> event_list{event()};
-      // Make this stream wait until event_ is completed.
-      stream.queue().ext_oneapi_submit_barrier(event_list);
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_wait(
-            c10::kXPU,
-            reinterpret_cast<uintptr_t>(event_.get()),
-            reinterpret_cast<uintptr_t>(&stream.queue()));
-      }
-    }
-  }
-
-  double elapsed_time(const XPUEvent& other) const {
-    TORCH_CHECK(
-        isCreated() && other.isCreated(),
-        "Both events must be recorded before calculating elapsed time.");
-    TORCH_CHECK(
-        query() && other.query(),
-        "Both events must be completed before calculating elapsed time.");
-    TORCH_CHECK(
-        enable_timing_ && other.enable_timing_,
-        "Both events must be created with argument 'enable_timing=True'.");
-
-    using namespace sycl::info::event_profiling;
-    // Block until both of the recorded events are completed.
-    uint64_t end_time_ns = other.event().get_profiling_info<command_end>();
-    uint64_t start_time_ns = event().get_profiling_info<command_end>();
-    // Return the eplased time in milliseconds.
-    return 1e-6 *
-        (static_cast<double>(end_time_ns) - static_cast<double>(start_time_ns));
-  }
-
-  void synchronize() const {
-    if (isCreated()) {
-      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
-      if (C10_UNLIKELY(interp)) {
-        (*interp)->trace_gpu_event_synchronization(
-            c10::kXPU, reinterpret_cast<uintptr_t>(event_.get()));
-      }
-      event().wait_and_throw();
-    }
-  }
-
- private:
-  void assignEvent(sycl::queue& queue) {
-    if (enable_timing_) {
-      event_ = std::make_unique<sycl::event>(
-          sycl::ext::oneapi::experimental::submit_profiling_tag(queue));
-    } else {
-      event_ = std::make_unique<sycl::event>(queue.ext_oneapi_submit_barrier());
-    }
-  }
-
-  void reassignEvent(sycl::queue& queue) {
-    event_.reset();
-    assignEvent(queue);
-  }
-
-  bool enable_timing_ = false;
-  c10::DeviceIndex device_index_ = -1;
-  // Only need to track the last event, as events in an in-order queue are
-  // executed sequentially.
-  std::unique_ptr<sycl::event> event_;
-};
-
-} // namespace c10::xpu

cmake/public/utils.cmake

@@ -478,7 +478,6 @@ function(torch_update_find_cuda_flags)
 endfunction()
 
 include(CheckCXXCompilerFlag)
-include(CheckCCompilerFlag)
 include(CheckLinkerFlag)
 
 ##############################################################################
@@ -502,24 +501,6 @@ function(append_cxx_flag_if_supported flag outputvar)
     endif()
 endfunction()
 
-function(append_c_flag_if_supported flag outputvar)
-    string(TOUPPER "HAS${flag}" _FLAG_NAME)
-    string(REGEX REPLACE "[=-]" "_" _FLAG_NAME "${_FLAG_NAME}")
-
-    # GCC silences unknown -Wno-XXX flags, so test the corresponding -WXXX.
-    if(CMAKE_C_COMPILER_ID STREQUAL "GNU")
-        string(REGEX REPLACE "^Wno-" "W" new_flag "${flag}")
-    else()
-        set(new_flag "${flag}")
-    endif()
-
-    check_c_compiler_flag("${new_flag}" ${_FLAG_NAME})
-    if(${_FLAG_NAME})
-        string(APPEND ${outputvar} " ${flag}")
-        set(${outputvar} "${${outputvar}}" PARENT_SCOPE)
-    endif()
-endfunction()
-
 function(target_compile_options_if_supported target flag)
   set(_compile_options "")
   append_cxx_flag_if_supported("${flag}" _compile_options)

docs/source/accelerator.md

@@ -40,6 +40,7 @@
     :nosignatures:
 
      empty_cache
+     get_memory_info
      max_memory_allocated
      max_memory_reserved
      memory_allocated

