Update (base update)

[ghstack-poisoned]

.ci/docker/build.sh

@@ -195,13 +195,16 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3)
+  pytorch-linux-jammy-xpu-n-py3 | pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.2
     NINJA_VERSION=1.9.0
     TRITON=yes
+    if [[ $tag =~ "benchmarks" ]]; then
+      INDUCTOR_BENCHMARKS=yes
+    fi
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10

.ci/docker/common/install_acl.sh

@@ -3,7 +3,7 @@
 
 set -eux
 
-ACL_VERSION=${ACL_VERSION:-"v25.02"}
+ACL_VERSION=${ACL_VERSION:-"v52.6.0"}
 ACL_INSTALL_DIR="/acl"
 
 # Clone ACL

.ci/docker/common/install_conda.sh

@@ -49,12 +49,20 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     export SYSROOT_DEP="sysroot_linux-64=2.17"
   fi
 
+# Install correct Python version
+# Also ensure sysroot is using a modern GLIBC to match system compilers
+if [ "$ANACONDA_PYTHON_VERSION" = "3.14" ]; then
+  as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
+             python="3.14.0" \
+             ${SYSROOT_DEP} \
+             -c conda-forge
+else
   # Install correct Python version
   # Also ensure sysroot is using a modern GLIBC to match system compilers
   as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
              python="$ANACONDA_PYTHON_VERSION" \
              ${SYSROOT_DEP}
-
+fi
   # libstdcxx from conda default channels are too old, we need GLIBCXX_3.4.30
   # which is provided in libstdcxx 12 and up.
   conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.3.24
+NVSHMEM_VERSION=3.4.5
 
 function install_cuda {
   version=$1

.ci/docker/common/install_rocm.sh

@@ -40,11 +40,7 @@ EOF
 
     # Default url values
     rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
-    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
-
-    # Add amdgpu repository
     UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
-    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
 
     # Add rocm repository
     wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -

.ci/docker/common/install_rocm_magma.sh

@@ -12,8 +12,8 @@ function do_install() {
 
     rocm_version_nodot=${rocm_version//./}
 
-    # https://github.com/icl-utk-edu/magma/pull/65
+    # post merge of https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
+    MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
     magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
 
     rocm_dir="/opt/rocm"

.ci/docker/manywheel/build.sh

@@ -97,7 +97,7 @@ case ${image} in
     manylinux2_28-builder:xpu)
         TARGET=xpu_final
         GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=11"
+        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13"
         MANY_LINUX_VERSION="2_28"
         ;;
     *)

.ci/docker/requirements-ci.txt

@@ -138,10 +138,12 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #test_binary_ufuncs.py
 numpy==1.22.4; python_version == "3.10"
 numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
-numpy==2.1.2; python_version >= "3.13"
+numpy==2.1.2; python_version >= "3.13" and python_version < "3.14"
+numpy==2.3.4; python_version >= "3.14"
 
 pandas==2.0.3; python_version < "3.13"
-pandas==2.2.3; python_version >= "3.13"
+pandas==2.2.3; python_version >= "3.13" and python_version < "3.14"
+pandas==2.3.3; python_version >= "3.14"
 
 #onnxruntime
 #Description: scoring engine for Open Neural Network Exchange (ONNX) models
@@ -153,7 +155,8 @@ opt-einsum==3.3
 #Pinned versions: 3.3
 #test that import: test_linalg.py
 
-optree==0.13.0
+optree==0.13.0 ; python_version < "3.14"
+optree==0.17.0 ; python_version >= "3.14"
 #Description: A library for tree manipulation
 #Pinned versions: 0.13.0
 #test that import: test_vmap.py, test_aotdispatch.py, test_dynamic_shapes.py,
@@ -252,7 +255,8 @@ scikit-image==0.22.0
 #test that import:
 
 scipy==1.10.1 ; python_version <= "3.11"
-scipy==1.14.1 ; python_version >= "3.12"
+scipy==1.14.1 ; python_version > "3.11" and python_version < "3.14"
+scipy==1.16.2 ; python_version >= "3.14"
 # Pin SciPy because of failing distribution tests (see #60347)
 #Description: scientific python
 #Pinned versions: 1.10.1
@@ -324,7 +328,8 @@ pywavelets==1.7.0 ; python_version >= "3.12"
 #Pinned versions: 1.4.1
 #test that import:
 
-lxml==5.3.0
+lxml==5.3.0 ; python_version < "3.14"
+lxml==6.0.2 ; python_version >= "3.14"
 #Description: This is a requirement of unittest-xml-reporting
 
 PyGithub==2.3.0
@@ -334,7 +339,9 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:
 
-onnx==1.19.1
+onnx==1.19.1 ; python_version < "3.14"
+# Unpin once Python 3.14 is supported. See  onnxruntime issue 26309.
+onnx==1.18.0 ; python_version == "3.14"
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -359,7 +366,7 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py
 
 # To build PyTorch itself
-pyyaml==6.0.2
+pyyaml==6.0.3
 pyzstd
 setuptools==78.1.1
 packaging==23.1

.ci/docker/ubuntu-xpu/Dockerfile

@@ -54,12 +54,15 @@ ENV OPENSSL_DIR /opt/openssl
 RUN rm install_openssl.sh
 
 ARG INDUCTOR_BENCHMARKS
+ARG ANACONDA_PYTHON_VERSION
+ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
+COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -100,6 +100,8 @@ COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
+# Only build aoti cpp tests when INDUCTOR_BENCHMARKS is set to True
+ENV BUILD_AOT_INDUCTOR_TEST ${INDUCTOR_BENCHMARKS}
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,7 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.9.5"
+    "uv==0.9.6"
 ]
 
 [tool.setuptools]

.ci/magma-rocm/Makefile

@@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash
 
 DOCKER_CMD ?= docker
-DESIRED_ROCM ?= 7.0
+DESIRED_ROCM ?= 7.1
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
@@ -16,6 +16,7 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma-rocm/build_magma.sh
 
 .PHONY: all
+all: magma-rocm71
 all: magma-rocm70
 all: magma-rocm64
 
@@ -24,6 +25,11 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
+.PHONY: magma-rocm71
+magma-rocm71: DESIRED_ROCM := 7.1
+magma-rocm71:
+	$(DOCKER_RUN)
+
 .PHONY: magma-rocm70
 magma-rocm70: DESIRED_ROCM := 7.0
 magma-rocm70:

.ci/magma-rocm/build_magma.sh

@@ -6,8 +6,8 @@ set -eou pipefail
 # The script expects DESIRED_CUDA and PACKAGE_NAME to be set
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 
-# https://github.com/icl-utk-edu/magma/pull/65
+# post merge of https://github.com/icl-utk-edu/magma/pull/65
-MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
+MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
 
 # Folders for the build
 PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
@@ -20,7 +20,7 @@ mkdir -p ${PACKAGE_DIR} ${PACKAGE_OUTPUT}/linux-64 ${PACKAGE_BUILD} ${PACKAGE_RE
 
 # Fetch magma sources and verify checksum
 pushd ${PACKAGE_DIR}
-git clone https://github.com/jeffdaily/magma
+git clone https://github.com/icl-utk-edu/magma
 pushd magma
 git checkout ${MAGMA_VERSION}
 popd

.ci/pytorch/build.sh

@@ -426,7 +426,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]]; then
   # export test times so that potential sharded tests that'll branch off this build will use consistent data
   # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
-  python tools/stats/export_test_times.py
+  PYTHONPATH=. python tools/stats/export_test_times.py
 fi
 # don't do this for bazel or s390x or riscv64 as they don't use sccache
 if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then

.ci/pytorch/test.sh

@@ -460,28 +460,18 @@ test_inductor_shard() {
     --verbose
 }
 
-test_inductor_aoti() {
+test_inductor_aoti_cpp() {
-  # docker build uses bdist_wheel which does not work with test_aot_inductor
-  # TODO: need a faster way to build
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
     # We need to hipify before building again
     python3 tools/amd_build/build_amd.py
   fi
   if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
-    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python -m pip install --no-build-isolation -v -e .)
     # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
   else
-    BUILD_COMMAND=(python -m pip install --no-build-isolation -v -e .)
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}")
   fi
 
-  # aoti cmake custom command requires `torch` to be installed
-  # initialize the cmake build cache and install torch
-  /usr/bin/env "${BUILD_COMMAND[@]}"
-  # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
-  /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
-
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
@@ -582,6 +572,8 @@ fi
 
 if [[ "${TEST_CONFIG}" == *cpu* ]]; then
   DYNAMO_BENCHMARK_FLAGS+=(--device cpu)
+elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
+  DYNAMO_BENCHMARK_FLAGS+=(--device xpu)
 else
   DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
 fi
@@ -675,6 +667,8 @@ test_perf_for_dashboard() {
     device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
+  elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
+    device=xpu
   fi
 
   for mode in "${modes[@]}"; do
@@ -1767,7 +1761,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   else
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
-    if [[ "${TEST_CONFIG}" != *cpu* ]]; then
+    if [[ "${TEST_CONFIG}" != *cpu* && "${TEST_CONFIG}" != *xpu* ]]; then
       install_torchrec_and_fbgemm
     fi
     PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
@@ -1776,7 +1770,7 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
   PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti
+    test_inductor_aoti_cpp
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision

.ci/pytorch/windows/internal/install_python.bat

@@ -7,12 +7,9 @@ if "%DESIRED_PYTHON%" == "3.13t" (
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.13t"
-) else if "%DESIRED_PYTHON%"=="3.14" (
-    echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
 ) else if "%DESIRED_PYTHON%"=="3.14t" (
     echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.14t"
 ) else (

.github/actions/diskspace-cleanup/action.yml

@@ -27,7 +27,9 @@ runs:
             docker system prune -af
             diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
             if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
-                echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
+                diskspace_cutoff_int=$((diskspace_cutoff + 0))
+                difference=$((100 - diskspace_cutoff_int))
+                echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
                 echo "$msg"
                 exit 1
             else

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-69bbe7363897764f9e758d851cd0340147d27f94
+3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-1752fe6809b74921644866275ab80244b96e80bc
+cfbc5c2f1c798991715a6b06bb3ce46478c4487c

.github/merge_rules.yaml

@@ -540,6 +540,26 @@
   - Lint
   - pull
 
+- name: PrivateUse1
+  patterns:
+  - torch/accelerator/**
+  - torch/utils/backend_registration.py
+  - torch/csrc/acc/**
+  - torch/csrc/DeviceAccelerator.*
+  - torch/csrc/profiler/standalone/privateuse1_observer.*
+  - aten/src/ATen/DeviceAccelerator.*
+  - aten/src/ATen/core/GeneratorForPrivateuseone.*
+  - aten/src/ATen/detail/PrivateUse1HooksInterface.*
+  - docs/source/accelerator/**
+  - test/cpp_extensions/open_registration_extension/torch_openreg/**
+  approved_by:
+  - albanD
+  - fffrog
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: superuser
   patterns:
   - '*'

.github/pytorch-probot.yml

@@ -19,6 +19,7 @@ ciflow_push_tags:
 - ciflow/inductor-perf-test-nightly-rocm-mi300
 - ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
+- ciflow/inductor-perf-test-nightly-xpu
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
 - ciflow/linux-aarch64
@@ -26,6 +27,7 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
+- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic

.github/scripts/generate_binary_build_matrix.py

@@ -11,11 +11,17 @@ architectures:
     * Latest XPU
 """
 
+import json
 import os
+import re
+from pathlib import Path
 from typing import Optional
 
 
-# NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
+SCRIPT_DIR = Path(__file__).absolute().parent
+REPO_ROOT = SCRIPT_DIR.parent.parent
+
+
 CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
@@ -31,8 +37,7 @@ CUDA_ARCHES_CUDNN_VERSION = {
     "13.0": "9",
 }
 
-# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
+ROCM_ARCHES = ["7.0", "7.1"]
-ROCM_ARCHES = ["6.4", "7.0"]
 
 XPU_ARCHES = ["xpu"]
 
@@ -56,7 +61,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
@@ -73,7 +78,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
@@ -90,7 +95,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
@@ -107,7 +112,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
         "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | "
         "nvidia-nvtx==13.0.85; platform_system == 'Linux' | "
         "nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
         "nvidia-cufile==1.15.1.6; platform_system == 'Linux'"
@@ -137,9 +142,48 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
 }
 
 
-def get_nccl_wheel_version(arch_version: str) -> str:
+# Used by tools/nightly.py
-    import re
+PYTORCH_NIGHTLY_PIP_INDEX_URL = "https://download.pytorch.org/whl/nightly"
+NIGHTLY_SOURCE_MATRIX = {
+    "cpu": dict(
+        name="cpu",
+        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cpu",
+        supported_platforms=["Linux", "macOS", "Windows"],
+        accelerator="cpu",
+    )
+}
+CUDA_NIGHTLY_SOURCE_MATRIX = {
+    f"cuda-{major}.{minor}": dict(
+        name=f"cuda-{major}.{minor}",
+        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cu{major}{minor}",
+        supported_platforms=["Linux", "Windows"],
+        accelerator="cuda",
+    )
+    for major, minor in (map(int, version.split(".")) for version in CUDA_ARCHES)
+}
+ROCM_NIGHTLY_SOURCE_MATRIX = {
+    f"rocm-{major}.{minor}": dict(
+        name=f"rocm-{major}.{minor}",
+        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/rocm{major}.{minor}",
+        supported_platforms=["Linux"],
+        accelerator="rocm",
+    )
+    for major, minor in (map(int, version.split(".")) for version in ROCM_ARCHES)
+}
+XPU_NIGHTLY_SOURCE_MATRIX = {
+    "xpu": dict(
+        name="xpu",
+        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/xpu",
+        supported_platforms=["Linux"],
+        accelerator="xpu",
+    )
+}
+NIGHTLY_SOURCE_MATRIX.update(CUDA_NIGHTLY_SOURCE_MATRIX)
+NIGHTLY_SOURCE_MATRIX.update(ROCM_NIGHTLY_SOURCE_MATRIX)
+NIGHTLY_SOURCE_MATRIX.update(XPU_NIGHTLY_SOURCE_MATRIX)
 
+
+def get_nccl_wheel_version(arch_version: str) -> str:
     requirements = map(
         str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
     )
@@ -147,17 +191,14 @@ def get_nccl_wheel_version(arch_version: str) -> str:
 
 
 def read_nccl_pin(arch_version: str) -> str:
-    from pathlib import Path
+    nccl_pin_path = (
-
+        REPO_ROOT
-    nccl_pin_path = os.path.join(
+        / ".ci"
-        Path(__file__).absolute().parents[2],
+        / "docker"
-        ".ci",
+        / "ci_commit_pins"
-        "docker",
+        / f"nccl-cu{arch_version[:2]}.txt"
-        "ci_commit_pins",
-        f"nccl-cu{arch_version[:2]}.txt",
     )
-    with open(nccl_pin_path) as f:
+    return nccl_pin_path.read_text().strip()
-        return f.read().strip()
 
 
 def validate_nccl_dep_consistency(arch_version: str) -> None:
@@ -165,7 +206,8 @@ def validate_nccl_dep_consistency(arch_version: str) -> None:
     wheel_ver = get_nccl_wheel_version(arch_version)
     if not nccl_release_tag.startswith(f"v{wheel_ver}"):
         raise RuntimeError(
-            f"{arch_version} NCCL release tag version {nccl_release_tag} does not correspond to wheel version {wheel_ver}"
+            f"{arch_version} NCCL release tag version {nccl_release_tag} "
+            f"does not correspond to wheel version {wheel_ver}"
         )
 
 
@@ -412,7 +454,14 @@ def generate_wheels_matrix(
     return ret
 
 
-validate_nccl_dep_consistency("13.0")
+arch_version = ""
-validate_nccl_dep_consistency("12.9")
+for arch_version in CUDA_ARCHES:
-validate_nccl_dep_consistency("12.8")
+    validate_nccl_dep_consistency(arch_version)
-validate_nccl_dep_consistency("12.6")
+del arch_version
+
+
+if __name__ == "__main__":
+    # Used by tools/nightly.py
+    (SCRIPT_DIR / "nightly_source_matrix.json").write_text(
+        json.dumps(NIGHTLY_SOURCE_MATRIX, indent=4) + "\n"
+    )

.github/workflows/_xpu-test.yml

@@ -38,6 +38,10 @@ on:
         default: ""
         description: |
           List of tests to include (empty string implies default list)
+      dashboard-tag:
+        required: false
+        type: string
+        default: ""
       disable-monitor:
         description: |
           [Experimental] Disable utilization monitoring for tests.
@@ -58,6 +62,11 @@ on:
         required: false
         type: number
         default: 1
+    secrets:
+      HUGGING_FACE_HUB_TOKEN:
+        required: false
+        description: |
+          HF Auth token to avoid rate limits when downloading models or datasets from hub
 permissions:
   id-token: write
   contents: read
@@ -196,6 +205,8 @@ jobs:
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
+          DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
+          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           # Fetch aws credential from IMDs
@@ -246,6 +257,8 @@ jobs:
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
             -e ZE_AFFINITY_MASK \
+            -e HUGGING_FACE_HUB_TOKEN \
+            -e DASHBOARD_TAG \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \

.github/workflows/build-almalinux-images.yml

@@ -36,7 +36,7 @@ jobs:
     runs-on: linux.9xlarge.ephemeral
     strategy:
       matrix:
-        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "cpu"]
+        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm7.0", "rocm7.1", "cpu"]
     steps:
       - name: Build docker image
         uses: pytorch/pytorch/.github/actions/binary-docker-build@main

.github/workflows/build-libtorch-images.yml

@@ -52,8 +52,8 @@ jobs:
           { tag: "cuda12.9" },
           { tag: "cuda12.8" },
           { tag: "cuda12.6" },
-          { tag: "rocm6.4"  },
           { tag: "rocm7.0"  },
+          { tag: "rocm7.1"  },
           { tag: "cpu"      },
         ]
     steps:

.github/workflows/build-magma-rocm-linux.yml

@@ -34,7 +34,7 @@ jobs:
       id-token: write
     strategy:
       matrix:
-        rocm_version: ["70", "64"]
+        rocm_version: ["71", "70"]
     steps:
       - name: Checkout PyTorch
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

.github/workflows/build-manywheel-images.yml

@@ -54,8 +54,8 @@ jobs:
           { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm7.0",           runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "rocm7.1",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28_aarch64-builder",  tag: "cpu-aarch64",       runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "xpu",               runner: "linux.9xlarge.ephemeral" },

.github/workflows/build-triton-wheel.yml

@@ -55,7 +55,7 @@ jobs:
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
           - device: "rocm"
-            rocm_version: "7.0"
+            rocm_version: "7.1"
             runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
           - device: "cuda"
             rocm_version: ""
@@ -159,12 +159,7 @@ jobs:
             WITH_CLANG_LDD="--with-clang-ldd"
           fi
 
