Update on "[ATen] Fix CUDA reduction warp shuffle order"

Typical warp shuffle reduction has the following pattern:
<img width="1138" height="501" alt="image" src="https://github.com/user-attachments/assets/3bd176dc-0ad2-4df6-90c7-06e467337166" />

which is exhibited in Triton generated by torch.compile:
<img width="663" height="403" alt="image" src="https://github.com/user-attachments/assets/7f9f36cd-b9eb-44c1-879e-b469668a2ea8" />

Switch the warp shuffle order to make bitwise equivalence between the 2 easier.
PTX difference between old and new, we see a few extra instructions: https://www.diffchecker.com/h6ly3INC/

Comparing the performance on different reduction operations, we see minimal differences. New represents the changes in this PR, old represents the past warp shuffle order:
```
Tensor Shape              Operation            New all dims (ms)       New dim=0 (ms)      New dim=1 (ms)     Old all dims (ms)    Old dim=0 (ms)      Old dim=1 (ms)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1024, 1024)              mean                 0.015817             0.016259             0.013642             0.015990             0.016258             0.013631             
(1024, 1024)              sum                  0.015917             0.015906             0.013359             0.015707             0.016266             0.013226             
(1024, 1024)              min                  0.016021             0.024625             0.015631             0.015761             0.024485             0.015317             
(1024, 1024)              max                  0.016349             0.024971             0.015972             0.015771             0.025001             0.015314             
(1024, 1024)              argmin               0.018070             0.024448             0.015578             0.018135             0.025370             0.015322             
(1024, 1024)              argmax               0.018427             0.024859             0.015932             0.018164             0.024452             0.015639             
(1024, 1024)              var                  0.020078             0.026413             0.020295             0.020199             0.026381             0.020214             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(2048, 2048)              mean                 0.023826             0.023726             0.022273             0.023236             0.023776             0.022248             
(2048, 2048)              sum                  0.023840             0.023355             0.021974             0.023294             0.023354             0.021884             
(2048, 2048)              min                  0.024519             0.041263             0.024620             0.023292             0.041491             0.024358             
(2048, 2048)              max                  0.024509             0.041670             0.024277             0.023334             0.041231             0.024395             
(2048, 2048)              argmin               0.026125             0.041282             0.024567             0.026772             0.041773             0.024296             
(2048, 2048)              argmax               0.026117             0.041487             0.024572             0.026412             0.041477             0.024273             
(2048, 2048)              var                  0.026603             0.048581             0.031308             0.027587             0.048603             0.030860             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(4096, 4096)              mean                 0.053927             0.057070             0.054073             0.053028             0.057544             0.053935             
(4096, 4096)              sum                  0.053604             0.057410             0.054451             0.053076             0.057033             0.054266             
(4096, 4096)              min                  0.054293             0.109122             0.058363             0.053821             0.108689             0.058382             
(4096, 4096)              max                  0.054258             0.108035             0.058703             0.053492             0.110552             0.058376             
(4096, 4096)              argmin               0.056805             0.111167             0.058301             0.056836             0.112325             0.058292             
(4096, 4096)              argmax               0.056488             0.110958             0.058636             0.056844             0.111000             0.057928             
(4096, 4096)              var                  0.058936             0.141755             0.068693             0.059735             0.141284             0.068500             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(8192, 8192)              mean                 0.145552             0.148082             0.138647             0.145364             0.147818             0.138207             
(8192, 8192)              sum                  0.145985             0.147900             0.138714             0.145755             0.148031             0.138616             
(8192, 8192)              min                  0.146566             0.205359             0.192739             0.145611             0.205237             0.182335             
(8192, 8192)              max                  0.146526             0.204844             0.193050             0.146073             0.205457             0.182697             
(8192, 8192)              argmin               0.150190             0.206605             0.192543             0.150654             0.206847             0.182007             
(8192, 8192)              argmax               0.150481             0.206368             0.192535             0.150845             0.206430             0.182022             
(8192, 8192)              var                  0.150884             0.184546             0.203900             0.151594             0.184172             0.197983             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1, 1024, 128)            mean                 0.014293             0.008119             0.014533             0.013861             0.008022             0.014449             
(1, 1024, 128)            sum                  0.014039             0.007877             0.014111             0.014219             0.008227             0.014045             
(1, 1024, 128)            min                  0.014159             0.011354             0.023493             0.014271             0.010862             0.023644             
(1, 1024, 128)            max                  0.014154             0.011027             0.023368             0.014259             0.011234             0.023692             
(1, 1024, 128)            argmin               0.016403             0.005677             0.023328             0.016273             0.005683             0.024073             
(1, 1024, 128)            argmax               0.016734             0.005675             0.023437             0.016580             0.005318             0.023331             
(1, 1024, 128)            var                  0.018338             0.009549             0.025538             0.018528             0.009391             0.024777             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(5, 1024, 128)            mean                 0.014873             0.010131             0.015546             0.015123             0.010131             0.015481             
(5, 1024, 128)            sum                  0.015334             0.009673             0.015824             0.014736             0.009671             0.015438             
(5, 1024, 128)            min                  0.015047             0.013252             0.024573             0.014803             0.013163             0.024551             
(5, 1024, 128)            max                  0.015050             0.013339             0.024197             0.014810             0.013525             0.024230             
(5, 1024, 128)            argmin               0.017341             0.012737             0.024306             0.017471             0.012379             0.024991             
(5, 1024, 128)            argmax               0.017345             0.012411             0.024421             0.017422             0.012471             0.024237             
(5, 1024, 128)            var                  0.019973             0.011453             0.026188             0.020050             0.011438             0.026282             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(10, 1024, 128)           mean                 0.016976             0.011575             0.016831             0.016722             0.011927             0.017173             
(10, 1024, 128)           sum                  0.017039             0.011841             0.017159             0.016385             0.011860             0.016753             
(10, 1024, 128)           min                  0.017036             0.015331             0.026770             0.016944             0.015205             0.027166             
(10, 1024, 128)           max                  0.017369             0.015348             0.027077             0.016531             0.015716             0.026819             
(10, 1024, 128)           argmin               0.019203             0.014447             0.026813             0.018994             0.014497             0.027313             
(10, 1024, 128)           argmax               0.019563             0.014795             0.027140             0.019460             0.014912             0.026733             
(10, 1024, 128)           var                  0.020529             0.014316             0.030405             0.020719             0.013960             0.029964             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(100, 1024, 128)          mean                 0.045046             0.039168             0.046082             0.044839             0.039217             0.045782             
(100, 1024, 128)          sum                  0.045094             0.039150             0.045777             0.044496             0.039542             0.046083             
(100, 1024, 128)          min                  0.045768             0.054466             0.076244             0.044915             0.053943             0.076599             
(100, 1024, 128)          max                  0.045748             0.054459             0.076188             0.044931             0.053949             0.076856             
(100, 1024, 128)          argmin               0.048275             0.054046             0.076647             0.048694             0.054105             0.077004             
(100, 1024, 128)          argmax               0.048267             0.054395             0.077401             0.048691             0.054131             0.076751             
(100, 1024, 128)          var                  0.049710             0.043254             0.083077             0.050971             0.043251             0.082378             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1000, 1000, 100)         mean                 0.202312             0.196723             0.197765             0.201774             0.196641             0.197459             
(1000, 1000, 100)         sum                  0.202651             0.196682             0.197736             0.202175             0.196313             0.197523             
(1000, 1000, 100)         min                  0.203022             0.264762             0.269200             0.202729             0.264129             0.268694             
(1000, 1000, 100)         max                  0.202864             0.264396             0.269388             0.202486             0.263896             0.268720             
(1000, 1000, 100)         argmin               0.226727             0.263781             0.268651             0.226597             0.264676             0.268983             
(1000, 1000, 100)         argmax               0.226412             0.264469             0.269090             0.226570             0.264595             0.269178             
(1000, 1000, 100)         var                  0.243223             0.204079             0.216096             0.241942             0.204079             0.215925             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(10000, 100)              mean                 0.016193             0.020277             0.014316             0.016152             0.020324             0.013712             
(10000, 100)              sum                  0.016289             0.020237             0.014034             0.016168             0.020265             0.013708             
(10000, 100)              min                  0.016046             0.030872             0.019609             0.016208             0.030867             0.018627             
(10000, 100)              max                  0.016369             0.030835             0.019257             0.016218             0.030861             0.018209             
(10000, 100)              argmin               0.017957             0.031171             0.019517             0.018050             0.031556             0.018077             
(10000, 100)              argmax               0.017961             0.031658             0.019521             0.018060             0.031564             0.018087             
(10000, 100)              var                  0.020393             0.035652             0.019339             0.020144             0.035987             0.019171             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(100000, 10)              mean                 0.015718             0.016576             0.016555             0.015999             0.016246             0.014869             
(100000, 10)              sum                  0.015833             0.016247             0.016572             0.016007             0.016627             0.014872             
(100000, 10)              min                  0.015888             0.020510             0.023920             0.015671             0.020821             0.021417             
(100000, 10)              max                  0.015889             0.020479             0.023918             0.016077             0.020386             0.021421             
(100000, 10)              argmin               0.018233             0.020863             0.023647             0.017574             0.020864             0.021103             
(100000, 10)              argmax               0.017896             0.020527             0.023296             0.017569             0.020447             0.021098             
(100000, 10)              var                  0.020005             0.024198             0.024372             0.020075             0.024167             0.022415             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1023, 1023, 1023)        mean                 1.874816             1.963506             1.903909             1.873279             1.963859             1.903230             
(1023, 1023, 1023)        sum                  1.875030             1.965716             1.902458             1.873566             1.960730             1.901642             
(1023, 1023, 1023)        min                  1.878563             2.473455             2.179092             1.875174             2.482086             2.183027             
(1023, 1023, 1023)        max                  1.879128             2.474803             2.178895             1.874831             2.482253             2.183884             
(1023, 1023, 1023)        argmin               1.921800             2.476629             2.174831             1.923987             2.472641             2.170453             
(1023, 1023, 1023)        argmax               1.922605             2.476688             2.177927             1.923366             2.472808             2.172979             
(1023, 1023, 1023)        var                  1.972606             3.088695             2.758797             1.978679             3.095658             2.762243             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1023, 1023, 255)         mean                 0.489984             0.500954             0.492957             0.489891             0.500654             0.491971             
(1023, 1023, 255)         sum                  0.490228             0.500764             0.492289             0.489624             0.501089             0.492824             
(1023, 1023, 255)         min                  0.491457             0.563560             0.553334             0.490355             0.564709             0.554754             
(1023, 1023, 255)         max                  0.491396             0.563628             0.553345             0.490017             0.565004             0.554947             
(1023, 1023, 255)         argmin               0.503666             0.561512             0.551831             0.503845             0.560972             0.551017             
(1023, 1023, 255)         argmax               0.503602             0.561185             0.551407             0.504328             0.561267             0.551448             
(1023, 1023, 255)         var                  0.510844             0.709452             0.701630             0.512693             0.710365             0.701965             
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1023, 1023, 377)         mean                 0.707439             0.727646             0.712019             0.706769             0.727101             0.711632             
(1023, 1023, 377)         sum                  0.707780             0.727453             0.711554             0.706807             0.726656             0.711729             
(1023, 1023, 377)         min                  0.709423             0.819809             0.794379             0.707847             0.822086             0.796664             
(1023, 1023, 377)         max                  0.709297             0.819780             0.794308             0.707566             0.821913             0.796690             
(1023, 1023, 377)         argmin               0.725028             0.817088             0.791695             0.726039             0.816445             0.790828             
(1023, 1023, 377)         argmax               0.725301             0.817011             0.791420             0.726040             0.816917             0.791143             
(1023, 1023, 377)         var                  0.740859             1.034165             1.006712             0.743413             1.035506             1.007638             
```