docs/source/conf.py

@@ -382,6 +382,20 @@ coverage_ignore_functions = [
     # torch.ao.quantization.backend_config.tensorrt
     "get_tensorrt_backend_config",
     "get_tensorrt_backend_config_dict",
+    # torch.ao.quantization.backend_config.utils
+    "entry_to_pretty_str",
+    "get_fused_module_classes",
+    "get_fuser_method_mapping",
+    "get_fusion_pattern_to_extra_inputs_getter",
+    "get_fusion_pattern_to_root_node_getter",
+    "get_module_to_qat_module",
+    "get_pattern_to_dtype_configs",
+    "get_pattern_to_input_type_to_index",
+    "get_qat_module_classes",
+    "get_root_module_to_quantized_reference_module",
+    "pattern_to_human_readable",
+    "remove_boolean_dispatch_from_name",
+    # torch.ao.quantization.backend_config.x86
     "get_x86_backend_config",
     # torch.ao.quantization.fuse_modules
     "fuse_known_modules",
@@ -412,6 +426,25 @@ coverage_ignore_functions = [
     "insert_observers_for_model",
     "prepare",
     "propagate_dtypes_for_known_nodes",
+    # torch.ao.quantization.fx.utils
+    "all_node_args_except_first",
+    "all_node_args_have_no_tensors",
+    "assert_and_get_unique_device",
+    "collect_producer_nodes",
+    "create_getattr_from_value",
+    "create_node_from_old_node_preserve_meta",
+    "get_custom_module_class_keys",
+    "get_linear_prepack_op_for_dtype",
+    "get_new_attr_name_with_prefix",
+    "get_non_observable_arg_indexes_and_types",
+    "get_qconv_prepack_op",
+    "get_skipped_module_name_and_classes",
+    "graph_module_from_producer_nodes",
+    "maybe_get_next_module",
+    "node_arg_is_bias",
+    "node_arg_is_weight",
+    "return_arg_list",
+    # torch.ao.quantization.pt2e.graph_utils
     "bfs_trace_with_node_process",
     "find_sequential_partitions",
     "get_equivalent_types",
@@ -827,10 +860,80 @@ coverage_ignore_functions = [
     "get_latency_of_one_partition",
     "get_latency_of_partitioned_graph",
     "get_partition_to_latency_mapping",
+    # torch.fx.experimental.proxy_tensor
+    "decompose",
+    "disable_autocast_cache",
+    "disable_proxy_modes_tracing",
+    "dispatch_trace",
+    "extract_val",
+    "fake_signature",
+    "fetch_sym_proxy",
+    "fetch_object_proxy",
+    "get_innermost_proxy_mode",
+    "get_isolated_graphmodule",
+    "get_proxy_slot",
+    "get_torch_dispatch_modes",
+    "has_proxy_slot",
+    "is_sym_node",
+    "maybe_handle_decomp",
+    "proxy_call",
+    "set_meta",
+    "set_original_aten_op",
+    "set_proxy_slot",
+    "snapshot_fake",
+    "thunkify",
+    "track_tensor",
+    "track_tensor_tree",
+    "wrap_key",
+    "wrapper_and_args_for_make_fx",
+    # torch.fx.experimental.recording
     "record_shapeenv_event",
     "replay_shape_env_events",
     "shape_env_check_state_equal",
+    # torch.fx.experimental.sym_node
+    "ceil_impl",
+    "floor_ceil_helper",
+    "floor_impl",
+    "method_to_operator",
+    "sympy_is_channels_last_contiguous_2d",
+    "sympy_is_channels_last_contiguous_3d",
+    "sympy_is_channels_last_strides_2d",
+    "sympy_is_channels_last_strides_3d",
+    "sympy_is_channels_last_strides_generic",
+    "sympy_is_contiguous",
+    "sympy_is_contiguous_generic",
+    "to_node",
+    "wrap_node",
     "sym_sqrt",
+    # torch.fx.experimental.symbolic_shapes
+    "bind_symbols",
+    "cast_symbool_to_symint_guardless",
+    "create_contiguous",
+    "error",
+    "eval_guards",
+    "eval_is_non_overlapping_and_dense",
+    "expect_true",
+    "find_symbol_binding_fx_nodes",
+    "free_symbols",
+    "free_unbacked_symbols",
+    "fx_placeholder_targets",
+    "fx_placeholder_vals",
+    "guard_bool",
+    "guard_float",
+    "guard_int",
+    "guard_scalar",
+    "has_hint",
+    "has_symbolic_sizes_strides",
+    "is_channels_last_contiguous_2d",
+    "is_channels_last_contiguous_3d",
+    "is_channels_last_strides_2d",
+    "is_channels_last_strides_3d",
+    "is_contiguous",
+    "is_non_overlapping_and_dense_indicator",
+    "is_nested_int",
+    "is_symbol_binding_fx_node",
+    "is_symbolic",
+    # torch.fx.experimental.unification.core
     "reify",
     # torch.fx.experimental.unification.match
     "edge",
@@ -868,6 +971,24 @@ coverage_ignore_functions = [
     "reverse_dict",
     # torch.fx.experimental.unification.multipledispatch.variadic
     "isvariadic",
+    # torch.fx.experimental.unification.unification_tools
+    "assoc",
+    "assoc_in",
+    "dissoc",
+    "first",
+    "get_in",
+    "getter",
+    "groupby",
+    "itemfilter",
+    "itemmap",
+    "keyfilter",
+    "keymap",
+    "merge",
+    "merge_with",
+    "update_in",
+    "valfilter",
+    "valmap",
+    # torch.fx.experimental.unification.utils
     "freeze",
     "hashable",
     "raises",