-          if [[ "${BUILD_DEVICE}" == xpu ]]; then
-            docker exec -t "${container_name}" bash -c "dnf install -y gcc-toolset-13-gcc-c++"
-            docker exec -t "${container_name}" bash -c "source /opt/rh/gcc-toolset-13/enable && ${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
-          else
           docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE $WITH_CLANG_LDD"
-          fi
 
           if [[ ("${{ matrix.device }}" == "cuda" || "${{ matrix.device }}" == "xpu") ]]; then
             docker exec -t "${container_name}"  bash -c "auditwheel repair --plat ${PLATFORM} //artifacts/*.whl"

.github/workflows/docker-builds.yml

@@ -57,6 +57,7 @@ jobs:
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.10-clang12,
           pytorch-linux-jammy-py3.13-clang12,
+          pytorch-linux-jammy-py3.14-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-jammy-rocm-n-py3-benchmarks,
@@ -66,6 +67,7 @@ jobs:
           pytorch-linux-jammy-py3.12-halide,
           pytorch-linux-jammy-xpu-n-1-py3,
           pytorch-linux-jammy-xpu-n-py3,
+          pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,

.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml

@@ -132,7 +132,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -178,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -270,7 +270,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +381,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -427,7 +427,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -473,7 +473,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -519,7 +519,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -676,7 +676,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -722,7 +722,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -768,7 +768,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -879,7 +879,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -925,7 +925,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -971,7 +971,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1017,7 +1017,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1128,7 +1128,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1174,7 +1174,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1220,7 +1220,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1266,7 +1266,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1377,7 +1377,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1423,7 +1423,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1469,7 +1469,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1515,7 +1515,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1626,7 +1626,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1672,7 +1672,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1718,7 +1718,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1764,7 +1764,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/generated-linux-binary-libtorch-nightly.yml

@@ -384,124 +384,6 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  libtorch-rocm6_4-shared-with-deps-release-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: libtorch-rocm6_4-shared-with-deps-release
-      build_environment: linux-binary-libtorch
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-rocm6_4-shared-with-deps-release-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - libtorch-rocm6_4-shared-with-deps-release-build
-      - get-label-type
-    runs-on: linux.rocm.gpu.mi250
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
-      GPU_ARCH_TYPE: rocm
-      SKIP_ALL_TESTS: 1
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Setup ROCm
-        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
-        name: Download Build Artifacts
-        with:
-          name: libtorch-rocm6_4-shared-with-deps-release
-          path: "${{ runner.temp }}/artifacts/"
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: ROCm set GPU_FLAG
-        run: |
-          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
-      - name: configure aws credentials
-        id: aws_creds
-        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
-        uses: aws-actions/configure-aws-credentials@v4
-        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: Calculate docker image
-        id: calculate-docker-image
-        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-        with:
-          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
-          docker-image-name: libtorch-cxx11-builder
-          custom-tag-prefix: rocm6.4
-          docker-build-dir: .ci/docker
-          working-directory: pytorch
-      - 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: Test Pytorch binary
-        uses: ./pytorch/.github/actions/test-pytorch-binary
-        env:
-          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Teardown ROCm
-        uses: ./.github/actions/teardown-rocm
-  libtorch-rocm6_4-shared-with-deps-release-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: libtorch-rocm6_4-shared-with-deps-release-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm6.4
-      GPU_ARCH_VERSION: "6.4"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-rocm6_4-shared-with-deps-release
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
-
   libtorch-rocm7_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -619,3 +501,121 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+
+  libtorch-rocm7_1-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm7.1
+      GPU_ARCH_VERSION: "7.1"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      timeout-minutes: 300
+      build_name: libtorch-rocm7_1-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-rocm7_1-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-rocm7_1-shared-with-deps-release-build
+      - get-label-type
+    runs-on: linux.rocm.gpu.mi250
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm7.1
+      GPU_ARCH_VERSION: "7.1"
+      GPU_ARCH_TYPE: rocm
+      SKIP_ALL_TESTS: 1
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Setup ROCm
+        uses: ./.github/actions/setup-rocm
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-rocm7_1-shared-with-deps-release
+          path: "${{ runner.temp }}/artifacts/"
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: ROCm set GPU_FLAG
+        run: |
+          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
+      - name: configure aws credentials
+        id: aws_creds
+        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
+        uses: aws-actions/configure-aws-credentials@v4
+        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: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
+          docker-image-name: libtorch-cxx11-builder
+          custom-tag-prefix: rocm7.1
+          docker-build-dir: .ci/docker
+          working-directory: pytorch
+      - 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: Test Pytorch binary
+        uses: ./pytorch/.github/actions/test-pytorch-binary
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Teardown ROCm
+        uses: ./.github/actions/teardown-rocm
+  libtorch-rocm7_1-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-rocm7_1-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm7.1
+      GPU_ARCH_VERSION: "7.1"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-rocm7_1-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml

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

@@ -0,0 +1,148 @@
+name: inductor-perf-nightly-xpu
+
+on:
+  push:
+    tags:
+      - ciflow/inductor-perf-test-nightly-xpu/*
+  schedule:
+    - cron: 30 17 * * *
+  workflow_dispatch:
+    inputs:
+      training:
+        description: Run training (on by default)?
+        required: false
+        type: boolean
+        default: true
+      inference:
+        description: Run inference (on by default)?
+        required: false
+        type: boolean
+        default: true
+      default:
+        description: Run inductor_default?
+        required: false
+        type: boolean
+        default: false
+      dynamic:
+        description: Run inductor_dynamic_shapes?
+        required: false
+        type: boolean
+        default: false
+      cppwrapper:
+        description: Run inductor_cpp_wrapper?
+        required: false
+        type: boolean
+        default: false
+      cudagraphs:
+        description: Run inductor_cudagraphs?
+        required: false
+        type: boolean
+        default: false
+      freezing_cudagraphs:
+        description: Run inductor_cudagraphs with freezing for inference?
+        required: false
+        type: boolean
+        default: false
+      aotinductor:
+        description: Run aot_inductor for inference?
+        required: false
+        type: boolean
+        default: false
+      maxautotune:
+        description: Run inductor_max_autotune?
+        required: false
+        type: boolean
+        default: false
+      benchmark_configs:
+        description: The list of configs used the benchmark
+        required: false
+        type: string
+        default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,cachebench
+
+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: read-all
+
+jobs:
+  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: ${{ github.ref_type }}
+      opt_out_experiments: lf
+
+  xpu-n-py3_10-inductor-benchmark-build:
+    name: xpu-n-py3.10-inductor-benchmark
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-xpu-n-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks
+      runner: linux.c7i.12xlarge
+      test-matrix: |
+        { include: [
+          { config: "inductor_huggingface_perf_xpu", shard: 1, num_shards: 5, runner: "linux.idc.xpu" },
+          { config: "inductor_huggingface_perf_xpu", shard: 2, num_shards: 5, runner: "linux.idc.xpu" },
+          { config: "inductor_huggingface_perf_xpu", shard: 3, num_shards: 5, runner: "linux.idc.xpu" },
+          { config: "inductor_huggingface_perf_xpu", shard: 4, num_shards: 5, runner: "linux.idc.xpu" },
+          { config: "inductor_huggingface_perf_xpu", shard: 5, num_shards: 5, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_timm_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
+          { config: "inductor_torchbench_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
+        ]}
+    secrets: inherit
+
+  xpu-n-py3_10-inductor-benchmark-test-nightly:
+    permissions:
+      id-token: write
+      contents: read
+    if: github.event_name != 'workflow_dispatch'
+    name: xpu-n-py3.10-inductor-benchmark
+    uses: ./.github/workflows/_xpu-test.yml
+    needs: xpu-n-py3_10-inductor-benchmark-build
+    with:
+      build-environment: linux-jammy-xpu-n-py3.10
+      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
+      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+      timeout-minutes: 720
+      # Disable monitor in perf tests for more investigation
+      disable-monitor: true
+      monitor-log-interval: 10
+      monitor-data-collect-interval: 2
+    secrets: inherit
+
+  xpu-n-py3_10-inductor-benchmark-test:
+    permissions:
+      id-token: write
+      contents: read
+    if: github.event_name == 'workflow_dispatch'
+    name: xpu-n-py3.10-inductor-test
+    uses: ./.github/workflows/_xpu-test.yml
+    needs: xpu-n-py3_10-inductor-benchmark-build
+    with:
+      build-environment: linux-jammy-xpu-n-py3.10
+      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
+      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
+      timeout-minutes: 720
+      disable-monitor: false
+      monitor-log-interval: 15
+      monitor-data-collect-interval: 4
+    secrets: inherit

.github/workflows/periodic-rocm-mi200.yml

@@ -0,0 +1,84 @@
+name: periodic-rocm-mi200
+
+on:
+  schedule:
+    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
+    # Also run less frequently on weekends.
+    - cron: 45 0,8,16 * * 1-5
+    - cron: 45 4 * * 0,6
+    - cron: 45 4,12,20 * * 1-5
+    - cron: 45 12 * * 0,6
+    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
+  push:
+    tags:
+      - ciflow/periodic/*
+      - ciflow/periodic-rocm-mi200/*
+    branches:
+      - release/*
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.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
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+        ]}
+    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/periodic.yml

@@ -204,37 +204,6 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
-  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
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-        ]}
-    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
-
   linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
     name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
     uses: ./.github/workflows/_linux-build.yml

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

@@ -6,6 +6,7 @@ on:
       - pull
       - trunk
       - periodic
+      - periodic-rocm-mi200
       - periodic-rocm-mi300
       - inductor
       - unstable

.github/workflows/xpu.yml

@@ -59,14 +59,18 @@ jobs:
       runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 2, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 2, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 3, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 3, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 4, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 4, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 5, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 5, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 6, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 6, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 7, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 7, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 8, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 8, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 9, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 10, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 11, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 12, num_shards: 12, runner: "linux.idc.xpu" },
         ]}
     secrets: inherit
 

.gitignore

@@ -143,6 +143,7 @@ scripts/release_notes/*.json
 sccache-stats*.json
 lint.json
 merge_record.json
+.github/scripts/nightly_source_matrix.json
 
 # These files get copied over on invoking setup.py
 torchgen/packaged/*

CMakeLists.txt

@@ -374,7 +374,7 @@ cmake_dependent_option(
   "Build the lazy Torchscript backend, not compatible with mobile builds" ON
   "NOT INTERN_BUILD_MOBILE" OFF)
 cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
-cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
+cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin folder"
                        OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                         "CPU_AARCH64" OFF)

README.md

@@ -1,4 +1,4 @@
-![PyTorch Logo](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
 
 --------------------------------------------------------------------------------
 
@@ -72,7 +72,7 @@ Elaborating Further:
 
 If you use NumPy, then you have used Tensors (a.k.a. ndarray).
 
-![Tensor illustration](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/tensor_illustration.png)
 
 PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the
 computation by a huge amount.
@@ -99,7 +99,7 @@ from several research papers on this topic, as well as current and past work suc
 While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date.
 You get the best of speed and flexibility for your crazy research.
 
-![Dynamic graph](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/dynamic_graph.gif)
 
 ### Python First
 

aten/src/ATen/CMakeLists.txt

@@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped).*")
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@@ -291,6 +291,7 @@ IF(USE_FBGEMM_GENAI)
 
     set(fbgemm_genai_cuh
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
       "${FBGEMM_GENAI_SRCS}/"
     )
 

aten/src/ATen/Context.cpp

@@ -825,6 +825,14 @@ void Context::setDisplayVmapFallbackWarnings(bool enabled) {
   display_vmap_fallback_warnings_ = enabled;
 }
 
+bool Context::warnOnAccumulateGradStreamMismatch() const {
+  return warn_on_accumulate_grad_stream_mismatch_;
+}
+
+void Context::setWarnOnAccumulateGradStreamMismatch(bool enabled) {
+  warn_on_accumulate_grad_stream_mismatch_ = enabled;
+}
+
 bool Context::isDefaultMobileCPUAllocatorSet() {
   return prev_allocator_ptr_ != nullptr;
 }

aten/src/ATen/Context.h

@@ -404,6 +404,9 @@ class TORCH_API Context {
   void setDisplayVmapFallbackWarnings(bool enabled);
   bool areVmapFallbackWarningsEnabled() const;
 
+  void setWarnOnAccumulateGradStreamMismatch(bool enabled);
+  bool warnOnAccumulateGradStreamMismatch() const;
+
   bool isDefaultMobileCPUAllocatorSet();
   void setDefaultMobileCPUAllocator();
   void unsetDefaultMobileCPUAllocator();
@@ -494,6 +497,7 @@ class TORCH_API Context {
   bool release_original_weights = false;
 #endif
   bool display_vmap_fallback_warnings_ = false;
+  bool warn_on_accumulate_grad_stream_mismatch_ = true;
   std::atomic<at::QEngine> quantized_engine = at::QEngine::NoQEngine;
   bool enable_sparse_tensor_invariant_checks = false;
   bool allow_fp16_reduction_cpu = false;

aten/src/ATen/core/CachingHostAllocator.h

@@ -677,8 +677,8 @@ struct CachingHostAllocatorImpl {
   // size. This allows us to quickly find a free block of the right size.
   // We use deque to store per size free list and guard the list with its own
   // mutex.
-  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
+  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>>
-      std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
+      free_list_{MAX_SIZE_INDEX};
 
   alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
   std::deque<std::pair<E, B*>> events_; // event queue paired with block

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

@@ -19,6 +19,13 @@ inline namespace CPU_CAPABILITY {
 #error "Big endian is not supported."
 #endif
 
+// GCC does not properly optimize bf16 operators
+#if defined(__ARM_FEATURE_BF16) && (__clang_major__ >= 19)
+#define BF16_ARITHMETIC_SUPPORTED() 1
+#else
+#define BF16_ARITHMETIC_SUPPORTED() 0
+#endif
+
 // Unlike the float16_t family of types, bfloat16_t is not available
 // when we're not targeting bfloat16 hardware support on some
 // platforms (but not Mac, so we have to be careful not to shadow the
@@ -352,18 +359,35 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
         other, &Vectorized<float>::name);                        \
   }
 
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
   Vectorized frac() const;
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
 
 #ifdef __ARM_FEATURE_BF16
+  // Flip sign bit
   Vectorized<c10::BFloat16> neg() const {
-    return -values;
+    return vreinterpretq_bf16_s16(vreinterpretq_s16_bf16(values) ^ (-32768));
   }
+  // Fast reciprocal is fine because we are truncating results
   Vectorized<c10::BFloat16> reciprocal() const {
-    return 1.0f / values;
+    auto x = vcvtq_low_f32_bf16(values);
+    auto y = vcvtq_high_f32_bf16(values);
+    x = vrecpeq_f32(x);
+    y = vrecpeq_f32(y);
+    return vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(x), y);
   }
+  // Clearing the sign bit
+  Vectorized<c10::BFloat16> abs() const {
+    return vreinterpretq_bf16_u16(vreinterpretq_u16_bf16(values) & 0x7FFF);
+  }
+#else
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
+#endif
+
+// These functions are optimized on clang-21+
+#if BF16_ARITHMETIC_SUPPORTED() && (__clang_major__ >= 21)
   Vectorized<c10::BFloat16> operator==(
       const Vectorized<c10::BFloat16>& other) const {
     return values == other.values;
@@ -394,8 +418,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
     return values >= other.values;
   }
 #else
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator==)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator!=)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<)
@@ -451,7 +473,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator+(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x + y;
@@ -464,7 +486,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator-(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x - y;
@@ -477,7 +499,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator*(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x * y;
@@ -490,7 +512,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator/(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x / y;
@@ -607,7 +629,7 @@ Vectorized<c10::BFloat16> inline fmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -627,7 +649,7 @@ Vectorized<c10::BFloat16> inline fnmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -643,7 +665,7 @@ Vectorized<c10::BFloat16> inline fmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -659,7 +681,7 @@ Vectorized<c10::BFloat16> inline fnmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
+#if BF16_ARITHMETIC_SUPPORTED()
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;

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

@@ -6,9 +6,9 @@ namespace at::vec {
 inline namespace CPU_CAPABILITY {
 #if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
 
-// Enable auto-vectorization for GCC-13+ and clang-17+
+// Enable auto-vectorization for clang-17+
 // GCC-12 has a bug: gcc.gnu.org/bugzilla/show_bug.cgi?id=117001
-#if __GNUC__ > 12 || (defined(__clang__) && (__clang_major__ >= 17))
+#if defined(__clang__) && (__clang_major__ >= 17)
 
 template <typename from_type, typename to_type>
 inline void convertImpl(
@@ -21,12 +21,46 @@ inline void convertImpl(
   }
 }
 
+template <typename to_type>
+inline void convertFromBool(
+    const bool* __restrict src,
+    to_type* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dst[i] = srcPtr[i] != 0 ? static_cast<to_type>(1) : static_cast<to_type>(0);
+  }
+}
+
+template <typename from_type>
+inline void convertToBool(
+    const from_type* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = src[i] != static_cast<from_type>(0) ? 1 : 0;
+  }
+}
+
 #define CONVERT_TEMPLATE(from_type, to_type)                           \
   template <>                                                          \
   inline void convert(const from_type* src, to_type* dst, int64_t n) { \
     return convertImpl<from_type, to_type>(src, dst, n);               \
   }
 