[ghstack-poisoned]

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -8,6 +8,8 @@ if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0"
 elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
 elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi

.ci/docker/build.sh

@@ -113,6 +113,7 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
+    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -181,7 +182,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
+    PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950;gfx1100"
     if [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
     fi
@@ -344,7 +345,7 @@ docker build \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
-       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx1100}" \
+       --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}" \
        --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
        --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \
@@ -361,6 +362,7 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
+       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+deb42f2a8e48f5032b4a98ee781a15fa87a157cf

.ci/docker/ci_commit_pins/triton.txt

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

.ci/docker/common/install_cpython.sh

@@ -83,10 +83,6 @@ function build_cpython {
         py_suffix=${py_ver::-1}
         py_folder=$py_suffix
     fi
-    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
-    if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0rc2"
-    fi
     wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
     do_cpython_build $py_ver Python-$py_suffix
 

.ci/docker/common/install_mingw.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+set -ex
+
+# Install MinGW-w64 for Windows cross-compilation
+apt-get update
+apt-get install -y g++-mingw-w64-x86-64-posix
+
+echo "MinGW-w64 installed successfully"
+x86_64-w64-mingw32-g++ --version

.ci/docker/common/install_onnx.sh

@@ -20,7 +20,7 @@ pip_install \
 
 pip_install coloredlogs packaging
 pip_install onnxruntime==1.23.0
-pip_install onnxscript==0.5.3
+pip_install onnxscript==0.5.4
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

.ci/docker/libtorch/build.sh

@@ -39,9 +39,13 @@ case ${DOCKER_TAG_PREFIX} in
         DOCKER_GPU_BUILD_ARG=""
         ;;
     rocm*)
+        # we want the patch version of 7.0 instead
+        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -75,9 +75,13 @@ case ${image} in
         DOCKERFILE_SUFFIX="_cuda_aarch64"
         ;;
     manylinux2_28-builder:rocm*)
+        # we want the patch version of 7.0 instead
+        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

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

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

.ci/docker/ubuntu/Dockerfile

@@ -103,6 +103,11 @@ 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 torchbench.txt
 
+ARG INSTALL_MINGW
+COPY ./common/install_mingw.sh install_mingw.sh
+RUN if [ -n "${INSTALL_MINGW}" ]; then bash ./install_mingw.sh; fi
+RUN rm install_mingw.sh
+
 ARG TRITON
 ARG TRITON_CPU
 

.ci/lumen_cli/cli/lib/common/git_helper.py

@@ -57,8 +57,8 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
         logger.info("Successfully cloned %s", target)
         return r, commit
 
-    except GitCommandError as e:
-        logger.error("Git operation failed: %s", e)
+    except GitCommandError:
+        logger.exception("Git operation failed")
         raise
 
 

.ci/lumen_cli/cli/lib/core/vllm/lib.py

@@ -143,7 +143,7 @@ def sample_vllm_test_library():
                 "pytest -v -s compile/test_decorator.py",
             ],
         },
-        "vllm_languagde_model_test_extended_generation_28_failure_test": {
+        "vllm_language_model_test_extended_generation_28_failure_test": {
             "title": "Language Models Test (Extended Generation) 2.8 release failure",
             "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
             "package_install": [

.ci/lumen_cli/cli/lib/core/vllm/vllm_build.py

@@ -63,7 +63,7 @@ class VllmBuildParameters:
     # DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
     use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
     dockerfile_path: Path = env_path_field(
-        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
+        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile"
     )
 
     # the cleaning script to remove torch dependencies from pip

.ci/manywheel/build_cuda.sh

@@ -187,19 +187,22 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             export USE_CUFILE=0
         else
             DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
                 "/usr/local/cuda/lib64/libcublas.so.12"
                 "/usr/local/cuda/lib64/libcublasLt.so.12"
                 "/usr/local/cuda/lib64/libcudart.so.12"
                 "/usr/local/cuda/lib64/libnvrtc.so.12"
                 "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
             DEPS_SONAME+=(
-                "libnvToolsExt.so.1"
                 "libcublas.so.12"
                 "libcublasLt.so.12"
                 "libcudart.so.12"
                 "libnvrtc.so.12"
                 "libcupti.so.12")
+
+            if [[ $CUDA_VERSION != 12.9* ]]; then
+                DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
+                DEPS_SONAME+=("libnvToolsExt.so.1")
+            fi
         fi
     else
         echo "Using nvidia libs from pypi."

.ci/pytorch/build.sh

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

.ci/pytorch/macos-test.sh

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

.ci/pytorch/test.sh

@@ -337,13 +337,13 @@ test_python() {
 
 test_python_smoke() {
   # Smoke tests for H100/B200
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
 test_python_smoke_b200() {
   # Targeted smoke tests for B200 - staged approach to avoid too many failures
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
+  time python test/run_test.py --include test_matmul_cuda test_scaled_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
@@ -485,6 +485,22 @@ test_inductor_aoti() {
   /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
 }
 
+test_inductor_aoti_cross_compile_for_windows() {
+
+  TEST_REPORTS_DIR=$(pwd)/test/test-reports
+  mkdir -p "$TEST_REPORTS_DIR"
+
+  # Set WINDOWS_CUDA_HOME environment variable
+  WINDOWS_CUDA_HOME="$(pwd)/win-torch-wheel-extracted"
+  export WINDOWS_CUDA_HOME
+
+  echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
+  echo "Contents:"
+  ls -lah "$(pwd)/win-torch-wheel-extracted/lib/x64/" || true
+
+  python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
+}
+
 test_inductor_cpp_wrapper_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -838,7 +854,7 @@ test_dynamo_benchmark() {
       elif [[ "${suite}" == "timm_models" ]]; then
         export TORCHBENCH_ONLY_MODELS="inception_v3"
       elif [[ "${suite}" == "torchbench" ]]; then
-        export TORCHBENCH_ONLY_MODELS="hf_Bert"
+        export TORCHBENCH_ONLY_MODELS="BERT_pytorch"
       fi
     fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
@@ -869,13 +885,13 @@ test_inductor_torchbench_smoketest_perf() {
   mkdir -p "$TEST_REPORTS_DIR"
 
   python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --float16 --training \
-    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only hf_Bert \
+    --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" --only BERT_pytorch \
     --output "$TEST_REPORTS_DIR/inductor_training_smoketest.csv"
   # The threshold value needs to be actively maintained to make this check useful
   python benchmarks/dynamo/check_perf_csv.py -f "$TEST_REPORTS_DIR/inductor_training_smoketest.csv" -t 1.4
 