docs/source/fx.experimental.md

@@ -12,37 +12,6 @@ These APIs are experimental and subject to change without notice.
 .. autoclass:: torch.fx.experimental.sym_node.DynamicInt
 ```
 
-## torch.fx.experimental.sym_node
-
-```{eval-rst}
-.. currentmodule:: torch.fx.experimental.sym_node
-```
-
-```{eval-rst}
-.. automodule:: torch.fx.experimental.sym_node
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    is_channels_last_contiguous_2d
-    is_channels_last_contiguous_3d
-    is_channels_last_strides_2d
-    is_channels_last_strides_3d
-    is_contiguous
-    is_non_overlapping_and_dense_indicator
-    method_to_operator
-    sympy_is_channels_last_contiguous_2d
-    sympy_is_channels_last_contiguous_3d
-    sympy_is_channels_last_strides_2d
-    sympy_is_channels_last_strides_3d
-    sympy_is_channels_last_strides_generic
-    sympy_is_contiguous
-    sympy_is_contiguous_generic
-```
-
 ## torch.fx.experimental.symbolic_shapes
 
 ```{eval-rst}
@@ -100,25 +69,6 @@ These APIs are experimental and subject to change without notice.
     rebind_unbacked
     resolve_unbacked_bindings
     is_accessor_node
-    cast_symbool_to_symint_guardless
-    create_contiguous
-    error
-    eval_guards
-    eval_is_non_overlapping_and_dense
-    find_symbol_binding_fx_nodes
-    free_symbols
-    free_unbacked_symbols
-    fx_placeholder_targets
-    fx_placeholder_vals
-    guard_bool
-    guard_float
-    guard_int
-    guard_scalar
-    has_hint
-    has_symbolic_sizes_strides
-    is_nested_int
-    is_symbol_binding_fx_node
-    is_symbolic
 ```
 
 ## torch.fx.experimental.proxy_tensor
@@ -141,46 +91,4 @@ These APIs are experimental and subject to change without notice.
     get_proxy_mode
     maybe_enable_thunkify
     maybe_disable_thunkify
-    decompose
-    disable_autocast_cache
-    disable_proxy_modes_tracing
-    extract_val
-    fake_signature
-    fetch_object_proxy
-    fetch_sym_proxy
-    has_proxy_slot
-    is_sym_node
-    maybe_handle_decomp
-    proxy_call
-    set_meta
-    set_original_aten_op
-    set_proxy_slot
-    snapshot_fake
 ```
-
-## torch.fx.experimental.unification.unification_tools
-
-```{eval-rst}
-.. currentmodule:: torch.fx.experimental.unification.unification_tools
-```
-
-```{eval-rst}
-.. automodule:: torch.fx.experimental.unification.unification_tools
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    assoc
-    assoc_in
-    dissoc
-    first
-    keyfilter
-    keymap
-    merge
-    merge_with
-    update_in
-    valfilter
-    valmap