+#define CONVERT_FROM_BOOL_TEMPLATE(to_type)                       \
+  inline void convert(const bool* src, to_type* dst, int64_t n) { \
+    return convertFromBool<to_type>(src, dst, n);                 \
+  }
+
+#define CONVERT_TO_BOOL_TEMPLATE(from_type)                         \
+  inline void convert(const from_type* src, bool* dst, int64_t n) { \
+    return convertToBool<from_type>(src, dst, n);                   \
+  }
+
 CONVERT_TEMPLATE(uint8_t, uint8_t)
 CONVERT_TEMPLATE(uint8_t, int8_t)
 CONVERT_TEMPLATE(uint8_t, int16_t)
@@ -34,6 +68,7 @@ CONVERT_TEMPLATE(uint8_t, int32_t)
 CONVERT_TEMPLATE(uint8_t, int64_t)
 CONVERT_TEMPLATE(uint8_t, float)
 CONVERT_TEMPLATE(uint8_t, double)
+CONVERT_TO_BOOL_TEMPLATE(uint8_t)
 CONVERT_TEMPLATE(int8_t, uint8_t)
 CONVERT_TEMPLATE(int8_t, int8_t)
 CONVERT_TEMPLATE(int8_t, int16_t)
@@ -41,6 +76,7 @@ CONVERT_TEMPLATE(int8_t, int32_t)
 CONVERT_TEMPLATE(int8_t, int64_t)
 CONVERT_TEMPLATE(int8_t, float)
 CONVERT_TEMPLATE(int8_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int8_t)
 CONVERT_TEMPLATE(int16_t, uint8_t)
 CONVERT_TEMPLATE(int16_t, int8_t)
 CONVERT_TEMPLATE(int16_t, int16_t)
@@ -48,6 +84,7 @@ CONVERT_TEMPLATE(int16_t, int32_t)
 CONVERT_TEMPLATE(int16_t, int64_t)
 CONVERT_TEMPLATE(int16_t, float)
 CONVERT_TEMPLATE(int16_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int16_t)
 CONVERT_TEMPLATE(int32_t, uint8_t)
 CONVERT_TEMPLATE(int32_t, int8_t)
 CONVERT_TEMPLATE(int32_t, int16_t)
@@ -55,6 +92,7 @@ CONVERT_TEMPLATE(int32_t, int32_t)
 CONVERT_TEMPLATE(int32_t, int64_t)
 CONVERT_TEMPLATE(int32_t, float)
 CONVERT_TEMPLATE(int32_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int32_t)
 CONVERT_TEMPLATE(int64_t, uint8_t)
 CONVERT_TEMPLATE(int64_t, int8_t)
 CONVERT_TEMPLATE(int64_t, int16_t)
@@ -62,6 +100,7 @@ CONVERT_TEMPLATE(int64_t, int32_t)
 CONVERT_TEMPLATE(int64_t, int64_t)
 CONVERT_TEMPLATE(int64_t, float)
 CONVERT_TEMPLATE(int64_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int64_t)
 CONVERT_TEMPLATE(float, uint8_t)
 CONVERT_TEMPLATE(float, int8_t)
 CONVERT_TEMPLATE(float, int16_t)
@@ -69,6 +108,7 @@ CONVERT_TEMPLATE(float, int32_t)
 CONVERT_TEMPLATE(float, int64_t)
 CONVERT_TEMPLATE(float, float)
 CONVERT_TEMPLATE(float, double)
+CONVERT_TO_BOOL_TEMPLATE(float)
 CONVERT_TEMPLATE(double, uint8_t)
 CONVERT_TEMPLATE(double, int8_t)
 CONVERT_TEMPLATE(double, int16_t)
@@ -76,6 +116,14 @@ CONVERT_TEMPLATE(double, int32_t)
 CONVERT_TEMPLATE(double, int64_t)
 CONVERT_TEMPLATE(double, float)
 CONVERT_TEMPLATE(double, double)
+CONVERT_TO_BOOL_TEMPLATE(double)
+CONVERT_FROM_BOOL_TEMPLATE(uint8_t)
+CONVERT_FROM_BOOL_TEMPLATE(int8_t)
+CONVERT_FROM_BOOL_TEMPLATE(int16_t)
+CONVERT_FROM_BOOL_TEMPLATE(int32_t)
+CONVERT_FROM_BOOL_TEMPLATE(int64_t)
+CONVERT_FROM_BOOL_TEMPLATE(float)
+CONVERT_FROM_BOOL_TEMPLATE(double)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
 #define CONVERT_FROM_FP16_TEMPLATE(to_type)                            \
@@ -107,6 +155,41 @@ CONVERT_TO_FP16_TEMPLATE(int32_t)
 CONVERT_TO_FP16_TEMPLATE(int64_t)
 CONVERT_TO_FP16_TEMPLATE(float)
 CONVERT_TO_FP16_TEMPLATE(double)
+
+inline void convertBoolToFp16Impl(
+    const bool* __restrict src,
+    at::Half* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0 ? 1.0 : 0;
+  }
+}
+
+template <>
+inline void convert(const bool* src, at::Half* dst, int64_t n) {
+  return convertBoolToFp16Impl(src, dst, n);
+}
+
+inline void convertFp16ToBoolImpl(
+    const at::Half* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src);
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0.0 ? 1 : 0;
+  }
+}
+
+template <>
+inline void convert(const at::Half* src, bool* dst, int64_t n) {
+  return convertFp16ToBoolImpl(src, dst, n);
+}
+
 #endif
 #ifdef __ARM_FEATURE_BF16
 CONVERT_TEMPLATE(bfloat16_t, uint8_t)
@@ -124,6 +207,44 @@ CONVERT_TEMPLATE(int32_t, bfloat16_t)
 CONVERT_TEMPLATE(int64_t, bfloat16_t)
 CONVERT_TEMPLATE(float, bfloat16_t)
 CONVERT_TEMPLATE(double, bfloat16_t)
+
+inline void convertBoolToBfloat16Impl(
+    const bool* __restrict src,
+    c10::BFloat16* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  uint16_t* dstPtr = reinterpret_cast<uint16_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  constexpr uint16_t kBf16One = 0x3f80; // 1.0 in bfloat16
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0 ? kBf16One : 0;
+  }
+}
+
+template <>
+inline void convert(const bool* src, c10::BFloat16* dst, int64_t n) {
+  return convertBoolToBfloat16Impl(src, dst, n);
+}
+
+inline void convertBfloat16ToBoolImpl(
+    const c10::BFloat16* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    // Check if all non-sign bits are 0
+    bool isBf16Zero = (srcPtr[i] & 0x7fff) == 0;
+    dstPtr[i] = isBf16Zero ? 0 : 1;
+  }
+}
+
+template <>
+inline void convert(const c10::BFloat16* src, bool* dst, int64_t n) {
+  return convertBfloat16ToBoolImpl(src, dst, n);
+}
+
 #endif
 
 #endif

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

@@ -309,7 +309,7 @@ class Vectorized<float> {
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
   // Implementation copied from Arm Optimized Routine
   // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
-  Vectorized<float> exp_u20() const {
+  inline Vectorized<float> vexpq_f32_u20() const {
     // bail out to sleef if it's a special case:
     // i.e. there's an input s.t. |input| > 87.3....
     const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
@@ -348,6 +348,9 @@ class Vectorized<float> {
 
     return vfmaq_f32(scale, poly, scale);
   }
+  Vectorized<float> exp_u20() const {
+    return vexpq_f32_u20();
+  }
   Vectorized<float> fexp_u20() const {
     return exp_u20();
   }
@@ -634,7 +637,7 @@ inline Vectorized<float> Vectorized<float>::erf() const {
   // - exp(- x * x)
   auto pow_2 = (*this) * (*this);
   auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.exp();
+  auto tmp4 = neg_pow_2.vexpq_f32_u20();
   auto tmp5 = tmp4 ^ neg_zero_vec;
   // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
   auto tmp6 = t * tmp5;

aten/src/ATen/cuda/CUDAGreenContext.cpp

@@ -1,8 +1,20 @@
 #include <ATen/cuda/CUDAGreenContext.h>
 
+#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
+#include <c10/cuda/driver_api.h>
+#include <stdexcept>
+#include <vector>
+#define HAS_CUDA_GREEN_CONTEXT() 1
+#else
+#define HAS_CUDA_GREEN_CONTEXT() 0
+// Suppress unsued private field warnings as this class is not supposed to be called
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-private-field")
+#endif
+
 namespace at::cuda {
-  GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
+
-#if CUDA_HAS_GREEN_CONTEXT
+GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
+#if HAS_CUDA_GREEN_CONTEXT()
   int driver_version;
   C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
   TORCH_CHECK(
@@ -68,11 +80,11 @@ namespace at::cuda {
   std::unique_ptr<GreenContext> GreenContext::create(
       uint32_t num_sms,
       std::optional<uint32_t> device_id) {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     if (!device_id.has_value()) {
       device_id = at::cuda::current_device();
     }
-    return std::make_unique<GreenContext>(device_id.value(), num_sms);
+    return std::unique_ptr<GreenContext>(new GreenContext(device_id.value(), num_sms));
 #else
     TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
 #endif
@@ -80,7 +92,7 @@ namespace at::cuda {
 
   // Implement move operations
   GreenContext::GreenContext(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     device_id_ = std::exchange(other.device_id_, -1);
     green_ctx_ = std::exchange(other.green_ctx_, nullptr);
     context_ = std::exchange(other.context_, nullptr);
@@ -91,7 +103,7 @@ namespace at::cuda {
   }
 
   GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     if (this != &other) {
       // Clean up current resources
       if (green_ctx_) {
@@ -120,7 +132,7 @@ namespace at::cuda {
   }
 
   GreenContext::~GreenContext() noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     C10_CUDA_DRIVER_CHECK(
         c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
 #else
@@ -128,25 +140,9 @@ namespace at::cuda {
 #endif
   }
 
-  // Get the underlying CUDA context
-  CUcontext GreenContext::getContext() const {
-#if CUDA_HAS_GREEN_CONTEXT
-    return context_;
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx GreenContext::getGreenContext() const {
-    return green_ctx_;
-  }
-#endif
-
   // Make this context current
   void GreenContext::setContext() {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     auto current_stream = c10::cuda::getCurrentCUDAStream();
     parent_stream_ = current_stream.stream();
 
@@ -175,7 +171,7 @@ namespace at::cuda {
   }
 
   void GreenContext::popContext() {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
     // see above note about stream being hardcoded to the default stream
     at::cuda::CUDAEvent ev;
     ev.record(c10::cuda::getCurrentCUDAStream());

aten/src/ATen/cuda/CUDAGreenContext.h

@@ -1,53 +1,38 @@
 #pragma once
 #include <ATen/cuda/CUDAEvent.h>
-
-#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
 #include <cuda.h>
-#include <memory>
+
-#include <stdexcept>
+// Forward declare green context as opaque ptr
-#include <vector>
+typedef struct CUgreenCtx_st* CUgreenCtx;
-#define CUDA_HAS_GREEN_CONTEXT 1
-#else
-#define CUDA_HAS_GREEN_CONTEXT 0
-#endif
 
 namespace at::cuda {
 
 class TORCH_CUDA_CPP_API GreenContext {
  public:
-  GreenContext(uint32_t device_id, uint32_t num_sms);
+  // Green context creation
-
+  static std::unique_ptr<GreenContext> create(
-  static std::unique_ptr<GreenContext> create(uint32_t num_sms, std::optional<uint32_t> device_id);
+      uint32_t num_sms,
+      std::optional<uint32_t> device_id);
+  ~GreenContext() noexcept;
 
   // Delete copy constructor and assignment
   GreenContext(const GreenContext&) = delete;
   GreenContext& operator=(const GreenContext&) = delete;
 
-  // Implement move operations
-  GreenContext(GreenContext&& other) noexcept;
-  GreenContext& operator=(GreenContext&& other) noexcept;
-  ~GreenContext() noexcept;
-
-  // Get the underlying CUDA context
-  CUcontext getContext() const;
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx getGreenContext() const;
-#endif
-
   // Make this context current
   void setContext();
 
   void popContext();
 
  private:
-#if CUDA_HAS_GREEN_CONTEXT
+  GreenContext(uint32_t device_id, uint32_t num_sms);
+  // Implement move operations
+  GreenContext(GreenContext&& other) noexcept;
+  GreenContext& operator=(GreenContext&& other) noexcept;
+
   int32_t device_id_ = -1;
   CUgreenCtx green_ctx_ = nullptr;
   CUcontext context_ = nullptr;
   cudaStream_t parent_stream_ = nullptr;
-#endif
 };
 } // namespace at::cuda

aten/src/ATen/cuda/CUDASparse.h

@@ -7,17 +7,6 @@
 #endif
 
 
-#if defined(USE_ROCM)
-// hipSparse const API added in v2.4.0
-#if HIPSPARSE_VERSION >= 200400
-#define AT_USE_HIPSPARSE_GENERIC_API() 1
-#else
-#define AT_USE_HIPSPARSE_GENERIC_API() 1
-#endif
-#else // USE_ROCM
-#define AT_USE_HIPSPARSE_GENERIC_API() 0
-#endif // USE_ROCM
-
 // cuSparse Generic API spsv function was added in CUDA 11.3.0
 #if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
 #define AT_USE_CUSPARSE_GENERIC_SPSV() 1

aten/src/ATen/cuda/Sleep.cu

@@ -1,6 +1,7 @@
 #include <ATen/cuda/CUDAContextLight.h>
 #include <ATen/cuda/Sleep.h>
 
+#include <c10/cuda/CUDACachingAllocator.h>
 #include <c10/cuda/CUDAException.h>
 #include <c10/cuda/CUDAStream.h>
 
@@ -24,8 +25,22 @@ __global__ void spin_kernel(int64_t cycles) {
 #endif
   }
 }
+
+thread_local int *flag = nullptr;
+
+__global__ void busy_wait_for_flag_kernel(int *flag) {
+  atomicExch(flag, 1);
+  while (atomicAdd(flag, 0) == 1) {
+    // do nothing
+  }
 }
 
+__global__ void clear_flag_kernel(int *flag) {
+  atomicExch(flag, 0);
+}
+
+} // anonymous namespace
+
 void sleep(int64_t cycles) {
   dim3 grid(1);
   dim3 block(1);
@@ -33,6 +48,26 @@ void sleep(int64_t cycles) {
   C10_CUDA_KERNEL_LAUNCH_CHECK();
 }
 
+void busy_wait_for_flag() {
+  if (!flag) {
+    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
+  }
+  dim3 grid(1);
+  dim3 block(1);
+  busy_wait_for_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
+void clear_flag() {
+  if (!flag) {
+    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
+  }
+  dim3 grid(1);
+  dim3 block(1);
+  clear_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
 #ifdef USE_ROCM
 __global__ void flush_icache_kernel()
 {

aten/src/ATen/cuda/Sleep.h

@@ -7,6 +7,11 @@ namespace at::cuda {
 // enqueues a kernel that spins for the specified number of cycles
 TORCH_CUDA_CU_API void sleep(int64_t cycles);
 
+// enqueues a kernel that spins until a flag is cleared by a
+// corresponding call to clear_flag()
+TORCH_CUDA_CU_API void busy_wait_for_flag();
+TORCH_CUDA_CU_API void clear_flag();
+
 // flushes instruction cache for ROCm; no-op for CUDA
 TORCH_CUDA_CU_API void flush_icache();
 

aten/src/ATen/cuda/detail/BLASConstants.cu

@@ -2,8 +2,6 @@
 #include <ATen/Tensor.h>
 #include <ATen/cuda/Exceptions.h>
 
-#include <mutex>
-
 namespace at {
 namespace cuda {
 namespace detail {
@@ -12,39 +10,36 @@ __device__ __constant__ float cublas_one_device;
 __device__ __constant__ float cublas_zero_device;
 
 float *get_cublas_device_one() {
-  static c10::once_flag init_flag;
+  static float *ptr = nullptr;
-
+  static auto init_flag = [&]() {
-  c10::call_once(init_flag, []() {
     const float one = 1.f;
     AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_one_device, &one, sizeof(float)));
-  });
-
-  float *ptr;
     AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
+    return true;
+  }();
+
   return ptr;
 }
 
 float *get_cublas_device_zero() {
-  static c10::once_flag init_flag;
+  static float *ptr = nullptr;
-
+  static auto init_flag = [&]() {
-  c10::call_once(init_flag, []() {
     const float zero = 0.f;
     AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_zero_device, &zero, sizeof(float)));
-  });
-
-  float *ptr;
     AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
+    return true;
+  }();
+
   return ptr;
 }
 
 float *get_user_alpha_ptr() {
   static float *alpha_ptr;
 
-  static c10::once_flag init_flag;
+  static bool init_flag [[maybe_unused]] = []() {
-
-  c10::call_once(init_flag, []() {
     AT_CUDA_CHECK(cudaMalloc(&alpha_ptr, sizeof(float)));
-  });
+    return true;
+  }();
 
   return alpha_ptr;
 }

aten/src/ATen/cuda/tunable/Tunable.cpp

@@ -580,7 +580,7 @@ std::ofstream& TuningContext::GetUntunedFile(){
       filename.append(device);
     }
 
-    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::trunc);
+    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::app);
   }
   return untuned_file_;
 }

aten/src/ATen/detail/MTIAHooksInterface.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <c10/core/CachingDeviceAllocator.h>
 #include <c10/core/Device.h>
 #include <c10/util/Exception.h>
 
@@ -151,6 +152,36 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
   }
 
   virtual bool isAvailable() const override;
+
+  /* MTIAGraph related APIs */
+  virtual int64_t mtiagraphCreate(bool keep_graph = false) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+    return -1;
+  }
+
+  virtual void mtiagraphCaptureBegin(int64_t handle, MempoolId_t pool) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
+
+  virtual void mtiagraphCaptureEnd(int64_t handle) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
+
+  virtual void mtiagraphInstantiate(int64_t handle) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
+
+  virtual void mtiagraphReplay(int64_t handle) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
+
+  virtual void mtiagraphReset(int64_t handle) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
+
+  virtual MempoolId_t mtiagraphPool(int64_t handle) const {
+    FAIL_MTIAHOOKS_FUNC(__func__);
+  }
 };
 
 struct TORCH_API MTIAHooksArgs {};

aten/src/ATen/functorch/BatchRulesViews.cpp

@@ -534,20 +534,20 @@ Tensor trace_decomp(const Tensor& tensor) {
 std::tuple<Tensor, std::optional<int64_t>> tril_batch_rule(
     const Tensor& self,
     std::optional<int64_t> self_bdim,
-    int64_t diagonal = 0) {
+    c10::SymInt diagonal = 0) {
   TORCH_CHECK(self.dim() >= 2, "tril: The input tensor must have at least 2 dimensions.");
   auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::tril(self_, diagonal);
+  auto result = at::tril_symint(self_, std::move(diagonal));
   return std::make_tuple(std::move(result), 0);
 }
 
 std::tuple<Tensor, std::optional<int64_t>> triu_batch_rule(
     const Tensor& self,
     std::optional<int64_t> self_bdim,
-    int64_t diagonal = 0) {
+    c10::SymInt diagonal = 0) {
   TORCH_CHECK(self.dim() >= 2, "triu: The input tensor must have at least 2 dimensions.");
   auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::triu(self_, diagonal);
+  auto result = at::triu_symint(self_, std::move(diagonal));
   return std::make_tuple(std::move(result), 0);
 }
 

aten/src/ATen/mps/MPSDevice.mm

@@ -1,7 +1,5 @@
 //  Copyright © 2022 Apple Inc.
 