   # Check memory compression ratio for a few models
-  for test in hf_Albert timm_vision_transformer; do
+  for test in BERT_pytorch yolov3; do
     python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --amp --training \
       --disable-cudagraphs --batch-size-file "$(realpath benchmarks/dynamo/torchbench_models_list.txt)" \
       --only $test --output "$TEST_REPORTS_DIR/inductor_training_smoketest_$test.csv"
@@ -900,7 +916,7 @@ test_inductor_set_cpu_affinity(){
   export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
   export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
 
-  if [[ "${TEST_CONFIG}" != *aarch64* ]]; then
+  if [[ "$(uname -m)" != "aarch64" ]]; then
     # Use Intel OpenMP for x86
     IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
     export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@@ -914,7 +930,7 @@ test_inductor_set_cpu_affinity(){
   cores=$((cpus / thread_per_core))
 
   # Set number of cores to 16 on aarch64 for performance runs
-  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
+  if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
@@ -1615,6 +1631,7 @@ test_operator_benchmark() {
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
   TEST_DIR=$(pwd)
+  ARCH=$(uname -m)
 
   test_inductor_set_cpu_affinity
 
@@ -1629,7 +1646,7 @@ test_operator_benchmark() {
   pip_install pandas
   python check_perf_csv.py \
       --actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
+      --expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
 test_operator_microbenchmark() {
@@ -1666,7 +1683,7 @@ if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
     python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   fi
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
-elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
   test_linux_aarch64
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
@@ -1717,6 +1734,8 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
+elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
+  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))

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

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

.circleci/scripts/binary_populate_env.sh

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

.flake8

@@ -7,16 +7,12 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907,B908,B910
-    # these ignores are from flake8-comprehensions; please fix!
-    C407,
-    # these ignores are from flake8-logging-format; please fix!
-    G100,G101,G200
+    B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
     SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # SIM104 is already covered by pyupgrade ruff

.github/ISSUE_TEMPLATE/ci-sev.md

@@ -8,6 +8,7 @@ assignees: ''
 ---
 
 > NOTE: Remember to label this issue with "`ci: sev`"
+>       If you want autorevert to be disabled, keep the ci: disable-autorevert label
 
  <!-- Add the `merge blocking` label to this PR to prevent PRs from being merged while this issue is open -->
 

.github/ISSUE_TEMPLATE/disable-autorevert.md

@@ -1,7 +1,7 @@
 ---
-name: DISABLE AUTOREVERT
+name: "D❌​\U0001F519​ ISABLE AUTOREVERT"
 about: Disables autorevert when open
-title: "❌​\U0001F519​ [DISABLE AUTOREVERT]"
+title: "[DISABLE AUTOREVERT]"
 labels: 'ci: disable-autorevert'
 assignees: ''
 

.github/actions/build-external-packages/action.yml

@@ -65,7 +65,7 @@ runs:
           cd .ci/lumen_cli
           python3 -m pip install -e .
         )
-        MAX_JOBS="$(nproc --ignore=6)"
+        MAX_JOBS="$(nproc --ignore=10)"
         export MAX_JOBS
 
         # Split the comma-separated list and build each target

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

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

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

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

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

@@ -111,3 +111,16 @@ runs:
         # This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
         # The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
         echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
+
+    - name: configure aws credentials
+      id: aws_creds
+      uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
+      with:
+        role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+        aws-region: us-east-1
+        role-duration-seconds: 18000
+
+    - name: Login to Amazon ECR
+      id: login-ecr
+      continue-on-error: true
+      uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1

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

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

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+69bbe7363897764f9e758d851cd0340147d27f94

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-966da7e46f65d6d49df3e31214470a4fe5cc8e66
+faffd5cf673615583da6517275e361cb3dbc77e6

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-0ad9951c416d33c5da4f7a504fb162cbe62386f5
+e5192819208c4d68194844b7dfafbc00020d0dea

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-2a9138a26ee257fef05310ad3fecf7c55fe80d73
+0fa6e3129e61143224663e1ec67980d12b7ec4eb

.github/ci_configs/vllm/Dockerfile

@@ -1,59 +1,71 @@
-# TODO(elainwy): remove this file after the torch nightly dockerfile is in sync in vllm repo
-# The vLLM Dockerfile is used to construct vLLM image against torch nightly and torch main that can be directly used for testing
-
 ARG CUDA_VERSION=12.8.1
 ARG PYTHON_VERSION=3.12
 
 # BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
 # by default, it uses the torch-nightly-base stage from this docker image
 ARG BUILD_BASE_IMAGE=torch-nightly-base
-
-# FINAL_BASE_IMAGE: used to set up vllm-instaled environment and build flashinfer,
-# by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
 # The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
 ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
 #################### TORCH NIGHTLY BASE IMAGE ####################
-# A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
 
 ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# Install Python and other dependencies
+# Install system dependencies and uv, then create Python virtual environment
 RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
+    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
-# Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
-RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+RUN apt-get install -y gcc-10 g++-10
+RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
+RUN <<EOF
+gcc --version
+EOF
 
-# install uv for faster pip installs
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv==0.8.4
+
+ENV UV_HTTP_TIMEOUT=500
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
+#################### TORCH NIGHTLY  BASE IMAGE ####################
+
+
+#################### BASE BUILD IMAGE ####################
+FROM ${BUILD_BASE_IMAGE} AS base
+USER root
+
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+
+# Only work with PyTorch manylinux builder
+ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
+
+# Install some system dependencies and double check python version
+RUN if command -v apt-get >/dev/null; then \
+        apt-get update -y \
+        && apt-get install -y ccache software-properties-common git wget sudo vim; \
+    else \
+        dnf install -y git wget sudo; \
+    fi \
+    && python3 --version && python3 -m pip --version
+
+# Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     python3 -m pip install uv==0.8.4
 
@@ -62,51 +74,17 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
-#################### TORCH NIGHTLY  BASE IMAGE ####################
-
-
-#################### BASE BUILD IMAGE ####################
-# A base image for building vLLM with torch nightly or torch wheels
-# prepare basic build environment
-FROM ${BUILD_BASE_IMAGE} AS base
-USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
-
-# Install uv for faster pip installs if not existed
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version >/dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
 WORKDIR /workspace
 
-# install build and runtime dependencies
+# Install build and runtime dependencies
 COPY requirements/common.txt requirements/common.txt
 COPY use_existing_torch.py use_existing_torch.py
 COPY pyproject.toml pyproject.toml
 
-# install build and runtime dependencies without stable torch version
+# Install build and runtime dependencies without stable torch version
 RUN python3 use_existing_torch.py
 
-# default mount file as placeholder, this just avoid the mount error
+# Default mount file as placeholder, this just avoid the mount error
 # change to a different vllm folder if this does not exist anymore
 ARG TORCH_WHEELS_PATH="./requirements"
 ARG PINNED_TORCH_VERSION
@@ -138,56 +116,36 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/common.txt
 
-# Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
-
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 
-RUN echo ${TORCH_CUDA_ARCH_LIST}
-RUN echo ${MAX_JOBS}
-RUN pip freeze | grep -E 'ninja'
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+    export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
+    git clone https://github.com/facebookresearch/xformers.git
 
-# Build xformers with cuda and torch nightly/wheel
-# following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
-ENV CCACHE_DIR=/root/.cache/ccache
+    pushd xformers
+    git checkout v0.0.32.post2
+    git submodule update --init --recursive
+    python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
+    popd
 
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    echo 'git clone xformers...' \
-    && git clone https://github.com/facebookresearch/xformers.git --recursive \
-    && cd xformers \
-    && git checkout ${XFORMERS_COMMIT} \
-    && git submodule update --init --recursive \
-    && echo 'finish git clone xformers...' \
-    && rm -rf build \
-    && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
-    && cd .. \
-    && rm -rf xformers
+    rm -rf xformers
+BASH
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl --verbose
+    uv pip install --system xformers-dist/*.whl
 
-# Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
-# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-
 RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
-
 #################### BASE BUILD IMAGE ####################
 
 
 #################### WHEEL BUILD IMAGE ####################
-# Image used to build vllm wheel
 FROM base AS build
 ARG TARGETPLATFORM
 
 COPY . .
-
 RUN python3 use_existing_torch.py
 
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -197,20 +155,17 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
 
-# Max jobs used by Ninja to build extensions
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=4
+ARG nvcc_threads=8
 ENV NVCC_THREADS=$nvcc_threads
-ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 
 ARG USE_SCCACHE
 ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 
-# if USE_SCCACHE is set, use sccache to speed up compilation
+# Use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
@@ -235,6 +190,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && sccache --show-stats; \
     fi
 
+ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
+
 ARG vllm_target_device="cuda"
 ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
@@ -248,17 +206,10 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
-
-RUN echo "[INFO] Listing current directory:" && \
-    ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
-    cat torch_build_versions.txt
-
 #################### WHEEL BUILD IMAGE ####################
 
 
 ################### VLLM INSTALLED IMAGE ####################
-# Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
 