docs/source/fx.md

@@ -1134,6 +1134,7 @@ The set of leaf modules can be customized by overriding
 .. py:module:: torch.fx.experimental.refinement_types
 .. py:module:: torch.fx.experimental.rewriter
 .. py:module:: torch.fx.experimental.schema_type_annotation
+.. py:module:: torch.fx.experimental.sym_node
 .. py:module:: torch.fx.experimental.unification.core
 .. py:module:: torch.fx.experimental.unification.dispatch
 .. py:module:: torch.fx.experimental.unification.match
@@ -1143,6 +1144,7 @@ The set of leaf modules can be customized by overriding
 .. py:module:: torch.fx.experimental.unification.multipledispatch.dispatcher
 .. py:module:: torch.fx.experimental.unification.multipledispatch.utils
 .. py:module:: torch.fx.experimental.unification.multipledispatch.variadic
+.. py:module:: torch.fx.experimental.unification.unification_tools
 .. py:module:: torch.fx.experimental.unification.utils
 .. py:module:: torch.fx.experimental.unification.variable
 .. py:module:: torch.fx.experimental.unify_refinements

docs/source/mtia.mtia_graph.md

@@ -1,21 +0,0 @@
-# torch.mtia.mtia_graph
-
-The MTIA backend is implemented out of the tree, only interfaces are defined here.
-
-```{eval-rst}
-.. automodule:: torch.mtia.mtia_graph
-```
-
-```{eval-rst}
-.. currentmodule:: torch.mtia.mtia_graph
-```
-
-```{eval-rst}
-.. autoclass:: MTIAGraph
-    :members:
-```
-
-```{eval-rst}
-.. autoclass:: graph
-    :members:
-```

docs/source/nn.attention.rst

@@ -14,10 +14,6 @@ Utils
 
     sdpa_kernel
     SDPBackend
-    register_flash_attention_impl
-    activate_flash_attention_impl
-    list_flash_attention_impls
-    current_flash_attention_impl
 
 Submodules
 ----------

docs/source/pytorch-api.md

@@ -29,7 +29,6 @@ mps
 xpu
 mtia
 mtia.memory
-mtia.mtia_graph
 meta
 torch.backends <backends>
 torch.export <export>

docs/source/quantization-support.md

@@ -134,23 +134,6 @@ Quantization to work with this as well.
     ObservationType
 ```
 
-## torch.ao.quantization.backend_config.utils
-```{eval-rst}
-.. currentmodule:: torch.ao.quantization.backend_config.utils
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-    :template: classtemplate.rst
-
-    entry_to_pretty_str
-    pattern_to_human_readable
-    remove_boolean_dispatch_from_name
-
-```
-
 ## torch.ao.quantization.fx.custom_config
 
 This module contains a few CustomConfig classes that's used in both eager mode and FX graph mode quantization
@@ -171,30 +154,6 @@ This module contains a few CustomConfig classes that's used in both eager mode a
     StandaloneModuleConfigEntry
 ```
 
-## torch.ao.quantization.fx.utils
-
-```{eval-rst}
-.. currentmodule:: torch.ao.quantization.fx.utils
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-    :template: classtemplate.rst
-
-    all_node_args_except_first
-    all_node_args_have_no_tensors
-    collect_producer_nodes
-    create_getattr_from_value
-    create_node_from_old_node_preserve_meta
-    graph_module_from_producer_nodes
-    maybe_get_next_module
-    node_arg_is_bias
-    node_arg_is_weight
-    return_arg_list
-```
-
 ## torch.ao.quantization.quantizer
 
 ```{eval-rst}