-#include <c10/util/CallOnce.h>
-
 #include <ATen/mps/IndexKernels.h>
 #include <ATen/mps/MPSAllocatorInterface.h>
 #include <ATen/mps/MPSDevice.h>
@@ -10,9 +8,6 @@
 
 namespace at::mps {
 
-static std::unique_ptr<MPSDevice> mps_device;
-static c10::once_flag mpsdev_init;
-
 static inline MTLLanguageVersion getMetalLanguageVersion(const id<MTLDevice>& device) {
   // MPS Advanced Indexing needs at least Metal 2.0 (support for Argument Buffers and function constants)
   // host_name attribute needs at least Metal 2.2 and ulong needs Metal 2.3 (supported on MacOS 11+
@@ -21,8 +16,8 @@ static inline MTLLanguageVersion getMetalLanguageVersion(const id<MTLDevice>& de
 }
 
 MPSDevice* MPSDevice::getInstance() {
-  c10::call_once(mpsdev_init, [] { mps_device = std::unique_ptr<MPSDevice>(new MPSDevice()); });
+  static MPSDevice mps_device;
-  return mps_device.get();
+  return &mps_device;
 }
 
 MPSDevice::~MPSDevice() {

aten/src/ATen/native/AveragePool2d.cpp

@@ -25,18 +25,19 @@ TORCH_PRECOMPUTE_META_FUNC(avg_pool2d)
   // #20866, #22032: Guarantee this for the official C++ API?
   TORCH_CHECK(kernel_size.size() == 1 || kernel_size.size() == 2,
     "avg_pool2d: kernel_size must either be a single int, or a tuple of two ints");
-  const int64_t kH = kernel_size[0];
+  const int kH = safe_downcast<int, int64_t>(kernel_size[0]);
-  const int64_t kW = kernel_size.size() == 1 ? kH : kernel_size[1];
+  const int kW = kernel_size.size() == 1 ? kH : safe_downcast<int, int64_t>(kernel_size[1]);
 
   TORCH_CHECK(stride.empty() || stride.size() == 1 || stride.size() == 2,
     "avg_pool2d: stride must either be omitted, a single int, or a tuple of two ints");
-  const int64_t dH = stride.empty() ? kH : stride[0];
+  const int dH = stride.empty() ? kH : safe_downcast<int, int64_t>(stride[0]);
-  const int64_t dW = stride.empty() ? kW : stride.size() == 1 ? dH : stride[1];
+  const int dW = stride.empty() ? kW :
+                 stride.size() == 1 ? dH : safe_downcast<int, int64_t>(stride[1]);
 
   TORCH_CHECK(padding.size() == 1 || padding.size() == 2,
     "avg_pool2d: padding must either be a single int, or a tuple of two ints");
-  const int64_t padH = padding[0];
+  const int padH = safe_downcast<int, int64_t>(padding[0]);
-  const int64_t padW = padding.size() == 1 ? padH : padding[1];
+  const int padW = padding.size() == 1 ? padH : safe_downcast<int, int64_t>(padding[1]);
 
   TORCH_CHECK(!divisor_override.has_value() || divisor_override.value() != 0,
     "divisor must be not zero");

aten/src/ATen/native/Convolution.cpp

@@ -410,8 +410,8 @@ struct ConvParams {
       return false;
     }
     static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
-    // broken on cuDNN 9.8
+    // broken on cuDNN 9.8 - 9.14
-    if (cudnn_version >= 90800) {
+    if (cudnn_version >= 90800 && cudnn_version < 91500) {
       if (cudnn_conv_suggest_memory_format(input, weight) == at::MemoryFormat::Contiguous &&
           (input.scalar_type() == at::kBFloat16 || input.scalar_type() == at::kHalf) &&
           weight.dim() == 5) {
@@ -689,6 +689,10 @@ static void check_shape_forward(const at::Tensor& input,
              ", but got bias of size ", at::symint::sizes<T>(bias), " instead");
 
     for (const auto i : c10::irange(2, k)) {
+      // T could be int64_t or SymInt, Specialized numeric_limts<SymInt> in c10/core/SymInt.h
+      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
+                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
+                  (std::numeric_limits<T>::max() / 2));
       input_shape.push_back(at::symint::size<T>(input, i) + 2 * padding[i-2]);
       // log new kernel size considering dilation
       kernel_shape.push_back(dilation[i-2] * (weight_sizes[i]-1) + 1);
@@ -715,6 +719,11 @@ static void check_shape_forward(const at::Tensor& input,
                "Kernel size: (", kernel_ss.str(), "). Kernel size can't be greater than actual input size");
     }
   } else { // transposed
+    for (const auto i : c10::irange(2, k)) {
+      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
+                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
+                  (std::numeric_limits<T>::max() / 2));
+    }
     TORCH_CHECK(at::symint::size<T>(input, 1) == weight_sizes[0],
              "Given transposed=", transposed, ", weight of size ", weight_sizes,
              ", expected input", at::symint::sizes<T>(input), " to have ", weight_sizes[0],

aten/src/ATen/native/ConvolutionTBC.cpp

@@ -52,8 +52,7 @@ Tensor conv_tbc(const Tensor& self, const Tensor& weight, const Tensor& bias, in
   for (const auto k : c10::irange(kw)) {
     int iShift = std::max(0, static_cast<int>(k - real_pad));
     int oShift = std::max(0, static_cast<int>(real_pad - k));
-    // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+    long t = std::min(ilen + real_pad - k, olen) - oShift;
-    int t = std::min(ilen + real_pad - k, olen) - oShift;
     // Note: gemm assumes column-major matrices
     // input    is l*m (row-major)
     // weight   is m*r (row-major)

aten/src/ATen/native/IndexingUtils.cpp

@@ -16,8 +16,7 @@ bool canUse32BitIndexMath(const TensorBase& t, int64_t max_elem) {
   auto linearId = elements - 1;
 
   // NOTE: Assumes all strides are positive, which is true for now
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  for (auto i = t.dim() - 1; i >= 0; --i) {
-  for (int i = t.dim() - 1; i >= 0; --i) {
     auto curDimIndex = linearId % t.sym_size(i);
     auto curDimOffset = curDimIndex * t.sym_stride(i);
     offset += curDimOffset;

aten/src/ATen/native/QuantizedLinear.cpp

@@ -68,7 +68,6 @@ Tensor fbgemm_linear_int8_weight_fp32_activation(
   const float* input_ptr = input_contig.const_data_ptr<float>();
 
   TORCH_CHECK(input.dim() >= 2);
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
   const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
   const int64_t K = input.size(input.dim() - 1);
   TORCH_CHECK(weight.dim() == 2);

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

@@ -160,10 +160,9 @@ struct Dist {
     // value of k.
     parallel_for(0, combs, internal::GRAIN_SIZE / (16 * m), [p, self_start, self_end, n, m, res_start](int64_t k, int64_t end) {
       const Vec pvec(p);
-      double n2 = n - .5;
+      double n2 = static_cast<double>(n) - .5;
       // The -1 accounts for floating point truncation issues
-      // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2.0 * static_cast<double>(k) - 1.0)));
-      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2 * k - 1)));
       int64_t j = k - n * i + i * (i + 1) / 2 + i + 1;
 
       const scalar_t * self_i = self_start + i * m;

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

@@ -139,7 +139,7 @@ void smooth_l1_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, dou
       }
     );
   } else {
-    AT_DISPATCH_ALL_TYPES(dtype, "smooth_l1_backward_cpu_out", [&] {
+    AT_DISPATCH_ALL_TYPES_AND(kHalf, dtype, "smooth_l1_backward_cpu_out", [&] {
     auto norm_val = norm.to<scalar_t>();
     scalar_t beta_val(beta);
     auto norm_val_vec = Vectorized<scalar_t>(norm_val);

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

@@ -170,10 +170,14 @@ static bool isInputCompliesAddmmCudaLt(Tensor& result, const Tensor& self, const
   #if defined(CUDA_VERSION) || defined(USE_ROCM)
   const auto scalar_type = mat1.scalar_type();
   return (beta.toComplexDouble() == 1.0
-    // self.dim() == 1 && result.dim() == 2 && self.sizes()[0] == mat2_sizes[1]
-    // is to use lt interface only when self is bias.
-    && self.dim() == 1 && self.sizes()[0] == mat2_sizes[1] && self.is_contiguous()
     && result.dim() == 2 && result.is_contiguous()
+    // Conditions for bias to be fusable
+    && (
+      self.is_contiguous() &&
+      // NOTE: fine to have 1-len dims to the left from the right-most one
+      (self.dim() == 1 || self.squeeze().dim() == 1) &&
+      self.sizes().back() == mat2_sizes[1]
+    )
     && ( // some dtype restrictions
       #ifndef USE_ROCM
       scalar_type == at::ScalarType::Double ||

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

@@ -208,6 +208,62 @@ _f8_f8_bf16_rowwise_grouped_mm(
 #endif
 }
 
+Tensor&
+_f4_f4_bf16_grouped_mm_fbgemm(
+      const Tensor& mat_a,
+      const Tensor& mat_b,
+      const Tensor& scale_a,
+      const std::optional<Tensor>& global_scale_a,
+      const Tensor& scale_b,
+      const std::optional<Tensor>& global_scale_b,
+      const std::optional<Tensor>& offs,
+      const std::optional<Tensor>& bias,
+      Tensor& out) {
+#if !defined(USE_ROCM) && defined(USE_FBGEMM_GENAI)
+  // Typing checks
+  TORCH_CHECK_VALUE(mat_a.scalar_type() == at::kFloat4_e2m1fn_x2,
+      "mat_a must be Float4_e2n1fn_2, got: ", mat_a.scalar_type());
+  TORCH_CHECK_VALUE(mat_b.scalar_type() == at::kFloat4_e2m1fn_x2,
+      "mat_b must be Float4_e2n1fn_2, got: ", mat_b.scalar_type());
+
+  std::optional<Tensor> combined_global_scale = std::nullopt;
+  if (global_scale_a.has_value() || global_scale_b.has_value()) {
+      // NVFP4
+      TORCH_CHECK_VALUE(global_scale_a.has_value() && global_scale_b.has_value(),
+          "For NVFP4 grouped gemm both of global_scale_{a,b} must have values")
+      TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e4m3fn,
+          "scale_a must be Float8_e4m3fn, got: ", scale_a.scalar_type());
+      TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e4m3fn,
+          "scale_b must be Float8_e4m3fn, got: ", scale_b.scalar_type());
+      TORCH_CHECK_VALUE(global_scale_a.value().scalar_type() == at::kFloat,
+          "global_scale_a must be Float, got: ", global_scale_a.value().scalar_type());
+      TORCH_CHECK_VALUE(global_scale_b.value().scalar_type() == at::kFloat,
+          "global_scale_b must be Float, got: ", global_scale_b.value().scalar_type());
+      combined_global_scale = global_scale_a.value().mul(global_scale_b.value());
+  } else {
+      // MXFP4
+      TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e8m0fnu,
+          "scale_a must be Float8_e8m0fnu, got: ", scale_a.scalar_type());
+      TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e8m0fnu,
+          "scale_b must be Float8_e8m0fnu, got: ", scale_b.scalar_type());
+  }
+
+  auto o = fbgemm_gpu::f4f4bf16_grouped_mm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs.value(),
+      out,
+      combined_global_scale
+  );
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(false, "nvfp4 grouped gemm is not supported without USE_FBGEMM_GENAI, and only for CUDA")
+#endif
+
+  return out;
+}
+
 void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
   // Checks scales for 2d or 3d target tensors (`mat`).
   if (mat.dim() == 2) {
@@ -245,7 +301,15 @@ void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int
   }
 }
 
-void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
+void _check_scales_blocked(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
+  // if {mx,nv}fp4, will need to modify K later
+  bool is_fp4 = (mat.scalar_type() == kFloat4_e2m1fn_x2);
+  int blocksize = 32;
+  // check for nvfp4 vs. mxfp4 to fix blocksize
+  if (is_fp4 && scale.scalar_type() == kFloat8_e4m3fn) {
+    blocksize = 16;
+  }
+
   // Checks scales for 2d or 3d target tensors (`mat`).
   if (mat.dim() == 2) {
     // For MXFP8, 2d tensors have variable size groups represented as subtensors,
@@ -253,17 +317,19 @@ void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim,
     // so we can't check the scale sizes without doing a d2h sync to get the group sizes here.
     TORCH_CHECK(
       scale.dim() == mat.dim(),
-      "for mxfp8, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(), " and scale.dim() = ", scale.dim(), " for arg ", arg_idx);
+      "for block-scaled, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(),
+      " and scale.dim() = ", scale.dim(), " for arg ", arg_idx
+    );
 
-    // LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/32, 4))
+    // LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/blocksize, 4))
-    // RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/32, 4))
+    // RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/blocksize, 4))
     //   * weight is transposed prior to the call, scale stays non-transposed.
     bool LHS = arg_idx == 0;
     int scale_dim_to_check = 0;
     int mat_dim_to_check = LHS ? 0 : 1;
     TORCH_CHECK(
         scale.size(scale_dim_to_check) >= mat.size(mat_dim_to_check),
-        "for mxfp8, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
+        "for block-scaled, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
         "must have scale.shape[", scale_dim_to_check, "] >= ", mat.size(mat_dim_to_check), " but got scale.shape=(", scale.size(0), ", ", scale.size(1), ")");
   } else {
     // For MXFP8, 3d tensors have static group sizes (stack of 2d tensors),
@@ -273,32 +339,40 @@ void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim,
     };
 
     // TODO: this is for 3d tensor in 2d-3d case specifically.
-    // We'll need to support 3d-3d and 3d-2d cases once mxfp8 grouped gemm supports them.
+    // We'll need to support 3d-3d and 3d-2d cases once mxfp8/nvfp4 grouped gemm supports them.
     int64_t G = mat.size(0);
     int64_t K = mat.size(1);
+    if (is_fp4) {
+      // FP4 packs 2 values into a single 8b word - the "real" K is 2x the
+      // reported K. Reverse that adjustment.
+      const int fp4_elems_per_byte = 2;
+      K *= fp4_elems_per_byte;
+    }
     int64_t N = mat.size(2);
-    int64_t blocked_scale_K = round_up(K/32, 4);
+    int64_t blocked_scale_K = round_up(K/blocksize, 4);
     int64_t blocked_scale_N = round_up(N, 128);
 
     // fbgemm expects stack of flattened blocked scales for 3d tensor, shape (G, blocked_scale_K * blocked_scale_N).
     TORCH_CHECK(
       scale.dim() == mat.dim() - 1,
-      "for mxfp8 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N), but scale is ", scale.dim(), "D for arg ", arg_idx
+      "for block-scaled 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N),",
+      "but scale is ", scale.dim(), "D for arg ", arg_idx
     );
     TORCH_CHECK(
       scale.size(0) == G && scale.size(1) == blocked_scale_K * blocked_scale_N,
-      "for mxfp8, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ") for arg ", arg_idx
+      "for block-scaled grouped GEMM, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ")",
+      " for arg ", arg_idx, ", got: ", scale.size(0), ", ", scale.size(1)
     );
   }
 }
 
 void check_scale(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
   bool using_fp8_rowwise = scale.scalar_type() == kFloat;
-  bool using_mxfp8 = scale.scalar_type() == at::kFloat8_e8m0fnu;
+  bool using_mx = scale.scalar_type() == at::kFloat8_e8m0fnu;
   if (using_fp8_rowwise) {
     _check_scales_fp8_rowwise(mat, scale, dim, arg_idx, scale_multiplier);
-  } else if (using_mxfp8) {
+  } else if (using_mx) {
-    _check_scales_mxfp8(mat, scale, dim, arg_idx);
+    _check_scales_blocked(mat, scale, dim, arg_idx);
   } else {
     TORCH_CHECK(false, "scale must be float32 or float8_e8m0fnu, but got ", scale.dtype());
   }
@@ -411,9 +485,11 @@ namespace {
 
 using acceptance_fn = std::function<bool(c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&, c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&)>;
 
-std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2> scale_grouped_kernel_dispatch = {{
+std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 4> scale_grouped_kernel_dispatch = {{
   { "rowwise_rowwise", scaled_blas::check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
-  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8}}};
+  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8},
+  { "mxfp4_mxfp4", scaled_blas::check_mxfp4_recipe, ScaledGemmImplementation::MXFP4_MXFP4},
+  { "nvfp4_nvfp4", scaled_blas::check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4}}};
 
 } // anonymous namespace
 
@@ -525,8 +601,9 @@ _scaled_grouped_mm_cuda_v2(
           out);
     }
     case ScaledGemmImplementation::MXFP8_MXFP8: {
-      _check_scales_mxfp8(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      // scale shape checks
-      _check_scales_mxfp8(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
       return _mx8_mx8_bf16_grouped_mm_fbgemm(
           mat_a,
           mat_b,
@@ -537,6 +614,36 @@ _scaled_grouped_mm_cuda_v2(
           offs.value(),
           out);
     }
+    case ScaledGemmImplementation::MXFP4_MXFP4: {
+      // scale shape checks
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      return _f4_f4_bf16_grouped_mm_fbgemm(
+          mat_a,
+          mat_b,
+          scale_a[0], /* block-scale A */
+          std::nullopt, /* global-scale A */
+          scale_b[0], /* block-scale B */
+          std::nullopt, /* global-scale B */
+          offs.value(),
+          std::nullopt, /* bias */
+          out);
+    }
+    case ScaledGemmImplementation::NVFP4_NVFP4: {
+      // scale shape checks
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      return _f4_f4_bf16_grouped_mm_fbgemm(
+          mat_a,
+          mat_b,
+          scale_a[0], /* block-scale A */
+          scale_a[1], /* global-scale A */
+          scale_b[0], /* block-scale B */
+          scale_b[1], /* global-scale B */
+          offs.value(),
+          std::nullopt, /* bias */
+          out);
+    }
     default:
       TORCH_CHECK_NOT_IMPLEMENTED(false,
           "_scaled_grouped_mm_cuda_v2 is in an inconsistent state - should never reach here");

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

@@ -13,7 +13,7 @@ __global__ void vectorized_gather_kernel(char * out, char * inp, index_t * idx,
     if (allow_neg_indices) {
         ind = (ind < 0) ? ind + ind_dim_size : ind;
     }
-    CUDA_KERNEL_ASSERT(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds");
+    CUDA_KERNEL_ASSERT_VERBOSE(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds", "Expected 0 <= index < ind_dim_size(%ld), but got index = %ld", ind_dim_size, ind);
     int32_t off = (blockDim.x * blockIdx.y + threadIdx.x) * Alignment; // off is guaranteed to be within int32 limits
     if (off >= slice_size) return;
     auto vec = at::native::memory::ld_vec<Alignment>(inp + ind * inp_stride + off);

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

@@ -794,6 +794,24 @@ void _check_deepseek_scale_stride(const Tensor& scale, const Tensor& t, const Sc
   }
 }
 