@@ -266,7 +217,7 @@ ARG CUDA_VERSION
 ARG PYTHON_VERSION
 ARG GET_PIP_URL
 
-# TODO (huydhn): Only work with PyTorch manylinux builder
+# Only work with PyTorch manylinux builder
 ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
 
 # prepare for environment starts
@@ -275,20 +226,19 @@ WORKDIR /workspace
 # Install Python and other dependencies
 RUN if command -v apt-get >/dev/null; then \
         apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
+        && apt-get install -y ccache software-properties-common git sudo vim python3-pip; \
     else \
-        dnf install -y git curl wget sudo; \
+        dnf install -y git wget sudo; \
     fi \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
+    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -s /opt/venv/bin/pip /usr/bin/pip \
     && python3 --version && python3 -m pip --version
 
-# Get the torch versions, and whls used in previous stagtes for consistency
+# Get the torch versions, and whls used in previous stage
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
@@ -297,33 +247,29 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
-# Install build and runtime dependencies, this is needed for flashinfer install
-COPY requirements/build.txt requirements/build.txt
-COPY use_existing_torch.py use_existing_torch.py
-RUN python3 use_existing_torch.py
-RUN cat requirements/build.txt
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version > /dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
+    python3 -m pip install uv==0.8.4
 
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
+# Install build and runtime dependencies, this is needed for flashinfer install
+COPY requirements/build.txt requirements/build.txt
+COPY use_existing_torch.py use_existing_torch.py
+RUN python3 use_existing_torch.py
+RUN cat requirements/build.txt
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt
 
-
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
-# Install torch, torchaudio and torchvision
-# if TORCH_WHEELS_PATH is default "./requirements", it will pull the nightly versions using pip using torch_build_versions.txt
-# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
+# Install torch, torchaudio and torchvision. If TORCH_WHEELS_PATH is default
+# to ./requirements, it will pull the nightly versions using pip. Otherwise,
+# it will use the local wheels from TORCH_WHEELS_PATH
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
@@ -344,18 +290,14 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install xformers wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
-# Build flashinfer from source.
+
+# Build FlashInfer from source
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
-# install package for build flashinfer
-# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
-
-RUN pip freeze | grep -E 'setuptools|packaging|build'
-
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-# Build flashinfer for torch nightly from source around 10 mins
+
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
 ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -367,7 +309,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     && cd .. \
     && rm -rf flashinfer
 
-# install flashinfer python
+# Install FlashInfer
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system wheels/flashinfer/*.whl --verbose
 
@@ -377,49 +319,6 @@ RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm
 ################### VLLM INSTALLED IMAGE ####################
 
 
-#################### UNITTEST IMAGE #############################
-FROM vllm-base as test
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
-
-COPY tests/ tests/
-COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-# Install build and runtime dependencies without stable torch version
-COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
-
-RUN python3 use_existing_torch.py
-
-# install packages
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
-# enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system hf_transfer
-ENV HF_HUB_ENABLE_HF_TRANSFER 1
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -e tests/vllm_test_utils
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/nightly_torch_test.txt
-
-# Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-
-# Logging to confirm all the packages are installed
-RUN pip freeze
-
-#################### UNITTEST IMAGE #############################
-
 #################### EXPORT STAGE ####################
 FROM scratch as export-wheels
 

.github/labeler.yml

@@ -133,3 +133,32 @@
 
 "ciflow/vllm":
 - .github/ci_commit_pins/vllm.txt
+
+"ciflow/b200":
+- test/test_matmul_cuda.py
+- test/test_scaled_matmul_cuda.py
+- test/inductor/test_fp8.py
+- aten/src/ATen/native/cuda/Blas.cpp
+- torch/**/*cublas*
+- torch/_inductor/kernel/mm.py
+- test/inductor/test_max_autotune.py
+- third_party/fbgemm
+
+"ciflow/h100":
+- test/test_matmul_cuda.py
+- test/test_scaled_matmul_cuda.py
+- test/inductor/test_fp8.py
+- aten/src/ATen/native/cuda/Blas.cpp
+- torch/**/*cublas*
+- torch/_inductor/kernel/mm.py
+- test/inductor/test_max_autotune.py
+- third_party/fbgemm
+
+"ciflow/rocm":
+- test/test_matmul_cuda.py
+- test/test_scaled_matmul_cuda.py
+- test/inductor/test_fp8.py
+- aten/src/ATen/native/cuda/Blas.cpp
+- torch/_inductor/kernel/mm.py
+- test/inductor/test_max_autotune.py
+- third_party/fbgemm

.github/pytorch-probot.yml

@@ -3,6 +3,7 @@ ciflow_tracking_issue: 64124
 ciflow_push_tags:
 - ciflow/b200
 - ciflow/b200-symm-mem
+- ciflow/b200-distributed
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
@@ -15,7 +16,8 @@ ciflow_push_tags:
 - ciflow/inductor-micro-benchmark
 - ciflow/inductor-micro-benchmark-cpu-x86
 - ciflow/inductor-perf-compare
-- ciflow/inductor-perf-test-nightly-rocm
+- ciflow/inductor-perf-test-nightly-rocm-mi300
+- ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
@@ -30,6 +32,7 @@ ciflow_push_tags:
 - ciflow/riscv64
 - ciflow/rocm
 - ciflow/rocm-mi300
+- ciflow/rocm-mi355
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.github/scripts/filter_test_configs.py

@@ -512,6 +512,8 @@ def perform_misc_tasks(
         "keep-going",
         branch == MAIN_BRANCH
         or bool(tag and re.match(r"^trunk/[a-f0-9]{40}$", tag))
+        # Pattern for tags created via manual run on HUD
+        or bool(tag and re.match(r"^ciflow/[^/]+/[a-f0-9]{40}$", tag))
         or check_for_setting(labels, pr_body, "keep-going"),
     )
     set_output(

.github/scripts/generate_binary_build_matrix.py

@@ -16,16 +16,18 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
+    "12.9": "12.9.1",
     "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
+    "12.9": "9",
     "13.0": "9",
 }
 
@@ -38,7 +40,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@@ -76,6 +78,23 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
     ),
+    "12.9": (
+        "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'"
+    ),
     "13.0": (
         "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
         "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
@@ -222,7 +241,11 @@ def generate_libtorch_matrix(
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
         elif os == "windows":
-            arches += CUDA_ARCHES
+            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
+            # in 2.10
+            windows_cuda_arches = CUDA_ARCHES.copy()
+            windows_cuda_arches.remove("12.9")
+            arches += windows_cuda_arches
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -286,7 +309,11 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            arches += CUDA_ARCHES + XPU_ARCHES
+            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
+            # in 2.10
+            windows_cuda_arches = CUDA_ARCHES.copy()
+            windows_cuda_arches.remove("12.9")
+            arches += windows_cuda_arches + XPU_ARCHES
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -322,7 +349,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["13.0", "12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -386,5 +413,6 @@ def generate_wheels_matrix(
 
 
 validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/trymerge.py

@@ -1092,7 +1092,7 @@ class GitHubPR:
         editor = node["editor"]
         return GitHubComment(
             body_text=node["bodyText"],
-            created_at=node["createdAt"] if "createdAt" in node else "",
+            created_at=node.get("createdAt", ""),
             author_login=node["author"]["login"],
             author_url=node["author"].get("url", None),
             author_association=node["authorAssociation"],
@@ -2042,10 +2042,6 @@ def validate_revert(
             f"[{', '.join(allowed_reverters)}], but instead is {author_association}."
         )
 
-    # Raises exception if matching rule is not found, but ignores all status checks
-    find_matching_merge_rule(
-        pr, repo, skip_mandatory_checks=True, skip_internal_checks=True
-    )
     commit_sha = get_pr_commit_sha(repo, pr)
     return (author_login, commit_sha)
 

.github/templates/linux_binary_build_workflow.yml.j2

@@ -177,6 +177,9 @@ jobs:
     runs-on: linux.rocm.gpu.mi250
     timeout-minutes: !{{ common.timeout_minutes }}
     !{{ upload.binary_env(config) }}
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

.github/templates/macos_binary_build_workflow.yml.j2

@@ -26,9 +26,8 @@ name: !{{ build_environment }}
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
+          python-version: "!{{ py_ver.strip('t') + ('.4' if '3.14' not in py_ver else '.0') }}"
           freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
 {%- endmacro %}
 

.github/workflows/_docs.yml

@@ -72,7 +72,7 @@ jobs:
             # Let's try to figure out how this can be improved
             timeout-minutes: 360
           - docs_type: python
-            runner: ${{ inputs.runner_prefix }}linux.2xlarge
+            runner: ${{ inputs.runner_prefix }}linux.c7i.2xlarge
             # It takes less than 30m to finish python docs unless there are issues
             timeout-minutes: 30
     # Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)

.github/workflows/_linux-build.yml

@@ -37,7 +37,7 @@ on:
       runner:
         required: false
         type: string
-        default: "linux.2xlarge"
+        default: "linux.c7i.2xlarge"
         description: |
           Label of the runner this job should run on.
       test-matrix:

.github/workflows/_linux-test.yml

@@ -224,6 +224,46 @@ jobs:
         continue-on-error: true
         uses: ./.github/actions/download-td-artifacts
 
+      - name: Download Windows torch wheel for cross-compilation
+        if: matrix.win_torch_wheel_artifact != ''
+        uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
+        with:
+          name: ${{ matrix.win_torch_wheel_artifact }}
+          path: win-torch-wheel
+
+      - name: Extract Windows wheel and setup CUDA libraries
+        if: matrix.win_torch_wheel_artifact != ''
+        shell: bash
+        run: |
+          set -x
+
+          # Find the wheel file
+          WHEEL_FILE=$(find win-torch-wheel -name "*.whl" -type f | head -n 1)
+          if [ -z "$WHEEL_FILE" ]; then
+            echo "Error: No wheel file found in win-torch-wheel directory"
+            exit 1
+          fi
+          echo "Found wheel file: $WHEEL_FILE"
+
+          # Unzip the wheel file
+          unzip -q "$WHEEL_FILE" -d win-torch-wheel-extracted
+          echo "Extracted wheel contents"
+
+          # Setup CUDA libraries (cuda.lib and cudart.lib) directory
+          mkdir -p win-torch-wheel-extracted/lib/x64
+          if [ -f "win-torch-wheel/cuda.lib" ]; then
+            mv win-torch-wheel/cuda.lib win-torch-wheel-extracted/lib/x64/
+            echo "Moved cuda.lib to win-torch-wheel-extracted/lib/x64/"
+          fi
+          if [ -f "win-torch-wheel/cudart.lib" ]; then
+            mv win-torch-wheel/cudart.lib win-torch-wheel-extracted/lib/x64/
+            echo "Moved cudart.lib to win-torch-wheel-extracted/lib/x64/"
+          fi
+
+          # Verify CUDA libraries are present
+          echo "CUDA libraries:"
+          ls -la win-torch-wheel-extracted/lib/x64/ || echo "No CUDA libraries found"
+
       - name: Parse ref
         id: parse-ref
         run: .github/scripts/parse_ref.py
@@ -389,8 +429,6 @@ jobs:
             "${DOCKER_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
-          # Propagate download.pytorch.org IP to container
-          grep download.pytorch.org /etc/hosts | docker exec -i "${container_name}" sudo bash -c "/bin/cat >> /etc/hosts"
           echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"
 
           if [[ ${BUILD_ENVIRONMENT} == *"s390x"* ]]; then

.github/workflows/_rocm-test.yml

@@ -102,19 +102,6 @@ jobs:
             exit 1
           fi
 
-      - name: configure aws credentials
-        id: aws_creds
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-          aws-region: us-east-1
-          role-duration-seconds: 18000
-
-      - name: Login to Amazon ECR
-        id: login-ecr
-        continue-on-error: true
-        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
-
       - name: Calculate docker image
         id: calculate-docker-image
         uses: pytorch/test-infra/.github/actions/calculate-docker-image@main

.github/workflows/_win-build.yml

@@ -168,6 +168,31 @@ jobs:
         run: |
           .ci/pytorch/win-build.sh
 
+      # Collect Windows torch libs and CUDA libs for cross-compilation
+      - name: Collect Windows CUDA libs for cross-compilation
+        if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu'
+        shell: bash
+        run: |
+          set -ex
+
+          # Create directory structure if does not exist
+          mkdir -p /c/${{ github.run_id }}/build-results
+
+          # Copy CUDA libs
+          CUDA_PATH="/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${{ inputs.cuda-version }}"
+
+          if [ -f "${CUDA_PATH}/lib/x64/cuda.lib" ]; then
+            cp "${CUDA_PATH}/lib/x64/cuda.lib" /c/${{ github.run_id }}/build-results/
+          fi
+
+          if [ -f "${CUDA_PATH}/lib/x64/cudart.lib" ]; then
+            cp "${CUDA_PATH}/lib/x64/cudart.lib" /c/${{ github.run_id }}/build-results/
+          fi
+
+          # List collected files
+          echo "Collected CUDA libs:"
+          ls -lah /c/${{ github.run_id }}/build-results/*.lib
+
       # Upload to github so that people can click and download artifacts
       - name: Upload artifacts to s3
         if: steps.build.outcome != 'skipped'

.github/workflows/b200-distributed.yml

@@ -0,0 +1,62 @@
+name: CI for distributed tests on B200
+
+on:
+  pull_request:
+    paths:
+      - .github/workflows/b200-distributed.yml
+  workflow_dispatch:
+  push:
+    tags:
+      - ciflow/b200-distributed/*
+  schedule:
+    - cron: 46 8 * * *  # about 1:46am PDT
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+
+  get-label-type:
+    if: github.repository_owner == 'pytorch'
+    name: get-label-type
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    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: ${{ github.ref_type }}
+
+  linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200:
+    name: linux-jammy-cuda12.8-py3.10-gcc11-build-distributed-b200
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.12xlarge.memory
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
+      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
+      cuda-arch-list: '10.0'
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 2, runner: "linux.dgx.b200.8" },
+          { config: "distributed", shard: 2, num_shards: 2, runner: "linux.dgx.b200.8" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-cuda12_8-py3_10-gcc11-test-distributed-b200:
+    name: linux-jammy-cuda12.8-py3.10-gcc11-test-b200
+    uses: ./.github/workflows/_linux-test.yml
+    needs:
+      - linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200
+    with:
+      timeout-minutes: 1200
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.test-matrix }}
+      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+    secrets: inherit

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

@@ -46,10 +46,12 @@ jobs:
       fail-fast: false
       matrix:
         include: [
-          { name: "manylinux2_28-builder",          tag: "cuda13.0",         runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "cuda13.0",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.8",          runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "cuda12.9",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.6",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda13.0",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { 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" },

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

@@ -27,9 +27,8 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
         platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
+        device: [ 'cu128', 'cu129', 'cu130' ]
         include:
           - platform: manylinux_2_28_x86_64
             device: cu128
@@ -39,6 +38,10 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
             runner: linux.12xlarge.memory
+          - platform: manylinux_2_28_x86_64
+            device: cu130
+            manylinux-image: 'pytorch/manylinux2_28-builder:cuda13.0'
+            runner: linux.12xlarge.memory
           - platform: manylinux_2_28_aarch64
             device: cu128
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
@@ -47,6 +50,11 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
             runner: linux.arm64.r7g.12xlarge.memory
+        exclude:
+          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
+          # xformers is update to support 13.0
+          - platform: manylinux_2_28_aarch64
+            device: cu130
     name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
@@ -169,7 +177,12 @@ jobs:
       fail-fast: false
       matrix:
         platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
+        device: [ 'cu128', 'cu129', 'cu130' ]
+        exclude:
+          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
+          # xformers is update to support 13.0
+          - platform: manylinux_2_28_aarch64
+            device: cu130
     env:
       PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}

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

@@ -204,6 +204,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -407,6 +453,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -610,6 +702,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -813,6 +951,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1016,6 +1200,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1219,6 +1449,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1422,6 +1698,52 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda-aarch64-12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.arm64.r7g.12xlarge.memory
+      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'
+      timeout-minutes: 420
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda-aarch64-12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda-aarch64-12_9-build
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9-aarch64"
+      GPU_ARCH_TYPE: cuda-aarch64
+      DOCKER_IMAGE: manylinuxaarch64-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda-aarch64-12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda-aarch64-13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml

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

@@ -248,6 +248,74 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-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: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -358,6 +426,9 @@ jobs:
       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
@@ -473,6 +544,9 @@ jobs:
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

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

@@ -241,6 +241,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_10-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_10-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_10-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_10-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.10"
+      build_name: manywheel-py3_10-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -347,6 +413,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.10"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -459,6 +528,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.10"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -835,6 +907,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_11-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_11-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_11-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_11-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.11"
+      build_name: manywheel-py3_11-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -941,6 +1079,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.11"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1053,6 +1194,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.11"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1429,6 +1573,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_12-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_12-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_12-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_12-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.12"
+      build_name: manywheel-py3_12-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -1535,6 +1745,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.12"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -1647,6 +1860,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.12"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2023,6 +2239,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13"
+      build_name: manywheel-py3_13-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2129,6 +2411,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2241,6 +2526,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2617,6 +2905,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_13t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_13t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_13t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_13t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.13t"
+      build_name: manywheel-py3_13t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -2723,6 +3077,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.13t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -2835,6 +3192,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.13t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3211,6 +3571,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14"
+      build_name: manywheel-py3_14-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3317,6 +3743,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3429,6 +3858,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -3805,6 +4237,72 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  manywheel-py3_14t-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-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'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_14t-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_14t-cuda12_9-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: manywheel-py3_14t-cuda12_9-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      DESIRED_PYTHON: "3.14t"
+      build_name: manywheel-py3_14t-cuda12_9
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   manywheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -3911,6 +4409,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
       DESIRED_PYTHON: "3.14t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
@@ -4023,6 +4524,9 @@ jobs:
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm7.0
       DESIRED_PYTHON: "3.14t"
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm

.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml

@@ -63,7 +63,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.10.4"
           freethreaded: false