+void
+_check_deepseek_support() {
+#ifndef USE_ROCM
+  auto dprops = at::cuda::getCurrentDeviceProperties();
+  if (dprops->major != 9) {
+    // Only on Hopper GPUs
+    TORCH_CHECK_NOT_IMPLEMENTED(
+      dprops->major == 9,
+      "DeepSeek style (1x128, 128x128) scaling only supported in CUDA for SM90")
+  }
+  // Only in cublasLt >= 12.9
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    CUBLAS_VERSION < 120900 || cublasLtGetVersion() < 120900,
+    "DeepSeek style (1x128, 128x128) scaling requires cublasLt >= 12.9"
+  );
+#endif
+}
+
 Tensor&
 _scaled_block1x128_block1x128(
           const Tensor& mat_a, const Tensor& mat_b,
@@ -802,8 +820,12 @@ _scaled_block1x128_block1x128(
           const c10::ScalarType out_dtype,
           const bool use_fast_accum,
           Tensor& out) {
+#ifndef USE_ROCM
   // Restrictions:
   // A, B are FP8, scales are fp32, shape K//128
+  // CUDA: Only Hopper GPUs
+  _check_deepseek_support();
+
   TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
       mat_a.scalar_type(), mat_b.scalar_type());
   TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
@@ -821,6 +843,12 @@ _scaled_block1x128_block1x128(
   _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
 
   return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }
 
 Tensor&
@@ -831,10 +859,12 @@ _scaled_block128x128_block1x128(
           const c10::ScalarType out_dtype,
           const bool use_fast_accum,
           Tensor& out) {
+#ifndef USE_ROCM
   // Restrictions:
   // A, B are FP8, scales are fp32, shape K//128
-  std::cout << "mat_b: " << mat_b.dim() << ", " << mat_b.sizes() << ", " << mat_b.strides() << std::endl;
+  // CUDA: Only Hopper GPUs
-  std::cout << "scale_b: " << scale_b.dim() << ", " << scale_b.sizes() << ", " << scale_b.strides() << std::endl;
+  _check_deepseek_support();
+
   TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
       mat_a.scalar_type(), mat_b.scalar_type());
   TORCH_CHECK_VALUE(scale_a.sizes()[0] == ceil_div<int64_t>(mat_a.sizes()[0], 128) && scale_a.sizes()[1] == ceil_div<int64_t>(mat_a.sizes()[1], 128) && scale_a.scalar_type() == kFloat,
@@ -852,6 +882,12 @@ _scaled_block128x128_block1x128(
   _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
 
   return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }
 
 Tensor&
@@ -862,8 +898,12 @@ _scaled_block1x128_block128x128(
           const c10::ScalarType out_dtype,
           const bool use_fast_accum,
           Tensor& out) {
+#ifndef USE_ROCM
   // Restrictions:
   // A, B are FP8, scales are fp32, A: shape K//128, B: K//128, N//128
+  // CUDA: Only Hopper GPUs
+  _check_deepseek_support();
+
   TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
       mat_a.scalar_type(), mat_b.scalar_type());
   TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
@@ -881,6 +921,12 @@ _scaled_block1x128_block128x128(
   _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
 
   return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }
 
 Tensor&

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

@@ -160,8 +160,8 @@ struct _cuda_scatter_gather_internal_kernel {
       auto offsets = offset_calc.get(i);
 
       int64_t idx_dim = *(index_t*)(index_ptr + offsets[2]);
-      CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
+      CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
-        && "scatter gather kernel index out of bounds");
+        && "scatter gather kernel index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);
 
       f(
         (scalar_t*)(self_ptr + offsets[0]),
@@ -406,9 +406,8 @@ struct _cuda_scatter_fill_internal_kernel {
       auto offsets = offset_calc.get(i);
 
       int64_t idx_dim = *(index_t*)(index_ptr + offsets[1]);
-      CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
+      CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
-        && "index out of bounds"
+        && "index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);
-      );
 
       f(
         (scalar_t*)(self_ptr + offsets[0]),

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

@@ -12,14 +12,15 @@
 
 namespace at::native {
 
-#if AT_USE_JITERATOR()
+#if 0 && AT_USE_JITERATOR()
 constexpr char tan_name[] = "tan_impl";
 #endif
 
 void tan_kernel_cuda(TensorIteratorBase& iter) {
   auto common_dtype = iter.common_dtype();
   if (at::isComplexType(common_dtype)) {
-#if AT_USE_JITERATOR()
+    // Disabled due to accuracy issues
+#if 0 && AT_USE_JITERATOR()
     static const auto tan_string = jiterator_stringify(
         template <typename T> T tan_impl(T a) { return std::tan(a); });
     AT_DISPATCH_COMPLEX_TYPES_AND(

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

@@ -12,14 +12,15 @@
 
 namespace at::native {
 
-#if AT_USE_JITERATOR()
+#if 0 && AT_USE_JITERATOR()
 constexpr char tanh_name[] = "tanh_impl";
 #endif
 
 void tanh_kernel_cuda(TensorIteratorBase& iter) {
   auto common_dtype = iter.common_dtype();
   if (at::isComplexType(common_dtype)) {
-#if AT_USE_JITERATOR()
+    // Disabled due to accuracy issues
+#if 0 && AT_USE_JITERATOR()
     static const auto tanh_string = jiterator_stringify(
         template <typename T> T tanh_impl(T a) { return std::tanh(a); });
     AT_DISPATCH_COMPLEX_TYPES_AND(

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

@@ -141,7 +141,8 @@ WelfordDataLN cuWelfordOnlineSum(
   if constexpr (!rms_norm){
     U delta = val - curr_sum.mean;
     U new_count = curr_sum.count + 1.f;
-#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
+#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
     U new_mean = curr_sum.mean + delta * __builtin_amdgcn_rcpf(new_count);
 #else
     U new_mean = curr_sum.mean + delta * (1.f/new_count); //proper division is slow, this is less accurate but noticeably faster
@@ -163,7 +164,8 @@ WelfordDataLN cuWelfordCombine(
     U count = dataA.count + dataB.count;
     U mean, sigma2;
     if (count > decltype(dataB.count){0}) {
-#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
+#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
       auto coef = __builtin_amdgcn_rcpf(count);
 #else
       auto coef = 1.f/count; //NB we don't use --use_fast_math, but this is emulation, 1./count goes to intrinsic, `* coef` is multiplication, instead of slow fp division

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

@@ -86,6 +86,28 @@ struct zeta_functor {
   }
 };
 
+struct logaddexp_functor {
+  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
+  inline T operator()(const T a, const T b) {
+    return c10::metal::logaddexp(a, b);
+  }
+  template <typename T, enable_if_t<is_integral_v<T>, bool> = true>
+  inline float operator()(const T a, const T b) {
+    return c10::metal::logaddexp(float(a), float(b));
+  }
+};
+
+struct logaddexp2_functor {
+  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
+  inline T operator()(const T a, const T b) {
+    return c10::metal::logaddexp2(a, b);
+  }
+  template <typename T, enable_if_t<is_integral_v<T>, bool> = true>
+  inline float operator()(const T a, const T b) {
+    return c10::metal::logaddexp2(float(a), float(b));
+  }
+};
+
 struct xlog1py_functor {
   template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
   inline T operator()(const T a, const T b) {
@@ -377,6 +399,10 @@ REGISTER_FLOAT_BINARY_OP(fmin);
 REGISTER_FLOAT_BINARY_OP(nextafter);
 REGISTER_FLOAT_BINARY_OP(zeta);
 REGISTER_INT2FLOAT_BINARY_OP(zeta);
+REGISTER_FLOAT_BINARY_OP(logaddexp);
+REGISTER_INT2FLOAT_BINARY_OP(logaddexp);
+REGISTER_FLOAT_BINARY_OP(logaddexp2);
+REGISTER_INT2FLOAT_BINARY_OP(logaddexp2);
 REGISTER_FLOAT_BINARY_OP(xlog1py);
 REGISTER_INT2FLOAT_BINARY_OP(xlog1py);
 REGISTER_FLOAT_BINARY_OP(chebyshev_polynomial_t);
@@ -463,6 +489,8 @@ REGISTER_BINARY_OP(add, float2, float2);
 REGISTER_BINARY_OP(add, half2, half2);
 REGISTER_BINARY_OP(sub, float2, float2);
 REGISTER_BINARY_OP(sub, half2, half2);
+REGISTER_BINARY_OP(logaddexp, float2, float2);
+REGISTER_BINARY_OP(logaddexp, half2, half2);
 REGISTER_BINARY_ALPHA_OP(add_alpha, float2, float2, float2);
 REGISTER_BINARY_ALPHA_OP(add_alpha, half2, half2, half2);
 REGISTER_BINARY_ALPHA_OP(sub_alpha, float2, float2, float2);

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

@@ -89,6 +89,14 @@ static void zeta_mps_kernel(TensorIteratorBase& iter) {
   lib.exec_binary_kernel(iter, "zeta");
 }
 
+static void logaddexp_mps_kernel(TensorIteratorBase& iter) {
+  lib.exec_binary_kernel(iter, "logaddexp");
+}
+
+static void logaddexp2_mps_kernel(TensorIteratorBase& iter) {
+  lib.exec_binary_kernel(iter, "logaddexp2");
+}
+
 static void xlog1py_mps_kernel(TensorIteratorBase& iter) {
   TORCH_CHECK_TYPE(isFloatingType(iter.common_dtype()), "xlog1py_mps not implemented for non-floating types");
   lib.exec_binary_kernel(iter, "xlog1py");
@@ -211,6 +219,8 @@ REGISTER_DISPATCH(fmin_stub, &fmin_mps_kernel)
 REGISTER_DISPATCH(copysign_stub, &copysign_mps_kernel)
 REGISTER_DISPATCH(nextafter_stub, &nextafter_mps_kernel)
 REGISTER_DISPATCH(zeta_stub, &zeta_mps_kernel)
+REGISTER_DISPATCH(logaddexp_stub, &logaddexp_mps_kernel);
+REGISTER_DISPATCH(logaddexp2_stub, &logaddexp2_mps_kernel);
 REGISTER_DISPATCH(xlog1py_stub, &xlog1py_mps_kernel)
 REGISTER_DISPATCH(chebyshev_polynomial_t_stub, &chebyshev_polynomial_t_mps_kernel)
 REGISTER_DISPATCH(chebyshev_polynomial_u_stub, &chebyshev_polynomial_u_mps_kernel)

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

@@ -17,8 +17,6 @@
 #include <ATen/ops/ge_native.h>
 #include <ATen/ops/gt_native.h>
 #include <ATen/ops/le_native.h>
-#include <ATen/ops/logaddexp2_native.h>
-#include <ATen/ops/logaddexp_native.h>
 #include <ATen/ops/logical_and_native.h>
 #include <ATen/ops/logical_or_native.h>
 #include <ATen/ops/logical_xor_native.h>
@@ -277,30 +275,6 @@ TORCH_IMPL_FUNC(pow_Scalar_out_mps)(const Scalar& base, const Tensor& exp, const
   }
 }
 
-TORCH_IMPL_FUNC(logaddexp_out_mps)(const Tensor& self, const Tensor& other, const Tensor& output) {
-  mps::BinaryOpBlock logaddexp_op_block = ^BinaryOpFn(cachedGraph, primaryCastTensor, secondaryCastTensor) {
-    MPSGraph* mpsGraph = cachedGraph->graph();
-    MPSGraphTensor* sumTensor =
-        [mpsGraph additionWithPrimaryTensor:[mpsGraph exponentWithTensor:primaryCastTensor name:nil]
-                            secondaryTensor:[mpsGraph exponentWithTensor:secondaryCastTensor name:nil]
-                                       name:nil];
-    return [mpsGraph logarithmWithTensor:sumTensor name:nil];
-  };
-  mps::binaryOpTensor(self, other, output, "logaddexp_out_mps", logaddexp_op_block);
-}
-
-TORCH_IMPL_FUNC(logaddexp2_out_mps)(const Tensor& self, const Tensor& other, const Tensor& output) {
-  mps::BinaryOpBlock logaddexp2_op_block = ^BinaryOpFn(cachedGraph, primaryCastTensor, secondaryCastTensor) {
-    MPSGraph* mpsGraph = cachedGraph->graph();
-    MPSGraphTensor* sumTensor =
-        [mpsGraph additionWithPrimaryTensor:[mpsGraph exponentBase2WithTensor:primaryCastTensor name:nil]
-                            secondaryTensor:[mpsGraph exponentBase2WithTensor:secondaryCastTensor name:nil]
-                                       name:nil];
-    return [mpsGraph logarithmBase2WithTensor:sumTensor name:nil];
-  };
-  mps::binaryOpTensor(self, other, output, "logaddexp2_out_mps", logaddexp2_op_block);
-}
-
 TORCH_IMPL_FUNC(xlogy_out_mps)(const Tensor& self, const Tensor& other, const Tensor& output) {
   mps::BinaryOpBlock xlogy_op_block = ^BinaryOpFn(cachedGraph, primaryCastTensor, secondaryCastTensor) {
     MPSGraph* mpsGraph = cachedGraph->graph();

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

@@ -57,6 +57,7 @@ Tensor& random_mps_impl(Tensor& self,
   if (self.numel() == 0) {
     return self;
   }
+  at::assert_no_internal_overlap(self);
   // MPS random is broken for 5D+ tensors, see https://github.com/pytorch/pytorch/issues/147624
   const auto need_reshape = self.ndimension() > 4;
   auto mps_gen = get_generator_or_default<MPSGeneratorImpl>(gen, at::mps::detail::getDefaultMPSGenerator());
@@ -153,8 +154,16 @@ Tensor& random_mps_impl(Tensor& self,
       feeds[meanPlaceholder.getMPSGraphTensor()] = meanPlaceholder.getMPSGraphTensorData();
     }
 
-    Placeholder outputPlaceholder = Placeholder(cachedGraph->resultTensor, self);
+    // Handle non-contiguous output tensors by creating a contiguous temporary
+    const auto needs_gather = needsGather(self);
+    Tensor self_ = needs_gather ? at::empty_like(self, MemoryFormat::Contiguous) : self;
+    Placeholder outputPlaceholder = Placeholder(cachedGraph->resultTensor, self_);
     runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
+
+    // Copy results back to original non-contiguous output
+    if (needs_gather) {
+      self.copy_(self_);
+    }
   }
 
   return self;

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

@@ -617,6 +617,9 @@ Tensor& index_select_out_mps(const Tensor& self, int64_t dim, const Tensor& inde
   TORCH_CHECK(self.scalar_type() == output.scalar_type(),
               "index_select(): self and output must have the same scalar type");
   TORCH_CHECK(dim == 0 || dim < self.dim(), "index_select(): Indexing dim ", dim, " is out of bounds of tensor");
+  at::assert_no_internal_overlap(output);
+  at::assert_no_overlap(output, self);
+  at::assert_no_overlap(output, index);
   auto output_size = self.sizes().vec();
   if (self.dim() > 0) {
     output_size[dim] = num_indices;

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

@@ -370,7 +370,7 @@ static void nllnd_loss_backward_impl(Tensor& grad_input_arg,
                                                      onValue:-1.0f
                                                     offValue:0.0f
                                                         name:nil];
-      oneHotTensor = castMPSTensor(mpsGraph, oneHotTensor, inputTensor.dataType);
+      oneHotTensor = castMPSTensor(mpsGraph, oneHotTensor, [inputTensor dataType]);
       if (isWeightsArrayValid) {
         oneHotTensor = [mpsGraph multiplicationWithPrimaryTensor:oneHotTensor
                                                  secondaryTensor:weightTensor
@@ -705,6 +705,7 @@ static void smooth_l1_loss_template(const Tensor& input,
   TORCH_CHECK(beta >= 0, "smooth_l1_loss does not support negative values for beta.");
   TORCH_CHECK(input.is_mps());
   TORCH_CHECK(target.is_mps());
+  TORCH_CHECK_NOT_IMPLEMENTED(input.scalar_type() != kLong, "MPS doesn't know how to do square_i64");
   if ((input.numel() == 0) || (target.numel() == 0)) {
     reduction == Reduction::Mean ? output.fill_(std::numeric_limits<float>::quiet_NaN()) : output.zero_();
     return;
@@ -771,7 +772,7 @@ static void smooth_l1_loss_backward_impl(const Tensor& grad_output,
       MPSGraphTensor* targetTensor = mpsGraphRankedPlaceHolder(mpsGraph, target);
       MPSGraphTensor* gradOutputTensor = mpsGraphRankedPlaceHolder(mpsGraph, grad_output);
 
-      MPSGraphTensor* betaTensor = [mpsGraph constantWithScalar:beta dataType:MPSDataTypeFloat32];
+      MPSGraphTensor* betaTensor = [mpsGraph constantWithScalar:beta dataType:[inputTensor dataType]];
       // xn - yn
       MPSGraphTensor* diffTensor = [mpsGraph subtractionWithPrimaryTensor:inputTensor
                                                           secondaryTensor:targetTensor
@@ -797,7 +798,8 @@ static void smooth_l1_loss_backward_impl(const Tensor& grad_output,
                                                                   name:@"lossTensor"];
       MPSGraphTensor* outputTensor = lossTensor;
       if (reduction == Reduction::Mean) {
-        MPSGraphTensor* numelTensor = [mpsGraph constantWithScalar:(double)input.numel() dataType:MPSDataTypeFloat32];
+        MPSGraphTensor* numelTensor = [mpsGraph constantWithScalar:(double)input.numel()
+                                                          dataType:[lossTensor dataType]];
         outputTensor = [mpsGraph divisionWithPrimaryTensor:lossTensor secondaryTensor:numelTensor name:nil];
       }
       MPSGraphTensor* gradInputTensor = [mpsGraph multiplicationWithPrimaryTensor:outputTensor

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

@@ -84,6 +84,9 @@ std::tuple<Tensor&, Tensor&, Tensor&> batch_norm_mps_out(const Tensor& self,
                                                          Tensor& output,
                                                          Tensor& save_mean,
                                                          Tensor& save_var) {
+  TORCH_CHECK_NOT_IMPLEMENTED(self.scalar_type() != kLong, "Long batch norm is not supported with MPS");
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()),
+                              "Batch norm for complex is not supported for MPS");
   using namespace at::native::mps;
   struct CachedGraph : public MPSCachedGraph {
     CachedGraph(MPSGraph* graph) : MPSCachedGraph(graph) {}
@@ -918,6 +921,7 @@ std::tuple<Tensor, Tensor, Tensor> layer_norm_mps(const Tensor& input,
   // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
   const int axis = input_ndim - normalized_ndim;
   MPSStream* stream = getCurrentMPSStream();
+  TORCH_CHECK_NOT_IMPLEMENTED(input.scalar_type() != kLong, "Not implemented for long on MPS");
   @autoreleasepool {
     mps::dispatch_sync_with_rethrow(stream->queue(), ^() {
       // which kernel variant to use based on the normalized axis N size

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

@@ -1028,15 +1028,18 @@ TORCH_IMPL_FUNC(prod_out_mps)
 }
 
 TORCH_IMPL_FUNC(amax_out_mps)(const Tensor& input_t, IntArrayRef dim, bool keepdim, const Tensor& output_t) {
+  TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "amax is not defined for complex types");
   reduction_out_mps(input_t, dim, keepdim, std::nullopt, output_t, MPSReductionType::AMAX, "amax_out_mps");
 }
 
 TORCH_IMPL_FUNC(amin_out_mps)(const Tensor& input_t, IntArrayRef dim, bool keepdim, const Tensor& output_t) {
+  TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "amin is not defined for complex types");
   reduction_out_mps(input_t, dim, keepdim, std::nullopt, output_t, MPSReductionType::AMIN, "amin_out_mps");
 }
 
 TORCH_IMPL_FUNC(aminmax_out_mps)
 (const Tensor& input_t, std::optional<int64_t> dim_opt, bool keepdim, const Tensor& min_t, const Tensor& max_t) {
+  TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "aminmax is not defined for complex types");
   reduction_out_mps(input_t,
                     dim_opt.has_value() ? OptionalIntArrayRef({*dim_opt}) : std::nullopt,
                     keepdim,

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

@@ -31,6 +31,7 @@ void kthvalue_out_mps_impl(const Tensor& self, int64_t k, int64_t dim, Tensor& v
     indices.copy_(values.toType(at::ScalarType::Long));
     return;
   }
+  TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()), "kthvalue is not implemented for complex types");
   // issue #154890, raising error to prevent crash within MPSGraph until
   // workaround is implemented.
   TORCH_CHECK(self.dim() - dim <= 4, "On-going issue on MPSGraph topk when ndims() - axis > 4, see issue #154890");

aten/src/ATen/native/native_functions.yaml

@@ -3622,8 +3622,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: logaddexp_out
+    CPU, CUDA, MPS: logaddexp_out
-    MPS: logaddexp_out_mps
   tags: pointwise
 
 - func: logaddexp(Tensor self, Tensor other) -> Tensor
@@ -3635,8 +3634,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: logaddexp2_out
+    CPU, CUDA, MPS: logaddexp2_out
-    MPS: logaddexp2_out_mps
   tags: pointwise
 
 - func: logaddexp2(Tensor self, Tensor other) -> Tensor
@@ -8867,11 +8865,11 @@
   autogen: bitwise_right_shift.Scalar_Tensor_out
   tags: pointwise
 
-- func: tril_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
+- func: tril_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!)
   structured_delegate: tril.out
   variants: method
 
-- func: triu_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
+- func: triu_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!)
   structured_delegate: triu.out
   variants: method
 
@@ -8995,25 +8993,25 @@
 - func: cross(Tensor self, Tensor other, int? dim=None) -> Tensor
   variants: method, function
 
-- func: triu.out(Tensor self, int diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
+- func: triu.out(Tensor self, SymInt diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
     CPU: triu_cpu
     CUDA: triu_cuda
     MPS: triu_mps_out
 
-- func: triu(Tensor self, int diagonal=0) -> Tensor
+- func: triu(Tensor self, SymInt diagonal=0) -> Tensor
   structured_delegate: triu.out
   variants: method, function
 
-- func: tril.out(Tensor self, int diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
+- func: tril.out(Tensor self, SymInt diagonal=0, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
     CPU: tril_cpu
     CUDA: tril_cuda
     MPS: tril_mps_out
 
-- func: tril(Tensor self, int diagonal=0) -> Tensor
+- func: tril(Tensor self, SymInt diagonal=0) -> Tensor
   structured_delegate: tril.out
   variants: method, function
 

aten/src/ATen/native/quantized/cpu/UpSampleBilinear2d.cpp

@@ -73,8 +73,7 @@ void upsample_bilinear2d_out_frame(
   const auto rwidth = area_pixel_compute_scale<float>(
       input_width, output_width, align_corners, scales_w);
 
-  // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
+  float output_scale = static_cast<float>(output.q_scale() / input.q_scale());
-  float output_scale = output.q_scale() / input.q_scale();
 
   const int64_t input_q_zero_point = input.q_zero_point();
   const int64_t output_q_zero_point = output.q_zero_point();

aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp

@@ -148,7 +148,7 @@ Tensor qcat_nhwc_kernel(
           // Vectorized loop
           if (c + VLEN <= curr_C) {
             auto curr_scale_vec = Vectorized<float>(curr_scale);
-            auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
+            auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
             auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
             for (; c + VLEN <= curr_C; c += VLEN) {
               auto inp_vec = Vec::loadu(iptr + c);
@@ -174,7 +174,7 @@ Tensor qcat_nhwc_kernel(
           int64_t elem_size = curr_C - c;
           if ((VLEN == 4 * kVLEN) && elem_size >= kVLEN) {
             auto curr_scale_vec = Vectorized<float>(curr_scale);
-            auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
+            auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
             auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
             int64_t vec_num = elem_size / kVLEN;
             std::array<typename scalar_t::underlying, VLEN> buf_in{};
@@ -611,12 +611,10 @@ void qrelu_kernel(const Tensor& qx, Tensor& qy) {
 void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
                                    const Scalar& negval_) {
   int64_t i_zp = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float i_scale = static_cast<float>(qx.q_scale());
-  float i_scale = qx.q_scale();
 
   int64_t o_zp = out.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float o_scale = static_cast<float>(out.q_scale());
-  float o_scale = out.q_scale();
   float o_inv_scale = 1.0f / o_scale;
 
   float negval = negval_.to<float>();
@@ -627,8 +625,8 @@ void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
     Vec zero_vec = Vec(0.0f);
     Vec one_vec = Vec(1.0f);
 
-    Vec i_scale_vec = Vec((float)i_scale);
+    Vec i_scale_vec = Vec(i_scale);
-    Vec i_zp_vec = Vec((float)i_zp);
+    Vec i_zp_vec = Vec(i_zp);
     Vec i_scale_zp_neg_premul_vec = i_scale_vec * i_zp_vec.neg();
 
     Vec negval_vec = Vec(negval);
@@ -738,10 +736,9 @@ void qprelu_out_kernel(Tensor& out,
 
 void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float scale = static_cast<float>(qx.q_scale());
-  float scale = qx.q_scale();
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
   int64_t output_zero_point = zero_point;
   float output_scale = scale;
@@ -828,10 +825,9 @@ void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
 void qsigmoid_kernel(
     const Tensor& qx, Tensor& qy, double output_scale, int64_t output_zero_point ) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float scale = static_cast<float>(qx.q_scale());
-  float scale = qx.q_scale();
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qsigmoid", [&]() {
     float inv_output_scale = 1.0 / output_scale;
@@ -870,10 +866,9 @@ void qsigmoid_kernel(
 
 void qhardsigmoid_kernel(const Tensor& qx, Tensor& qy) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float scale = static_cast<float>(qx.q_scale());
-  float scale = qx.q_scale();
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qhardsigmoid", [&]() {
@@ -1029,13 +1024,10 @@ void qthreshold_kernel(
 
   // defines input and output scales and zero_points
   int64_t input_zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float input_scale = static_cast<float>(qx.q_scale());
-  float input_scale = qx.q_scale();
   int64_t output_zero_point = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float output_scale = static_cast<float>(qy.q_scale());
-  float output_scale = qy.q_scale();
+  float inv_output_scale = static_cast<float>(1.0 / output_scale);
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float inv_output_scale = 1.0 / output_scale;
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qthreshold", [&]() {
     qy = at::_empty_affine_quantized(
@@ -1096,8 +1088,7 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
 
   const auto o_scale = qy.q_scale();
   const auto o_zero_point = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  const float o_inv_scale = static_cast<float>(1.0 / o_scale);
-  const float o_inv_scale = 1.0 / o_scale;
 
   using fVec = Vectorized<float>;
   fVec i_scale_vec(i_scale);
@@ -1135,10 +1126,9 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
 
 void qtanh_kernel(const Tensor& qx, Tensor& qy) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float scale = static_cast<float>(qx.q_scale());
-  float scale = qx.q_scale();
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qtanh", [&]() {
@@ -1198,16 +1188,13 @@ void qelu_kernel(
   // they are NOT related to the quantization scale term
 
   int64_t i_zp = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float i_scale = static_cast<float>(qx.q_scale());
-  float i_scale = qx.q_scale();
 
   // In a future PR, we can improve on output scale and zero_point
   // selection.
   int64_t o_zp = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float o_scale = static_cast<float>(qy.q_scale());
-  float o_scale = qy.q_scale();
+  float inv_o_scale = static_cast<float>(1.0 / o_scale);
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float inv_o_scale = 1.0 / o_scale;
 
   float alpha_float = alpha.to<float>();
   float scale_coef = scale.to<float>();
@@ -1227,7 +1214,7 @@ void qelu_kernel(
     Vec scale_coef_vec = Vec(scale_coef);
     Vec input_scale_coef_vec = Vec(input_scale_coef);
     Vec i_scale_vec = Vec(i_scale);
-    Vec i_zero_point_vec = Vec((float)i_zp);
+    Vec i_zero_point_vec = Vec(i_zp);
     Vec i_scale_neg_zp_premul_vec = i_scale_vec * i_zero_point_vec.neg();
 
     cpu_kernel_vec(
@@ -1326,23 +1313,20 @@ void qadd_scalar_kernel(Tensor& out, const Tensor& self, const Scalar& other) {
 template <bool ReLUFused = false>
 void qadd_kernel(Tensor& out, const Tensor& self, const Tensor& other) {
   int64_t zero_point = out.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float scale = static_cast<float>(out.q_scale());
-  float scale = out.q_scale();
   float inv_scale = 1.0f / scale;
   int64_t self_zero_point = self.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float self_scale = static_cast<float>(self.q_scale());
-  float self_scale = self.q_scale();
   int64_t other_zero_point = other.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  float other_scale = static_cast<float>(other.q_scale());
-  float other_scale = other.q_scale();
 
   // Broadcast out the parameters here to amortize out that cost across
   // loop iterations.
   // TODO: we can optimize dequantization by doing a premultiplication
   // of the zero point by scale and doing FMA on scale*x_q - (scale*zero_point)
-  auto self_zero_point_vec = Vectorized<float>((float)self_zero_point);
+  auto self_zero_point_vec = Vectorized<float>(self_zero_point);
   auto self_scale_vec = Vectorized<float>(self_scale);
-  auto other_zero_point_vec = Vectorized<float>((float)other_zero_point);
+  auto other_zero_point_vec = Vectorized<float>(other_zero_point);
   auto other_scale_vec = Vectorized<float>(other_scale);
 
   auto self_scale_neg_zp_premul_vec = self_scale_vec * self_zero_point_vec.neg();
@@ -2965,7 +2949,7 @@ void quantized_normalize_kernel(
     const bool beta_null = beta_data == nullptr;
     int64_t x_zp = X.q_zero_point();
     float x_scale = X.q_scale();
-    fVec x_zp_vec((float)x_zp);
+    fVec x_zp_vec(x_zp);
     fVec one_vec(1.0f);
     fVec zero_vec(0.0f);
     float x_fake_scale = 1.0f;
@@ -3253,7 +3237,7 @@ void quantized_groupnorm_nhwc_kernel(
     const bool beta_null = beta_data == nullptr;
     int64_t x_zp = X.q_zero_point();
     float x_scale = X.q_scale();
-    fVec x_zp_vec((float)x_zp);
+    fVec x_zp_vec(x_zp);
     fVec one_vec(1.0f);
     fVec zero_vec(0.0f);
     float x_fake_scale = 1.0f;

aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp

@@ -414,7 +414,6 @@ at::Tensor& PackedLinearWeightFp16::apply_dynamic_impl(
   TORCH_CHECK(input.size(input.dim() - 1) == packed_weight_fp16.numRows())
   TORCH_CHECK(input.dim() >= 2);
 
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
   const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
   const int64_t N = packed_weight_fp16.numCols();
   std::vector<int64_t> output_sizes = input.sizes().vec();

aten/src/ATen/native/sparse/SparseTensor.cpp

@@ -467,6 +467,28 @@ Tensor sparse_coo_tensor(const Tensor& indices, const Tensor& values, IntArrayRe
       !options.has_layout() || options.layout() == kSparse,
       "expected sparse layout, but got layout ",
       options.layout());
+
+  if (indices.numel() > 0) {
+    Tensor min_indices =
+        std::get</* values */ 0>(indices.min(/* dim */ 1, /* keepdim */ false));
+    Tensor cpu_min_indices;
+    if (!indices.is_cpu()) {
+      cpu_min_indices = min_indices.to(at::DeviceType::CPU);
+    } else {
+      cpu_min_indices = min_indices;
+    }
+    auto cpu_min_indices_accessor = cpu_min_indices.accessor<int64_t, 1>();
+    for (const auto d : c10::irange(indices.size(0))) {
+      int64_t min_index_in_dim = cpu_min_indices_accessor[d];
+      TORCH_CHECK(
+          min_index_in_dim >= 0,
+          "found negative index ",
+          min_index_in_dim,
+          " for dim ",
+          d);
+    }
+  }
+
   return at::native::_sparse_coo_tensor_unsafe(
       indices,
       values,

aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp

@@ -22,6 +22,7 @@
 #else
 #include <ATen/ops/empty.h>
 #include <ATen/ops/empty_like.h>
+#include <ATen/ops/zeros_like.h>
 #include <ATen/ops/reshape.h>
 #include <ATen/ops/scalar_tensor.h>
 #include <ATen/ops/sum.h>
@@ -42,7 +43,6 @@ C10_DIAGNOSTIC_POP()
 #include <static_switch.h>
 #include <ATen/native/transformers/cuda/flash_attn/flash_api.h>
 
-
 #include <c10/util/Exception.h>
 
 namespace FLASH_NAMESPACE {
@@ -417,6 +417,26 @@ mha_fwd(const at::Tensor &q,         // batch_size x seqlen_q x num_heads x head
     const int head_size_og = sizes[3];
     const int seqlen_k = k.size(1);
     const int num_heads_k = k.size(2);
+
+    if (batch_size == 0) {
+        auto opts = q.options();
+        at::Tensor out = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
+        at::Tensor q_padded = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
+        at::Tensor k_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
+        at::Tensor v_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
+        at::Tensor softmax_lse = at::empty({0, num_heads, seqlen_q}, opts.dtype(at::kFloat));
+        at::Tensor rng_state = at::empty({2}, at::dtype(c10::kUInt64).device(at::kCUDA));
+        at::Tensor _unused = at::empty({}, at::dtype(c10::kUInt64).device(at::kCUDA));
+        at::Tensor p = at::empty({0}, opts);
+        if (return_softmax) {
+            auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
+            const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
+            const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
+            p = at::empty({0, num_heads, seqlen_q_rounded, seqlen_k_rounded}, opts);
+        }
+        return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), _unused, std::move(p)};
+    }
+
     TORCH_CHECK(batch_size > 0, "batch size must be positive");
     TORCH_CHECK(head_size_og % 8 == 0, "head_size must be a multiple of 8, this is ensured by padding!");
     TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
@@ -547,7 +567,7 @@ mha_fwd(const at::Tensor &q,         // batch_size x seqlen_q x num_heads x head
         q_padded = q_padded.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
         softmax_lse = softmax_lse.reshape({batch_size, num_heads_k * seqlen_q, 1});
     }
-    return {out, q_padded, k_padded, v_padded, softmax_lse, rng_state, _unused, p};
+    return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), std::move(_unused), std::move(p)};
 }
 
 std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
@@ -852,7 +872,6 @@ mha_bwd(const at::Tensor &dout,  // batch_size x seqlen_q x num_heads, x head_si
     TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
     TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
     TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
-    TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
 
     const auto sizes = q.sizes();
 
@@ -863,6 +882,20 @@ mha_bwd(const at::Tensor &dout,  // batch_size x seqlen_q x num_heads, x head_si
     const int head_size = sizes[3];
     const int seqlen_k = k.size(1);
     const int num_heads_k = k.size(2);
+
+    if (batch_size == 0) {
+        auto opts = q.options();
+        at::Tensor dq = at::empty_like(q);
+        at::Tensor dk = at::empty_like(k);
+        at::Tensor dv = at::empty_like(v);
+        auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
+        const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
+        at::Tensor softmax_d = at::empty({0, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
+        return {dq, dk, dv, softmax_d};
+    }
+
+    TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
+
     TORCH_CHECK(batch_size > 0, "batch size must be positive");
     TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
     TORCH_CHECK(head_size_og % 8 == 0, "head_size_og should be a multiple of 8, this is ensured by padding!");

benchmarks/dynamo/common.py

@@ -1837,6 +1837,10 @@ class BenchmarkRunner:
     def skip_models_for_cuda(self):
         return set()
 
+    @property
+    def skip_models_for_xpu(self):
+        return set()
+
     @property
     def skip_models_for_cpu(self):
         return set()
@@ -3927,6 +3931,8 @@ def run(runner, args, original_dir=None):
             runner.skip_models.update(runner.skip_models_for_cpu_aarch64)
     elif args.devices == ["cuda"]:
         runner.skip_models.update(runner.skip_models_for_cuda)
+    elif args.devices == ["xpu"]:
+        runner.skip_models.update(runner.skip_models_for_xpu)
 
     if not args.multiprocess:
         runner.skip_models.update(runner.skip_multiprocess_models)

benchmarks/dynamo/genai_layers/benchmark.py

@@ -56,6 +56,20 @@ def list_benchmarks():
     print(f"Available benchmarks: {list(BENCHMARK_REGISTRY.keys())}")
 
 
+def _run_benchmark(
+    benchmark_cls,
+    script_args,
+):
+    benchmark = benchmark_cls(script_args)
+    benchmark.benchmark()
+    benchmark.report_geomean_speedup()
+    if script_args.print_benchmark_result:
+        print(f"Benchmarking results {benchmark.name}:")
+        print(benchmark.profiling_results)
+    if script_args.visualize:
+        benchmark.visualize()
+
+
 def run_benchmark(
     benchmark_name: str,
     script_args,
@@ -71,10 +85,7 @@ def run_benchmark(
     print("=" * 60)
 
     benchmark_class = BENCHMARK_REGISTRY[benchmark_name]
-    benchmark = benchmark_class(script_args)
+    _run_benchmark(benchmark_class, script_args)
-    benchmark.benchmark()
-    if script_args.visualize:
-        benchmark.visualize()
 
     return True
 
@@ -87,10 +98,7 @@ def run_all_benchmarks(script_args):
 
     for name, cls in BENCHMARK_REGISTRY.items():
         print(f"\n{'=' * 20} {name.upper()} {'=' * 20}")
-        benchmark = cls(script_args)
+        _run_benchmark(cls, script_args)
-        benchmark.benchmark()
-        if script_args.visualize:
-            benchmark.visualize()
         print()
 
 
@@ -149,8 +157,43 @@ Examples:
         help="Whether to exit with an error message for accuracy failure",
     )
 
+    parser.add_argument(
+        "--print-benchmark-result",
+        action="store_true",
+        help="Whether to print the raw benchmarking result. Easier to quickly check the benchmark results on a server without GUI",
+    )
+
+    parser.add_argument(
+        "--custom-compile-name",
+        type=str,
+        default=None,
+        help="Name for the curve with customized compilation options",
+    )
+
+    parser.add_argument(
+        "--custom-compile-options",
+        type=str,
+        default=None,
+        help="Json string for the custom compile options.",
+    )
+
     args = parser.parse_args()
 