.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml

@@ -59,7 +59,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.10.4"
           freethreaded: false
@@ -169,7 +168,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.11.4"
           freethreaded: false
@@ -279,7 +277,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.12.4"
           freethreaded: false
@@ -389,7 +386,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.13.4"
           freethreaded: false
@@ -499,7 +495,6 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
           python-version: "3.13.4"
           freethreaded: true
@@ -609,9 +604,8 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
+          python-version: "3.14.0"
           freethreaded: false
       - name: Checkout PyTorch
         uses: actions/checkout@v4
@@ -719,9 +713,8 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v6
         with:
-          # TODO: Removeme once 3.14 is out
           # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "3.14.0-rc.2"
+          python-version: "3.14.0"
           freethreaded: true
       - name: Checkout PyTorch
         uses: actions/checkout@v4

.github/workflows/h100-distributed.yml

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

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

@@ -2,7 +2,7 @@ name: inductor-perf-nightly-h100
 
 on:
   schedule:
-    - cron: 15 0,12 * * 1-6
+    - cron: 15 0 * * 1-6
     - cron: 0 7 * * 0
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
@@ -130,7 +130,7 @@ jobs:
     name: test-periodically
     uses: ./.github/workflows/_linux-test.yml
     needs: build
-    if: github.event.schedule == '15 0,12 * * 1-6'
+    if: github.event.schedule == '15 0 * * 1-6'
     with:
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

.github/workflows/inductor-perf-test-nightly-rocm-mi300.yml

@@ -0,0 +1,132 @@
+name: inductor-perf-nightly-rocm-mi300
+
+on:
+  push:
+    tags:
+      - ciflow/inductor-perf-test-nightly-rocm-mi300/*
+  schedule:
+    - cron: 15 0 * * *
+  # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
+  # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
+  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: true
+      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_rocm_mi300,inductor_timm_perf_rocm_mi300,inductor_torchbench_perf_rocm_mi300
+
+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
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-build:
+    if: github.repository_owner == 'pytorch'
+    name: rocm-py3_10-inductor-benchmark-build
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      build-environment: linux-jammy-rocm-py3_10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      test-matrix: |
+        { include: [
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi300", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_timm_perf_rocm_mi300", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_torchbench_perf_rocm_mi300", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-inductor-benchmark-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: rocm-py3_10-inductor-benchmark-test
+    uses: ./.github/workflows/_rocm-test.yml
+    needs: linux-jammy-rocm-py3_10-inductor-benchmark-build
+    with:
+      build-environment: linux-jammy-rocm-py3_10
+      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-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

.github/workflows/inductor-perf-test-nightly-rocm-mi355.yml

@@ -1,11 +1,11 @@
-name: inductor-perf-nightly-rocm
+name: inductor-perf-nightly-rocm-mi355
 
 on:
   push:
     tags:
-      - ciflow/inductor-perf-test-nightly-rocm/*
+      - ciflow/inductor-perf-test-nightly-rocm-mi355/*
   schedule:
-    - cron: 0 7 * * 0,3
+    - cron: 15 0 * * *
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
   workflow_dispatch:
@@ -59,7 +59,7 @@ on:
         description: The list of configs used the benchmark
         required: false
         type: string
-        default: inductor_huggingface_perf_rocm,inductor_timm_perf_rocm,inductor_torchbench_perf_rocm
+        default: inductor_huggingface_perf_rocm_mi355,inductor_timm_perf_rocm_mi355,inductor_torchbench_perf_rocm_mi355
 
 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' }}
@@ -88,23 +88,27 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
         ]}
     secrets: inherit
 

.github/workflows/lint.yml

@@ -12,6 +12,7 @@ on:
       - landchecks/*
     tags:
       - ciflow/pull/*
+      - ciflow/trunk/*
   workflow_dispatch:
 
 permissions: read-all
@@ -32,10 +33,12 @@ jobs:
     name: Get changed files
     uses: ./.github/workflows/_get-changed-files.yml
     with:
-      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') }}
+      all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') || github.event_name == 'push' }}
 
   lintrunner-clang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    # Needed to prevent deduping on HUD
+    name: lintrunner-clang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to clangtidy / clangformat
     if: |
@@ -75,6 +78,7 @@ jobs:
   #       fails to find types when it should
   lintrunner-mypy:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    name: lintrunner-mypy-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     # Only run if there are changed files relevant to mypy
     if: |
@@ -99,6 +103,7 @@ jobs:
 
   lintrunner-noclang:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    name: lintrunner-noclang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
     needs: [get-label-type, get-changed-files]
     with:
       timeout: 120
@@ -113,9 +118,9 @@ jobs:
         CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
         echo "Running all other linters"
         if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
         else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
         fi
 
   quick-checks:

.github/workflows/operator_benchmark.yml

@@ -7,9 +7,11 @@ on:
   workflow_dispatch:
     inputs:
       test_mode:
-        required: false
-        type: string
-        default: 'short'
+        type: choice
+        options:
+          - 'short'
+          - 'long'
+          - 'all'
         description: tag filter for operator benchmarks, options from long, short, all
   schedule:
     # Run at 07:00 UTC every Sunday
@@ -28,38 +30,49 @@ permissions:
   contents: read
 
 jobs:
-  opbenchmark-build:
+  x86-opbenchmark-build:
     if: github.repository_owner == 'pytorch'
-    name: opbenchmark-build
+    name: x86-opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
-          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
+          { config: "cpu_operator_benchmark_${{ inputs.test_mode || 'short' }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
         ]}
     secrets: inherit
 
-  opbenchmark-on-demand-build:
-    if: ${{ github.event_name == 'workflow_dispatch' && github.repository_owner == 'pytorch' }}
-    name: opbenchmark-on-demand-build
-    uses: ./.github/workflows/_linux-build.yml
-    with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
-      test-matrix: |
-        { include: [
-          { config: "cpu_operator_benchmark_${{ inputs.test_mode }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
-        ]}
-    secrets: inherit
-
-  opbenchmark-test:
-    name: opbenchmark-test
+  x86-opbenchmark-test:
+    name: x86-opbenchmark-test
     uses: ./.github/workflows/_linux-test.yml
-    needs: opbenchmark-build
+    needs: x86-opbenchmark-build
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
-      docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
+      docker-image: ${{ needs.x86-opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.x86-opbenchmark-build.outputs.test-matrix }}
+    secrets: inherit
+
+  aarch64-opbenchmark-build:
+    if: github.repository_owner == 'pytorch'
+    name: aarch64-opbenchmark-build
+    uses: ./.github/workflows/_linux-build.yml
+    with:
+      build-environment: linux-jammy-aarch64-py3.10
+      runner: linux.arm64.m7g.4xlarge
+      docker-image-name: ci-image:pytorch-linux-jammy-aarch64-py3.10-gcc11
+      test-matrix: |
+        { include: [
+          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
+        ]}
+    secrets: inherit
+
+  aarch64-opbenchmark-test:
+    name: aarch64-opbenchmark-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs: aarch64-opbenchmark-build
+    with:
+      build-environment: linux-jammy-aarch64-py3.10
+      docker-image: ${{ needs.aarch64-opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.aarch64-opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/periodic.yml

@@ -182,11 +182,11 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11
       test-matrix: |
         { include: [
-          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
-          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
         ]}
     secrets: inherit

.github/workflows/pull.yml

@@ -127,6 +127,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
+      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/rocm-mi355.yml

@@ -1,6 +1,9 @@
 name: rocm-mi355
 
 on:
+  push:
+    tags:
+      - ciflow/rocm-mi355/*
   workflow_dispatch:
   schedule:
     - cron: 30 11,1 * * *  # about 4:30am PDT and 6:30pm PDT
@@ -42,12 +45,12 @@ jobs:
       sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
         ]}
     secrets: inherit
 
@@ -64,5 +67,7 @@ jobs:
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image: ${{ needs.linux-noble-rocm-py3_12-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-noble-rocm-py3_12-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
+      tests-to-include: >-
+                        ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor test_matmul_cuda test_scaled_matmul_cuda'
+                           || '' }}
     secrets: inherit

.github/workflows/slow.yml

@@ -140,6 +140,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
+      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-py3.10-clang18-asan
       docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

.github/workflows/trunk.yml

@@ -56,7 +56,7 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
       build-generates-artifacts: false
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: "linux.4xlarge"
+      runner: "linux.c7i.4xlarge"
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 1 },
@@ -180,13 +180,13 @@ jobs:
       disable-monitor: false
     secrets: inherit
 
-  win-vs2022-cuda12_6-py3-build:
-    name: win-vs2022-cuda12.6-py3
+  win-vs2022-cuda12_8-py3-build:
+    name: win-vs2022-cuda12.8-py3
     uses: ./.github/workflows/_win-build.yml
     needs: get-label-type
     with:
-      build-environment: win-vs2022-cuda12.6-py3
-      cuda-version: "12.6"
+      build-environment: win-vs2022-cuda12.8-py3
+      cuda-version: "12.8"
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     secrets: inherit
 