+    if args.custom_compile_options:
+        import json
+
+        try:
+            args.custom_compile_options = json.loads(args.custom_compile_options)
+        except json.decoder.JSONDecodeError as e:
+            raise RuntimeError(
+                f"Invalid json string for --custom-compile-options: {args.custom_compile_options}"
+            ) from e
+
+        if not args.custom_compile_options:
+            raise RuntimeError("Found no options for --custom-compile-options")
+        if not args.custom_compile_name:
+            raise RuntimeError("Missing label name for the custom compilation")
+
     # Handle list option
     if args.list:
         list_benchmarks()

benchmarks/dynamo/genai_layers/kernels.py

@@ -8,6 +8,15 @@ import torch
 import torch.nn.functional as F
 
 
+# more important shapes used by internal models
+extra_shapes_for_norm = (
+    (1152 * 500, 384),
+    (1152 * 500, 512),
+    (1152 * 1000, 384),
+    (1152 * 1000, 512),
+)
+
+
 class CrossEntropyForward(BenchmarkKernel):
     def __init__(self, script_args):
         super().__init__(script_args)
@@ -346,7 +355,7 @@ class RMSNormForward(BenchmarkKernel):
             (32768, 65536),
             (16384, 131072),
             (8192, 262144),
-        )
+        ) + extra_shapes_for_norm
 
     def get_memory_bytes(self, args, kwargs) -> int:
         x, w = args
@@ -438,8 +447,7 @@ class RMSNormBackward(BenchmarkKernel):
             (32768, 4096),
             (32768, 8192),
             (32768, 16384),
-            (32768, 32768),
+        ) + extra_shapes_for_norm
-        )
 
     def get_memory_bytes(self, args, kwargs) -> int:
         x, w, dy = args
@@ -553,7 +561,7 @@ class LayerNormForward(BenchmarkKernel):
             (32768, 16384),
             (32768, 32768),
             (32768, 65536),
-        )
+        ) + extra_shapes_for_norm
 
     def get_memory_bytes(self, args, kwargs) -> int:
         x, w = args
@@ -627,7 +635,7 @@ class LayerNormBackward(BenchmarkKernel):
             (32768, 16384),
             (32768, 32768),
             (32768, 65536),
-        )
+        ) + extra_shapes_for_norm
 
     def get_memory_bytes(self, args, kwargs) -> int:
         x, w, dy = args

benchmarks/dynamo/genai_layers/utils.py

@@ -6,6 +6,7 @@ from dataclasses import dataclass
 from typing import Any, Optional
 
 import matplotlib.pyplot as plt
+from scipy.stats import gmean
 
 import torch
 from torch._inductor.runtime.benchmarking import benchmarker
@@ -107,6 +108,18 @@ class BenchmarkKernel:
         for backend in self.available_backends:
             args_ref, kwargs_ref = self.clone_inputs(args, kwargs)
             res[backend] = getattr(self, backend)(args_ref, kwargs_ref)()
+
+        if (
+            "compiled" in self.available_backends
+            and self.script_args.custom_compile_options
+        ):
+            torch._dynamo.reset()  # cause recompile
+            with torch._inductor.config.patch(self.script_args.custom_compile_options):
+                args_ref, kwargs_ref = self.clone_inputs(args, kwargs)
+                res[self.script_args.custom_compile_name] = self.compiled(
+                    args_ref, kwargs_ref
+                )()
+
         gold = res["eager"]
 
         tol = {}
@@ -115,7 +128,7 @@ class BenchmarkKernel:
                 "atol": self.script_args.tolerance,
                 "rtol": self.script_args.tolerance,
             }
-        for backend in self.available_backends:
+        for backend in res:
             if backend == "eager":
                 continue
             try:
@@ -134,37 +147,83 @@ class BenchmarkKernel:
                     print("Exit right away since --exit-on-accuracy-failure is set")
                     sys.exit(1)
 
-    def benchmark_single_shape(
+    def benchmark_single_shape_for_backend(
-        self, args, kwargs=None, should_check_accuracy=True, setting: str = ""
+        self, backend, args, kwargs, setting, fn=None
-    ):
+    ) -> bool:
-        for backend in self.available_backends:
+        if fn is None:
+            fn = getattr(self, backend)
         args_ref, kwargs_ref = self.clone_inputs(args, kwargs)
         try:
-                avg_time = benchmark_kernel_in_milliseconds(
+            avg_time = benchmark_kernel_in_milliseconds(fn(args_ref, kwargs_ref))
-                    getattr(self, backend)(args_ref, kwargs_ref)
-                )
         except Exception as e:
             print(
                 f"Failed to run {backend} backend on {self.name} kernel for {setting} due to {e}"
             )
             self.available_backends.remove(backend)  # noqa: B909
-                continue
+            return False
         mem_bytes = self.get_memory_bytes(args_ref, kwargs_ref)
         perf = Performance(setting, avg_time, mem_bytes)
         print(f"{self.name} kernel on {backend} backend. {perf}")
         self.profiling_results[backend].append(perf)
+        return True
+
+    def benchmark_single_shape(
+        self, args, kwargs=None, should_check_accuracy=True, setting: str = ""
+    ):
+        for backend in self.available_backends:
+            self.benchmark_single_shape_for_backend(backend, args, kwargs, setting)
+        if (
+            "compiled" in self.available_backends
+            and self.script_args.custom_compile_options
+        ):
+            torch._dynamo.reset()  # cause recompile
+            with torch._inductor.config.patch(self.script_args.custom_compile_options):
+                status = self.benchmark_single_shape_for_backend(
+                    self.script_args.custom_compile_name,
+                    args,
+                    kwargs,
+                    setting,
+                    fn=self.compiled,
+                )
+            if not status:
+                self.script_args.custom_compile_options = (
+                    None  # once fail, don't run again
+                )
 
         if should_check_accuracy:
             self.check_accuracy(args, kwargs)
 
     def visualize(self) -> None:
+        device_name = torch.cuda.get_device_name(0)
         visualize_comparison(
             self.profiling_results,
-            title=f"{self.name}",
+            title=f"{self.name} ({device_name})",
             output_path=f"{self.name}_bench",
         )
         return
 
+    def report_geomean_speedup(self) -> None:
+        print(f"Geomean speedup for benchmark {self.name}")
+        eager_result = {
+            result.setting: result for result in self.profiling_results["eager"]
+        }
+        print(f"  eager {len(eager_result)} data points")
+        for backend, backend_result in self.profiling_results.items():
+            if backend == "eager":
+                continue
+            speeduplist = []
+            for result in backend_result:
+                eager_latency = eager_result[result.setting].latency
+                backend_latency = result.latency
+                speeduplist.append(
+                    eager_latency / backend_latency if backend_latency != 0 else 0.0
+                )
+
+            if len(speeduplist) > 0:
+                print(
+                    f"  {backend} {len(speeduplist)} data points, {gmean(speeduplist):.2f}x speedup"
+                )
+
 
 def get_backend_colors() -> dict[str, str]:
     """Get consistent color scheme for different backends."""
@@ -252,5 +311,6 @@ def visualize_comparison(
         os.makedirs("pics", exist_ok=True)
         full_path = os.path.join("pics", output_path + ".png")
         plt.savefig(full_path, dpi=300, bbox_inches="tight", facecolor="white")
+        print(f"Chart saved to {full_path}")
 
     plt.close()

benchmarks/dynamo/timm_models.py

@@ -74,7 +74,8 @@ REQUIRE_HIGHER_TOLERANCE = {
 REQUIRE_HIGHER_TOLERANCE_AMP = {}
 
 REQUIRE_EVEN_HIGHER_TOLERANCE = {
-    "beit_base_patch16_224",
+    "deit_base_distilled_patch16_224",
+    "vit_base_patch16_siglip_256",
 }
 
 # These models need higher tolerance in MaxAutotune mode
@@ -354,7 +355,9 @@ class TimmRunner(BenchmarkRunner):
         if is_training:
             from torch._inductor import config as inductor_config
 
-            if name in REQUIRE_EVEN_HIGHER_TOLERANCE or (
+            if name == "beit_base_patch16_224":
+                tolerance = 16 * 1e-2
+            elif name in REQUIRE_EVEN_HIGHER_TOLERANCE or (
                 inductor_config.max_autotune
                 and name in REQUIRE_EVEN_HIGHER_TOLERANCE_MAX_AUTOTUNE
             ):

benchmarks/dynamo/torchbench.py

@@ -124,6 +124,10 @@ class TorchBenchmarkRunner(BenchmarkRunner):
     def skip_models_for_cuda(self):
         return self._skip["device"]["cuda"]
 
+    @property
+    def skip_models_for_xpu(self):
+        return self._skip["device"]["xpu"]
+
     @property
     def skip_models_for_freezing_cuda(self):
         return self._skip["freezing"]["cuda"]

benchmarks/dynamo/torchbench.yaml

@@ -217,6 +217,9 @@ skip:
 
     cuda: []
 
+    xpu:
+      - *DETECTRON2_MODELS
+
   test:
     training:
       - *DETECTRON2_MODELS

benchmarks/transformer/config_utils.py

@@ -0,0 +1,157 @@
+"""Configuration utilities for parsing JSON and YAML config files."""
+
+import json
+import re
+
+
+def heads_input_type(s: str) -> tuple[int, int]:
+    """Convert string format 'Hq,Hkv' to tuple (Hq, Hkv)."""
+    try:
+        hq, hkv = map(int, s.split(","))
+        return hq, hkv
+    except Exception as e:
+        raise ValueError("Heads must be Hq,Hkv") from e
+
+
+default_config = {
+    "dynamic": False,
+    "calculate_bwd": False,
+    "dtype": "bfloat16",
+    "b": [2, 8, 16],
+    "nh": ["16,16", "16,2"],
+    "s": [512, 1024, 4096],
+    "d": [64, 128],
+    "mods": ["noop", "causal", "alibi", "sliding_window"],
+    "backend": ["efficient"],
+    "max_autotune": False,
+    "decoding": False,
+    "kv_size": None,
+    "throughput": True,
+    "save_path": None,
+    "output_json_for_dashboard": None,
+    "benchmark_name": "PyTorch operator microbenchmark",
+}
+
+
+def load_config_file(config_path: str) -> dict:
+    """Load configuration from JSON or YAML file.
+
+    Automatically converts 'nh' field from strings to tuples.
+
+    Args:
+        config_path: Path to the configuration file
+
+    Returns:
+        Dictionary containing the configuration
+
+    Raises:
+        FileNotFoundError: If config file doesn't exist
+        ValueError: If config file format is invalid
+    """
+    with open(config_path) as f:
+        config_str = f.read()
+
+    # Try to load as JSON first
+    try:
+        config = json.loads(config_str)
+    except json.JSONDecodeError:
+        # Fall back to YAML parsing
+        config = _parse_simple_yaml(config_str)
+
+    # Apply automatic conversions for 'nh' field
+    if "nh" in config and isinstance(config["nh"], list):
+        config["nh"] = [
+            heads_input_type(h) if isinstance(h, str) else h for h in config["nh"]
+        ]
+
+    return config
+
+
+def _parse_simple_yaml(yaml_str: str) -> dict:
+    """Simple YAML parser for basic configs (without external dependencies).
+
+    Supports:
+    - key: value pairs
+    - booleans (true/false)
+    - null values
+    - integers and floats
+    - strings (quoted and unquoted)
+    - lists in JSON format [item1, item2, ...]
+    - comments (lines starting with # or after #)
+
+    Args:
+        yaml_str: YAML content as string
+
+    Returns:
+        Dictionary containing parsed YAML content
+    """
+    config = {}
+
+    for line in yaml_str.split("\n"):
+        # Remove comments
+        line = line.split("#")[0].strip()
+
+        if not line or ":" not in line:
+            continue
+
+        key, value = line.split(":", 1)
+        key = key.strip()
+        value = value.strip()
+
+        # Parse value based on type
+        if value.lower() == "true":
+            config[key] = True
+        elif value.lower() == "false":
+            config[key] = False
+        elif value.lower() in ("null", "none", ""):
+            config[key] = None
+        elif value.startswith("[") and value.endswith("]"):
+            # Parse list - handle quoted strings properly
+            pattern = r'"([^"]+)"|\'([^\']+)\'|([^,\[\]\s]+)'
+            matches = re.findall(pattern, value[1:-1])  # Remove [ ]
+            parsed_items = []
+            for match in matches:
+                # match is a tuple of (double_quoted, single_quoted, unquoted)
+                item = match[0] or match[1] or match[2]
+                item = item.strip()
+                if item:
+                    try:
+                        parsed_items.append(int(item))
+                    except ValueError:
+                        parsed_items.append(item)
+            config[key] = parsed_items
+        elif value.startswith(('"', "'")):
+            config[key] = value.strip("\"'")
+        else:
+            # Try to parse as number
+            try:
+                config[key] = int(value)
+            except ValueError:
+                try:
+                    config[key] = float(value)
+                except ValueError:
+                    config[key] = value
+
+    return config
+
+
+def print_default_config(output_format: str) -> None:
+    """Print a default configuration template in JSON or YAML format.
+
+    Args:
+        output_format: Either "json" or "yaml"
+    """
+    if output_format == "json":
+        print(json.dumps(default_config, indent=2))
+    else:  # yaml
+        for key, value in default_config.items():
+            if value is None:
+                print(f"{key}: null")
+            elif isinstance(value, bool):
+                print(f"{key}: {str(value).lower()}")
+            elif isinstance(value, str):
+                print(f'{key}: "{value}"')
+            elif isinstance(value, list):
+                print(f"{key}: {json.dumps(value)}")
+            else:
+                print(f"{key}: {value}")

benchmarks/transformer/configs/config_basic.yaml

@@ -0,0 +1,29 @@
+# Basic benchmark configuration for PyTorch transformer benchmarks
+# Usage: python score_mod.py --config config_basic.yaml
+
+# Core parameters
+dynamic: false
+calculate_bwd: true
+dtype: "bfloat16"
+
+# Shape parameters - larger sweep
+b: [1, 2, 4, 8, 16]  # batch sizes
+nh: ["16,16", "16,2", "32,32", "32,4"]  # [query_heads,key_value_heads]
+s: [512, 1024, 2048, 4096, 8192]  # sequence lengths
+d: [64, 128]  # head dimensions (limited to 128 for Flash Attention/cuDNN compatibility)
+
+# All attention types
+mods: ["noop", "causal", "rel", "head_bias", "alibi", "sliding_window", "prefix_lm", "softcap"]
+
+# Multiple backends for comparison (SDPA + Flash Attention) - flex is always included internally
+backend: ["efficient", "math", "cudnn", "fav2"]
+max_autotune: true  # Enable torch.compile with max-autotune for optimal performance
+
+# Decoding and cache settings
+decoding: false
+kv_size: null
+
+# Metrics and output
+throughput: true  # Calculate memory bandwidth & TFLOPS
+save_path: "comprehensive_results.csv"  # Save to CSV
+output_json_for_dashboard: "attn_bench_basic.json"

benchmarks/transformer/score_mod.py

@@ -1,15 +1,19 @@
 import argparse
 import csv
+import gc
 import itertools
+import json
 import random
+import sys
 from collections import defaultdict
 from collections.abc import Callable
 from contextlib import nullcontext
 from dataclasses import asdict, dataclass
-from functools import partial
+from functools import partial, wraps
-from typing import Optional, Union
+from typing import Literal, Optional, Union
 
 import numpy as np
+from config_utils import heads_input_type, load_config_file, print_default_config
 from tabulate import tabulate
 from tqdm import tqdm
 
@@ -33,6 +37,96 @@ torch._dynamo.config.recompile_limit = 1000
 from torch._inductor.runtime.benchmarking import benchmarker
 
 
+def cleanup_memory():
+    """Aggressively free GPU memory"""
+    torch.cuda.empty_cache()
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+
+
+def safe_backend(backend_name=None, return_dict=False):
+    """Decorator that wraps backend functions with error handling
+
+    Args:
+        backend_name: Name of the backend for error messages
+        return_dict: If True, returns dict of results for all backends (for run_single_experiment)
+                     If False, returns single ExperimentResults (for individual backend functions)
+    """
+
+    def decorator(func):
+        @wraps(func)
+        def wrapper(config, *args, **kwargs):
+            try:
+                return func(config, *args, **kwargs)
+            except torch.OutOfMemoryError:
+                print(
+                    f"[SKIP] OOM for {backend_name or func.__name__} with shape {config.shape}"
+                )
+                cleanup_memory()
+            except RuntimeError as e:
+                error_msg = str(e)
+                if "out of resource" in error_msg or "OutOfMemoryError" in error_msg:
+                    print(
+                        f"[SKIP] Triton OOM for {backend_name or func.__name__} with shape {config.shape}"
+                    )
+                    cleanup_memory()
+                elif "No valid triton configs" in error_msg:
+                    print(
+                        f"[SKIP] No valid Triton config for {backend_name or func.__name__} with shape {config.shape}"
+                    )
+                else:
+                    print(
+                        f"[SKIP] Runtime error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
+                    )
+            except Exception as e:
+                print(
+                    f"[SKIP] Error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
+                )
+
+            # Return appropriate NaN result based on function type
+            if return_dict:
+                # For run_single_experiment: return dict with NaN for all backends
+                nan_result = ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                )
+                results = dict.fromkeys(config.backends, nan_result)
+                results["flex"] = ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                    sparsity=None,
+                )
+                return results
+            else:
+                # For individual backend functions: return single ExperimentResults
+                return ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                )
+
+        return wrapper
+
+    return decorator
+
+
+# Type definitions
+Backend = Literal["math", "efficient", "cudnn", "fav2", "fav3", "fakv", "og-eager"]
+AttentionType = Literal[
+    "noop",
+    "causal",
+    "rel",
+    "head_bias",
+    "alibi",
+    "sliding_window",
+    "document_mask",
+    "prefix_lm",
+    "softcap",
+]
+DtypeString = Literal["bfloat16", "float16", "float32"]
+SpeedupType = Literal["fwd", "bwd"]
+
+
 def benchmark_torch_function_in_microseconds(func: Callable, *args, **kwargs) -> float:
     # warmup
     for _ in range(5):
@@ -48,6 +142,7 @@ class ExperimentConfig:
     calculate_bwd_time: bool
     cal_bandwidth: bool
     backends: list[str]
+    max_autotune: bool
 
     def __post_init__(self):
         assert len(self.shape) == 6, (
@@ -62,6 +157,7 @@ class ExperimentConfig:
         d.pop("cal_bandwidth", None)
         d["shape(B,Hq,M,Hkv,N,D)"] = d.pop("shape")
         d.pop("backends", None)
+        d.pop("max_autotune", False)
         return d
 
 
@@ -209,6 +305,7 @@ def query_key_value_clones(
     return query_ref, key_ref, value_ref
 
 
+@safe_backend("SDPA")
 def run_single_backend_sdpa(
     config: ExperimentConfig,
     query: torch.Tensor,
@@ -223,6 +320,7 @@ def run_single_backend_sdpa(
     backend_context = get_backend_context(backend)
     with backend_context:
         _device = torch.device("cuda")
+
         eager_sdpa = generate_eager_sdpa(
             config.attn_type, config.shape, config.dtype, block_mask, score_mod
         )
@@ -290,6 +388,7 @@ def run_single_backend_sdpa(
             )
 
 
+@safe_backend("FlashAttention")
 def run_single_backend_FA(
     config: ExperimentConfig,
     query: torch.Tensor,
@@ -301,9 +400,9 @@ def run_single_backend_FA(
     mask_kwargs,
     backend: str,
 ) -> ExperimentResults:
-    assert backend in ["fav2", "fav3", "fakv"]
+    assert backend in ["fav3", "fakv"]
     # Generate callable for specific backend.
-    if backend in ["fav2", "fav3"]:
+    if backend in ["fav3"]:
         FA = generate_FA_callable(
             config.attn_type, config.shape, config.dtype, backend, **mask_kwargs
         )
@@ -354,10 +453,10 @@ def run_single_backend_FA(
     )
 
 
+@safe_backend("flex_attention", return_dict=True)
 def run_single_experiment(
     config: ExperimentConfig,
     dynamic=False,
-    max_autotune=False,
 ) -> dict[str, ExperimentResults]:
     device = torch.device("cuda")
     batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim = config.shape
@@ -377,7 +476,7 @@ def run_single_experiment(
     block_mask, mask_kwargs = generate_block_mask(config.attn_type, config.shape)
     kernel_options = get_kernel_options(config.attn_type, config.shape)
 
-    if max_autotune:
+    if config.max_autotune:
         compiled_sdpa = torch.compile(
             flex_attention, dynamic=dynamic, mode="max-autotune-no-cudagraphs"
         )
@@ -407,7 +506,7 @@ def run_single_experiment(
 
     results = {}
     for backend in config.backends:
-        if backend in ["fav2", "fav3", "fakv"]:
+        if backend in ["fav3", "fakv"]:
             results[backend] = run_single_backend_FA(
                 config,
                 query,
@@ -419,7 +518,7 @@ def run_single_experiment(
                 mask_kwargs,
                 backend,
             )
-        else:  # sdpa
+        else:  # sdpa (also supports fav2)
             results[backend] = run_single_backend_sdpa(
                 config,
                 query,
@@ -440,7 +539,7 @@ def run_single_experiment(
     sparsity = block_mask.sparsity() / 100.0 if block_mask is not None else 0.0
     sparsity = sparsity if config.attn_type != "document_mask" else 0.5
 
-    results["compiled"] = ExperimentResults(
+    results["flex"] = ExperimentResults(
         fwd_time=forward_compiled_time,
         bwd_time=backward_compile_time if config.calculate_bwd_time else None,
         sparsity=sparsity,
@@ -501,15 +600,15 @@ def calculate_tflops(config: ExperimentConfig, results: ExperimentResults) -> fl
     softmax_flops = M * N * 2  # Not counting online softmax overhead
     o_flops = M * D * N * 2
     # Not counting split k overhead
-    total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - results.sparsity)
+    sparsity = results.sparsity if results.sparsity is not None else 0.0
+    total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - sparsity)
     return total_flops / results.fwd_time / 1e6  # in TFLOPs/
 
 
 def get_average_speedups(results: list[Experiment], type: str, backend: str):
     # Calculate speedups
     speedups = [
-        calculate_speedup(r.results["compiled"], r.results[backend], type)
+        calculate_speedup(r.results["flex"], r.results[backend], type) for r in results
-        for r in results
     ]
 
     # Find indices of max and min speedups
@@ -537,7 +636,7 @@ def get_average_speedups(results: list[Experiment], type: str, backend: str):
 def print_results(results: list[Experiment], save_path: Optional[str] = None):
     table_data = defaultdict(list)
     for experiment in results:
-        backends = experiment.config.backends + ["compiled"]
+        backends = experiment.config.backends + ["flex"]
         for key, value in experiment.asdict().items():
             if key in backends:
                 if value.fwd_time:
@@ -550,45 +649,43 @@ def print_results(results: list[Experiment], save_path: Optional[str] = None):
     # Calculate speedups
     for backend in results[0].config.backends:
         fwd_speedups = [
-            calculate_speedup(r.results["compiled"], r.results[backend], type="fwd")
+            calculate_speedup(r.results["flex"], r.results[backend], type="fwd")
             for r in results
         ]
-        table_data[f"fwd_{backend}_speedup"] = fwd_speedups
+        table_data[f"fwd_speedup_flex_over_{backend}"] = fwd_speedups
 
     if results[0].config.calculate_bwd_time:
         for backend in results[0].config.backends:
             bwd_speedups = [
-                calculate_speedup(r.results["compiled"], r.results[backend], type="bwd")
+                calculate_speedup(r.results["flex"], r.results[backend], type="bwd")
                 for r in results
             ]
-            table_data[f"bwd_{backend}_speedup"] = bwd_speedups
+            table_data[f"bwd_speedup_flex_over_{backend}"] = bwd_speedups
 
     # Calculate mem + computational throughput
     if results[0].config.cal_bandwidth:
         fwd_bandwidth = [
-            calculate_bandwidth(r.config, r.results["compiled"], type="fwd")
+            calculate_bandwidth(r.config, r.results["flex"], type="fwd")
             for r in results
         ]
         table_data["fwd_mem_bw (TB/s)"] = fwd_bandwidth
-        fwd_tflops = [
+        fwd_tflops = [calculate_tflops(r.config, r.results["flex"]) for r in results]
-            calculate_tflops(r.config, r.results["compiled"]) for r in results
-        ]
         table_data["TFlops/s"] = fwd_tflops
 
     print(tabulate(table_data, headers="keys", tablefmt="github", floatfmt=".3f"))
 
     for backend in results[0].config.backends:
-        if np.isnan(table_data[f"fwd_{backend}_speedup"]).all():
+        if np.isnan(table_data[f"fwd_speedup_flex_over_{backend}"]).all():
             continue
         print("\n")
-        print(f"FWD Speedups vs. {backend}".center(125, "="))
+        print(f"FWD Speedup of Flex over {backend}".center(125, "="))
         print("\n")
         average_data = get_average_speedups(results, type="fwd", backend=backend)
         print(tabulate(average_data, headers="keys", tablefmt="github", floatfmt=".3f"))
 
         if results[0].config.calculate_bwd_time:
             print("\n")
-            print(f"BWD Speedups vs. {backend}".center(125, "="))
+            print(f"BWD Speedup of Flex over {backend}".center(125, "="))
             print("\n")
             average_data = get_average_speedups(results, type="bwd", backend=backend)
             print(
@@ -791,14 +888,14 @@ def get_backend_context(backend: str):
     Returns a context manager for the specified backend.
     Args:
         backend (str): The name of the backend to use.
-                       Valid options are 'fav2', 'cudnn', 'math', 'efficient', 'fav3', 'fakv', 'og-eager'.
+                       Valid options are 'math', 'efficient', 'cudnn', 'fav2', 'fav3', 'fakv', 'og-eager'.
     Returns:
         A context manager for the specified backend.
     Raises:
         ValueError: If an invalid backend is specified.
     """
     backends = {
-        "fav2": nullcontext(),
+        "fav2": sdpa_kernel(SDPBackend.FLASH_ATTENTION),
         "cudnn": sdpa_kernel(SDPBackend.CUDNN_ATTENTION),
         "math": sdpa_kernel(SDPBackend.MATH),
         "efficient": sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION),
@@ -820,15 +917,7 @@ def generate_FA_callable(
 ) -> Callable | None:
     if dtype not in [torch.float16, torch.bfloat16]:
         return None
-    if backend == "fav2":
+    if backend == "fav3":
-        try:
-            from flash_attn import flash_attn_func, flash_attn_varlen_func
-        except ImportError:
-            print(
-                "Flash attention 2 is not installed. Please install it to run fav2 backend. "
-            )
-            raise
-    elif backend == "fav3":
         try:
             from flash_attn.flash_attn_interface import (
                 flash_attn_func,
@@ -1034,6 +1123,7 @@ def generate_experiment_configs(
     kv_cache_size: list[int],
     cal_bandwidth: bool,
     backends: list[str],
+    max_autotune: bool,
 ) -> list[ExperimentConfig]:
     assert not (calculate_bwd and decoding), "Decoding does not support backward"
 
@@ -1077,52 +1167,333 @@ def generate_experiment_configs(
                 calculate_bwd_time=calculate_bwd,
                 cal_bandwidth=cal_bandwidth,
                 backends=backends,
+                max_autotune=max_autotune,
             )
         )
 
     return all_configs
 
 
-def main(args):
+def _output_json_for_dashboard(
+    experiments,
+    output_file,
+    benchmark_name="PyTorch operator microbenchmark",
+):
+    """
+    Write the result into JSON format for PyTorch OSS dashboard.
+    The JSON format is defined at
+    https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
+
+    Args:
+        experiments: List of experiment results
+        output_file: Path to output JSON file
+        benchmark_name: Name of the benchmark
+    """
+    if not experiments:
+        return
+
+    import math
+    import platform
+    from dataclasses import asdict, dataclass
+    from typing import Any, Optional
+
+    # Prepare headers and records for JSON output
+    records = []
+    for experiment in experiments:
+        config = experiment.config
+        results_dict = (
+            experiment.results
+        )  # This is a dict: backend -> ExperimentResults
+
+        # Process each backend result
+        for backend, results in results_dict.items():
+            # Skip backends that were not run (NaN results)
+            if math.isnan(results.fwd_time):
+                continue
+
+            # Extract data from experiment
+            test_name = f"{backend}_{config.attn_type}_"
+            input_config = f"shape: {config.shape}, dtype: {config.dtype}"
+
+            # Determine mode based on backward pass
+            mode = "training" if config.calculate_bwd_time else "inference"
+
+            # Extract dtype
+            dtype = (
+                str(config.dtype).split(".")[1]
+                if "." in str(config.dtype)
+                else str(config.dtype)
+            )
+
+            # Determine device
+            device = "cuda"
+
+            # Get device architecture
+            device_arch = (
+                torch.cuda.get_device_name(0)
+                if device == "cuda"
+                else platform.processor()
+                if device == "cpu"
+                else "unknown"
+            )
+
+            # Create dataclasses for JSON structure
+            @dataclass
+            class BenchmarkInfo:
+                name: str
+                mode: Optional[str]
+                dtype: str
+                extra_info: dict[str, Any]
+
+            @dataclass
+            class ModelInfo:
+                name: str
+                type: str
+                origins: list[str]
+                extra_info: dict[str, Any]
+
+            @dataclass
+            class MetricInfo:
+                name: str
+                unit: str
+                benchmark_values: list[float]
+                target_value: Optional[float]
+
+            @dataclass
+            class BenchmarkRecord:
+                benchmark: BenchmarkInfo
+                model: ModelInfo
+                metric: MetricInfo
+
+            # Benchmark extra info
+            benchmark_extra_info = {
+                "input_config": input_config,
+                "device": device,
+                "arch": device_arch,
+                "operator_name": backend,
+                "attn_type": config.attn_type,
+                "shape": str(config.shape),
+                "max_autotune": config.max_autotune,
+            }
+            # Add record for forward latency
+            record_fwd_latency = BenchmarkRecord(
+                benchmark=BenchmarkInfo(
+                    name=benchmark_name,
+                    mode=mode,
+                    dtype=dtype,
+                    extra_info=benchmark_extra_info,
+                ),
+                model=ModelInfo(
+                    name=test_name + str(config.shape),
+                    type="attention-benchmark",
+                    origins=["pytorch"],
+                    extra_info={
+                        "operator_name": backend,
+                        "attn_type": config.attn_type,
+                    },
+                ),
+                metric=MetricInfo(
+                    name="forward latency",
+                    unit="us",
+                    benchmark_values=[results.fwd_time],
+                    target_value=None,
+                ),
+            )
+            records.append(asdict(record_fwd_latency))
+
+            # Add record for forward memory bandwidth (if available)
+            if config.cal_bandwidth:
+                record_fwd_bandwidth = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="memory bandwidth",
+                        unit="TB/s",
+                        benchmark_values=[calculate_bandwidth(config, results, "fwd")],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_fwd_bandwidth))
+
+            # Add record for forward TFLOPS (if available)
+            if config.cal_bandwidth:
+                record_fwd_tflops = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="tflops",
+                        unit="TFLOPS/s",
+                        benchmark_values=[calculate_tflops(config, results)],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_fwd_tflops))
+
+            # Add record for backward latency (if available and not NaN)
+            if (
+                config.calculate_bwd_time
+                and results.bwd_time is not None
+                and not math.isnan(results.bwd_time)
+            ):
+                record_bwd_latency = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="backward latency",
+                        unit="us",
+                        benchmark_values=[results.bwd_time],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_bwd_latency))
+
+    # Write all records to the output file
+    with open(output_file, "w", encoding="utf-8") as f:
+        json.dump(records, f, indent=2)
+
+
+def main(
+    dynamic: bool = False,
+    calculate_bwd: bool = False,
+    dtype: DtypeString = "bfloat16",
+    b: list[int] | None = None,
+    nh: list[str] | None = None,
+    s: list[int] | None = None,
+    d: list[int] | None = None,
+    mods: list[AttentionType] | None = None,
+    backend: list[Backend] | None = None,
+    max_autotune: bool = False,
+    decoding: bool = False,
+    kv_size: Optional[list[int]] = None,
+    throughput: bool = True,
+    save_path: Optional[str] = None,
+    output_json_for_dashboard: Optional[str] = None,
+    benchmark_name: str = "PyTorch operator microbenchmark",
+) -> None:
+    """Run sweep over sizes and score mods for flex attention.
+
+    Usage Examples:
+        # Use a yml config file
+        python score_mod.py --config basic_config.yaml
+
+        # Use a json config file
+        python score_mod.py --config my_config.json
+
+        # Generate a config template
+        python score_mod.py --print-config json > my_config.json # For a json config
+        python score_mod.py --print-config yaml > my_config.yaml # For a yaml config
+
+        # Override config with CLI args
+        python score_mod.py --config my_config.json -dtype float16 --max-autotune
+
+        # Pure CLI usage
+        python score_mod.py -b 4 8 -s 1024 2048 -mods causal alibi --backend efficient
+
+    Args:
+        dynamic: Runs a dynamic shapes version of compiled flex attention
+        calculate_bwd: Calculate backward pass times
+        dtype: Data type for tensors (bfloat16, float16, float32)
+        b: Batch sizes to benchmark
+        nh: Number of query and key/value heads in format "Hq,Hkv"
+        s: Sequence lengths to benchmark
+        d: Head dimensions to benchmark
+        mods: Score modifications: noop, causal, rel, head_bias, alibi, sliding_window, document_mask, prefix_lm, softcap
+        backend: Backends for attention computation: math, efficient, cudnn, fav2, fav3, fakv, og-eager
+        max_autotune: Turn on max-autotune optimization
+        decoding: Benchmark decoding mode (query sequence length = 1)
+        kv_size: Key/value cache size in MiB (ignores batch size if specified)
+        throughput: Calculate kernel memory bandwidth & computational throughput (always True)
+        save_path: Path to save the results CSV file
+        output_json_for_dashboard: Path to save results in JSON format for PyTorch OSS dashboard
+        benchmark_name: Name of the benchmark for dashboard output
+    """
+    # Convert dtype string to torch dtype (if not already converted)
+    import torch
+
+    if isinstance(dtype, str):
+        dtype = getattr(torch, dtype)
+
+    # Always calculate throughput
+    throughput = True
+    print("Backend: ", backend)
     seed = 123
     np.random.seed(seed)
     torch.manual_seed(seed)
     results = []
-    for config in tqdm(
+    for experiment_count, config in enumerate(
+        tqdm(
             generate_experiment_configs(
-            args.calculate_bwd,
+                calculate_bwd,
-            args.dtype,
+                dtype,
-            args.b,
+                b,
-            args.nh,
+                nh,
-            args.s,
+                s,
-            args.d,
+                d,
-            args.mods,
+                mods,
-            args.decoding,
+                decoding,
-            args.kv_size,
+                kv_size,
-            args.throughput,
+                throughput,
-            args.backend,
+                backend,
+                max_autotune,
             )
+        ),
+        start=1,
     ):
         results.append(
             Experiment(
                 config,
                 run_single_experiment(
                     config,
-                    dynamic=args.dynamic,
+                    dynamic=dynamic,
-                    max_autotune=args.max_autotune,
                 ),
             )
         )
 
-    print_results(results, args.save_path)
+        # Periodic memory cleanup every 50 experiments
+        if experiment_count % 50 == 0:
+            cleanup_memory()
 
+    print_results(results, save_path)
 
-def heads_input_type(s):
+    # Output JSON for dashboard if requested
-    try:
+    if output_json_for_dashboard:
-        hq, hkv = map(int, s.split(","))
+        _output_json_for_dashboard(results, output_json_for_dashboard, benchmark_name)
-        return hq, hkv
-    except Exception as e:
-        raise argparse.ArgumentTypeError("Heads must be Hq,Hkv") from e
 
 
 if __name__ == "__main__":
@@ -1130,6 +1501,12 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Run sweep over sizes and score mods for flex attention"
     )
+    parser.add_argument(
+        "--config",
+        type=str,
+        help="Path to JSON config file. CLI args override config file values.",
+        default=None,
+    )
     parser.add_argument(
         "--dynamic",
         action="store_true",
@@ -1199,8 +1576,49 @@ Ignores -b batch size and calculate batch size from kv size instead when specifi
         default=["efficient"],
         help="Backend to use for attention computation",
     )
+    parser.add_argument(
+        "--output-json-for-dashboard",
+        type=str,
+        help="Path to save results in JSON format for PyTorch OSS dashboard",
+        default=None,
+    )
+    parser.add_argument(
+        "--benchmark-name",
+        type=str,
+        help="Name of the benchmark for dashboard output",
+        default="PyTorch operator microbenchmark",
+    )
+    parser.add_argument(
+        "--print-config",
+        type=str,
+        choices=["json", "yaml"],
+        help="Print a default config template in JSON or YAML format and exit",
+        default=None,
+    )
     # Parse arguments
     args = parser.parse_args()
+
+    # Handle --print-config
+    if args.print_config:
+        print_default_config(args.print_config)
+        sys.exit(0)
+
+    # Load and merge config if provided
+    if args.config:
+        config = load_config_file(args.config)
+
+        # Merge config with CLI args (CLI args take precedence)
+        json_args = argparse.Namespace()
+        json_args.__dict__ = config
+        args = parser.parse_args(namespace=json_args)
+
+    # Convert dtype string to torch dtype (only if it's still a string)
+    if isinstance(args.dtype, str):
         args.dtype = getattr(torch, args.dtype)
 
-    main(args)
+    # Remove config and print_config from args before passing to main
+    args_dict = vars(args)
+    args_dict.pop("config", None)
+    args_dict.pop("print_config", None)
+
+    main(**args_dict)

build_variables.bzl

@@ -482,6 +482,7 @@ inductor_core_resources = [
     "torch/csrc/inductor/aoti_torch/oss_proxy_executor.cpp",
     "torch/csrc/inductor/inductor_ops.cpp",
     "torch/csrc/jit/serialization/pickle.cpp",
+    "torch/csrc/shim_common.cpp",
 ]
 
 libtorch_core_sources = sorted(

c10/core/SymInt.h

@@ -556,3 +556,26 @@ inline SymBool sym_ge(const SymInt& a, const SymInt& b) {
 }
 
 } // namespace c10
+
+#include <limits>
+
+namespace std {
+
+template <>
+class numeric_limits<c10::SymInt> {
+ public:
+  static constexpr bool is_specialized = true;
+
+  static constexpr int64_t max() noexcept {
+    return std::numeric_limits<int64_t>::max();
+  }
+
+  static constexpr int64_t min() noexcept {
+    return std::numeric_limits<int64_t>::min();
+  }
+
+  static constexpr bool is_signed = true;
+  static constexpr bool is_integer = true;
+};
+
+} // namespace std

c10/cuda/CUDAStream.cpp

@@ -15,7 +15,6 @@ namespace c10::cuda {
 namespace {
 
 // Global stream state and constants
-c10::once_flag init_flag;
 DeviceIndex num_gpus = -1;
 constexpr int kStreamsPerPoolBits = 5;
 constexpr int kStreamsPerPool = 1 << kStreamsPerPoolBits;
@@ -226,7 +225,10 @@ void initDeviceStreamState(DeviceIndex device_index) {
 // Init front-end to ensure initialization only occurs once
 void initCUDAStreamsOnce() {
   // Inits default streams (once, globally)
-  c10::call_once(init_flag, initGlobalStreamState);
+  auto static init_flag [[maybe_unused]] = [] {
+    initGlobalStreamState();
+    return true;
+  }();
 
   if (current_streams) {
     return;