@@ -200,6 +200,23 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
+  # Test cross-compiled models with Windows libs extracted from wheel
+  cross-compile-linux-test:
+    name: cross-compile-linux-test
+    uses: ./.github/workflows/_linux-test.yml
+    needs:
+      - linux-jammy-cuda12_8-py3_10-gcc11-build
+      - get-label-type
+      - win-vs2022-cuda12_8-py3-build
+    with:
+      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
+      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
+        ]}
+    secrets: inherit
+
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml
@@ -249,3 +266,14 @@ jobs:
       docker-image: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-py3-clang12-executorch-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-py3_10-gcc11-full-debug-build-only:
+    name: linux-jammy-py3.10-gcc11-full-debug-build-only
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: linux.2xlarge.memory
+      build-environment: linux-jammy-py3.10-gcc11-full-debug-build-only
+      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
+    secrets: inherit

.github/workflows/vllm.yml

@@ -46,7 +46,7 @@ jobs:
       runner: linux.24xlarge.memory
       test-matrix: |
         { include: [
-          { config:  "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
+          { config: "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_basic_models_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_entrypoints_test", shard: 1, num_shards: 1,runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_regression_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
@@ -54,7 +54,7 @@ jobs:
           { config: "vllm_pytorch_compilation_unit_tests", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_multi_model_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
-          { config: "vllm_languagde_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
+          { config: "vllm_language_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
           { config: "vllm_distributed_test_2_gpu_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 0, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
           { config: "vllm_lora_test", shard: 1, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },

.github/workflows/xpu.yml

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

.gitignore

@@ -374,6 +374,7 @@ third_party/ruy/
 third_party/glog/
 
 # Virtualenv
+.venv/
 venv/
 
 # Log files
@@ -395,3 +396,4 @@ android/pytorch_android_torchvision/.cxx
 CLAUDE.local.md
 /test_*.py
 /debug_*.py
+CLAUDE_CONTEXT/

.lintrunner.toml

@@ -209,6 +209,46 @@ command = [
     '@{{PATHSFILE}}'
 ]
 
+
+[[linter]]
+code = 'PYREFLY'
+include_patterns = [
+    'torch/**/*.py',
+    'torch/**/*.pyi',
+    'torchgen/**/*.py',
+    'torchgen/**/*.pyi',
+    'functorch/**/*.py',
+    'functorch/**/*.pyi',
+]
+exclude_patterns = []
+command = [
+    'python3',
+    'tools/linter/adapters/pyrefly_linter.py',
+    '--config=pyrefly.toml',
+]
+init_command = [
+    'python3',
+    'tools/linter/adapters/pip_init.py',
+    '--dry-run={{DRYRUN}}',
+    'numpy==2.1.0 ; python_version >= "3.12"',
+    'expecttest==0.3.0',
+    'pyrefly==0.36.2',
+    'sympy==1.13.3',
+    'types-requests==2.27.25',
+    'types-pyyaml==6.0.2',
+    'types-tabulate==0.8.8',
+    'types-protobuf==5.29.1.20250403',
+    'types-setuptools==79.0.0.20250422',
+    'types-jinja2==2.11.9',
+    'types-colorama==0.4.6',
+    'filelock==3.18.0',
+    'junitparser==2.1.1',
+    'rich==14.1.0',
+    'optree==0.17.0',
+    'types-openpyxl==3.1.5.20250919',
+    'types-python-dateutil==2.9.0.20251008'
+]
+
 [[linter]]
 code = 'CLANGTIDY'
 include_patterns = [

CMakeLists.txt

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

CODEOWNERS

@@ -201,3 +201,17 @@ torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
 /torch/csrc/stable/ @janeyx99 @mikaylagawarecki
 /torch/headeronly/ @janeyx99
 /torch/header_only_apis.txt @janeyx99
+
+# FlexAttention
+/torch/nn/attention/flex_attention.py @drisspg
+/torch/_higher_order_ops/flex_attention.py @drisspg
+/torch/_inductor/kernel/flex/ @drisspg
+/torch/_inductor/codegen/cpp_flex_attention_template.py @drisspg
+/test/inductor/test_flex_attention.py @drisspg
+/test/inductor/test_flex_decoding.py @drisspg
+
+# Low Precision GEMMs
+/aten/src/ATen/native/cuda/Blas.cpp @drisspg @slayton58
+/aten/src/ATen/cuda/CUDABlas.cpp @drisspg @slayton58
+/aten/src/ATen/cuda/CUDABlas.h @drisspg @slayton58
+/test/test_scaled_matmul_cuda.py @drisspg @slayton58

aten/src/ATen/BlasBackend.h

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

aten/src/ATen/CMakeLists.txt

@@ -256,6 +256,7 @@ endif()
 IF(USE_FBGEMM_GENAI)
   set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
   set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
+
   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.
@@ -288,62 +289,69 @@ IF(USE_FBGEMM_GENAI)
 
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-    set(fbgemm_genai_mx8mx8bf16_grouped
+    set(fbgemm_genai_cuh
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/"
     )
 
-    target_include_directories(fbgemm_genai PUBLIC
+    target_include_directories(fbgemm_genai PRIVATE
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_mx8mx8bf16_grouped}
+      ${fbgemm_genai_cuh}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
-  else()
-    if(USE_ROCM)
-      # Only include the kernels we want to build to avoid increasing binary size.
-      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-      # Add additional HIPCC compiler flags for performance
-      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-        -mllvm
-        -amdgpu-coerce-illegal-types=1
-        -mllvm
-        -enable-post-misched=0
-        -mllvm
-        -greedy-reverse-local-assignment=1
-        -fhip-new-launch-api)
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  elseif(USE_ROCM)
+    # Only include the kernels we want to build to avoid increasing binary size.
+    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-      # Only compile for gfx942 for now.
-      # This is rather hacky, I could not figure out a clean solution :(
-      set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
-      string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
-      if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-        list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
-      endif()
-      set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
+    # Add additional HIPCC compiler flags for performance
+    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+      -mllvm
+      -amdgpu-coerce-illegal-types=1
+      -mllvm
+      -enable-post-misched=0
+      -mllvm
+      -greedy-reverse-local-assignment=1
+      -fhip-new-launch-api)
 
-      hip_add_library(
-        fbgemm_genai STATIC
-        ${fbgemm_genai_native_rocm_hip}
-        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-      set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
-      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-      target_include_directories(fbgemm_genai PUBLIC
-        # FBGEMM version of Composable Kernel is used due to some customizations
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-      )
+    # Only compile for gfx942 for now.
+    # This is rather hacky, I could not figure out a clean solution :(
+    set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
+    string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
+    if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+      list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
     endif()
+    set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
+
+    hip_add_library(
+      fbgemm_genai STATIC
+      ${fbgemm_genai_native_rocm_hip}
+      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+    set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
+    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+    target_include_directories(fbgemm_genai PRIVATE
+      # FBGEMM version of Composable Kernel is used due to some customizations
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+    )
+
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
   endif()
 endif()
 
@@ -692,12 +700,6 @@ if(USE_CUDA AND NOT USE_ROCM)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
 
-  # Add FBGEMM_GENAI include directories for torch_ops.h
-  if(USE_FBGEMM_GENAI)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  endif()
-
   if($ENV{ATEN_STATIC_CUDA})
     if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
       list(APPEND ATen_CUDA_DEPENDENCY_LIBS

aten/src/ATen/CPUApplyUtils.h

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

aten/src/ATen/Context.cpp

@@ -587,20 +587,33 @@ void Context::setROCmFAPreferredBackend(at::ROCmFABackend b) {
   rocm_fa_preferred_backend = b;
 }
 
-bool Context::allowFP16ReductionCuBLAS() const {
+CuBLASReductionOption Context::allowFP16ReductionCuBLAS() const {
   return allow_fp16_reduction_cublas;
 }
 
-void Context::setAllowFP16ReductionCuBLAS(bool b) {
-  allow_fp16_reduction_cublas = b;
+CuBLASReductionOption inline get_reduction_option(bool allow_reduced_precision, bool allow_splitk) {
+  TORCH_CHECK(
+      !(allow_reduced_precision && !allow_splitk),
+      "allow_splitk=False is not supported when reduced precision reductions are enabled");
+  if (allow_reduced_precision) {
+    return CuBLASReductionOption::AllowReducedPrecisionWithSplitK;
+  } else if (allow_splitk) {
+    return CuBLASReductionOption::DisallowReducedPrecisionAllowSplitK;
+  } else {
+    return CuBLASReductionOption::DisallowReducedPrecisionDisallowSplitK;
+  }
 }
 
-bool Context::allowBF16ReductionCuBLAS() const {
+void Context::setAllowFP16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_fp16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
+}
+
+CuBLASReductionOption Context::allowBF16ReductionCuBLAS() const {
   return allow_bf16_reduction_cublas;
 }
 
-void Context::setAllowBF16ReductionCuBLAS(bool b) {
-  allow_bf16_reduction_cublas = b;
+void Context::setAllowBF16ReductionCuBLAS(bool allow_reduced_precision, bool allow_splitk) {
+  allow_bf16_reduction_cublas = get_reduction_option(allow_reduced_precision, allow_splitk);
 }
 
 bool Context::allowFP16AccumulationCuBLAS() const {

aten/src/ATen/Context.h

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

aten/src/ATen/DLConvertor.cpp

@@ -389,37 +389,16 @@ void fillVersion<DLManagedTensorVersioned>(
 // constructed out of ATen tensor
 template <class T>
 T* toDLPackImpl(const Tensor& src) {
-  auto view = src;
-
-  // Detect whether there is need to normalize the strides
-  // Background: gh-83069
-  //
-  // However, normalizing strides can come at a high-cost
-  // to slow down toDLPack conversion 3x, so we
-  // only normalize if needed.
-  //
-  // The following code detects whether the src follows
-  // a continuous pattern. If the src follows such pattern (common-case)
-  // then we do not need to normalize the strides.
-  bool need_normalize_strides = src.dim() == 1 && src.size(0) == 1 && src.stride(0) != 1;
-  // less common case, try normalizing the strides
-  if (need_normalize_strides) {
-    // create a new tensor with possibly normalized strides
-    // gh-83069
-    auto shape = src.sizes();
-    view = src.as_strided(shape, {1}, src.storage_offset());
-  }
-
   ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
-  atDLMTensor->handle = view;
+  atDLMTensor->handle = src;
   atDLMTensor->tensor.manager_ctx = atDLMTensor;
   atDLMTensor->tensor.deleter = &deleter<T>;
-  atDLMTensor->tensor.dl_tensor.data = view.data_ptr();
+  atDLMTensor->tensor.dl_tensor.data = src.data_ptr();
   atDLMTensor->tensor.dl_tensor.device = torchDeviceToDLDevice(src.device());
   atDLMTensor->tensor.dl_tensor.ndim = static_cast<int32_t>(src.dim());
   atDLMTensor->tensor.dl_tensor.dtype = getDLDataType(src);
-  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(view.sizes().data());
-  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(view.strides().data());
+  atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(src.sizes().data());
+  atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(src.strides().data());
   atDLMTensor->tensor.dl_tensor.byte_offset = 0;
   fillVersion(&atDLMTensor->tensor);
 

aten/src/ATen/DLConvertor.h

@@ -52,16 +52,16 @@ struct DLPackTraits {};
 
 template <>
 struct DLPackTraits<DLManagedTensor> {
-  inline static const char* capsule = "dltensor";
-  inline static const char* used = "used_dltensor";
+  inline static constexpr const char* capsule = "dltensor";
+  inline static constexpr const char* used = "used_dltensor";
   inline static auto toDLPack = at::toDLPack;
   inline static auto fromDLPack = at::fromDLPack;
 };
 
 template <>
 struct DLPackTraits<DLManagedTensorVersioned> {
-  inline static const char* capsule = "dltensor_versioned";
-  inline static const char* used = "used_dltensor_versioned";
+  inline static constexpr const char* capsule = "dltensor_versioned";
+  inline static constexpr const char* used = "used_dltensor_versioned";
   inline static auto toDLPack = at::toDLPackVersioned;
   inline static auto fromDLPack = at::fromDLPackVersioned;
 };

aten/src/ATen/EmptyTensor.h

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

aten/src/ATen/InferSize.h

@@ -4,6 +4,7 @@
 #include <c10/core/ScalarType.h>
 #include <c10/core/SymIntArrayRef.h>
 #include <c10/util/DimVector.h>
+#include <c10/util/Exception.h>
 #include <optional>
 #include <sstream>
 #include <vector>
@@ -26,9 +27,7 @@ inline void infer_size_impl(
   std::optional<int64_t> infer_dim;
   for (int64_t dim = 0, ndim = shape.size(); dim != ndim; dim++) {
     if (TORCH_GUARD_OR_FALSE(sym_eq(shape[dim], -1))) {
-      if (infer_dim) {
-        throw std::runtime_error("only one dimension can be inferred");
-      }
+      TORCH_CHECK(!infer_dim, "only one dimension can be inferred");
       infer_dim = dim;
     } else {
       // in case of unbacked shape[dim] we assume it's not -1 and add a runtime

aten/src/ATen/LegacyBatchingRegistrations.cpp

@@ -58,7 +58,7 @@ namespace at {
 namespace{
 
 // PyTorch allows operations to specify dim 0 and dim -1 on a scalar tensor.
-static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
+bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
   return dim == 0 || dim == -1;
 }
 
@@ -365,7 +365,7 @@ Tensor select_batching_rule(const Tensor& self, int64_t dim, int64_t index) {
   return self_physical.getPhysicalToLogicalMap().apply(result);
 }
 
-static int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
+int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
   return maybe_wrap_dim(dim, static_cast<int64_t>(input_sizes.size())) + num_batch_dims;
 }
 
@@ -488,7 +488,7 @@ Tensor view_as_complex_batching_rule(const Tensor& self) {
 // Checks that the smallest batch stride is greater than the largest example
 // stride. This is something we can support but we choose not to because it's
 // potentially error prone.
-static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
+void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
   auto smallest_batch_stride = std::min_element(
       physical_strides.begin(), physical_strides.begin() + num_batch_dims);
   auto largest_example_stride = std::max_element(
@@ -508,7 +508,7 @@ static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t
 // given (sizes, strides, storage_offset) returns the maximum location that
 // can be indexed (or nullopt if such a location doesn't exist, e.g., tensors
 // with zero-size dims).
-static std::optional<int64_t> maximum_indexable_location(
+std::optional<int64_t> maximum_indexable_location(
     IntArrayRef sizes, IntArrayRef strides, int64_t storage_offset) {
   auto result = native::storage_size_for(sizes, strides);
   if (result == 0) {
@@ -521,7 +521,7 @@ static std::optional<int64_t> maximum_indexable_location(
 // This checks that the range of possible memory locations accessible by
 // x.as_strided(sizes, strides, maybe_storage_offset)
 // are within the bounds of possible memory locations accessible by x.
-static void checkBasicAsStridedValidForSlice(
+void checkBasicAsStridedValidForSlice(
     const Tensor& physical_tensor,
     int64_t num_batch_dims,
     IntArrayRef sizes,

aten/src/ATen/MapAllocator.cpp

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

aten/src/ATen/MapAllocator.h

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

aten/src/ATen/NestedTensorImpl.cpp

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

aten/src/ATen/NestedTensorImpl.h

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

aten/src/ATen/Parallel.h

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

aten/src/ATen/PythonTorchFunctionTLS.cpp

@@ -42,8 +42,14 @@ const PythonTorchFunctionTLS& PythonTorchFunctionTLS::get_state() {
 }
 
 bool torch_function_mode_enabled() {
-  return PythonTorchFunctionTLS::get_disabled_state() != TorchFunctionDisabledState::ALL_DISABLED &&
-         PythonTorchFunctionTLS::stack_len() > 0;
+  // Manually flatten because gcc is refusing to inline here.  Note
+  // that we are still calling __tls_get_addr twice here with GCC,
+  // presumably because of
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81501 (which says
+  // the fix ships in GCC 16), but forcing inlining still improves
+  // performance.
+  const auto& ptfs = pythonTorchFunctionState;
+  return ptfs.disabled_state_ != TorchFunctionDisabledState::ALL_DISABLED && !ptfs.stack_.empty();
 }
 
 // This is needed to disambiguate the ternary torch function disabled states

aten/src/ATen/PythonTorchFunctionTLS.h

@@ -27,6 +27,7 @@ struct TORCH_API PythonTorchFunctionTLS {
   TorchFunctionDisabledState disabled_state_ =
       TorchFunctionDisabledState::ENABLED;
   std::vector<std::shared_ptr<c10::SafePyObject>> stack_;
+  friend TORCH_API bool torch_function_mode_enabled();
 };
 
 TORCH_API bool torch_function_mode_enabled();

aten/src/ATen/SavedTensorHooks.cpp

@@ -13,7 +13,7 @@ namespace {
   // and left at true for the rest of the execution.
   // It's an optimization so that users who never use default hooks don't need to
   // read the thread_local variables pack_hook_ and unpack_hook_.
-  static bool is_initialized(false);
+  bool is_initialized(false);
 }
 
 static void assertSavedTensorHooksNotDisabled() {

aten/src/ATen/SparseCsrTensorImpl.cpp

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

aten/src/ATen/SparseCsrTensorImpl.h

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

aten/src/ATen/SparseTensorImpl.h

@@ -46,7 +46,9 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
 
  public:
   // Public for now...
-  explicit SparseTensorImpl(at::DispatchKeySet, const caffe2::TypeMeta);
+  explicit SparseTensorImpl(
+      at::DispatchKeySet /*key_set*/,
+      const caffe2::TypeMeta /*data_type*/);
 
   void release_resources() override;
 
@@ -384,8 +386,8 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
 
  private:
   explicit SparseTensorImpl(
-      at::DispatchKeySet,
-      const caffe2::TypeMeta,
+      at::DispatchKeySet /*key_set*/,
+      const caffe2::TypeMeta /*data_type*/,
       at::Tensor indices,
       at::Tensor values);
 

aten/src/ATen/TensorIndexing.h

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

aten/src/ATen/TensorIterator.cpp

@@ -56,7 +56,7 @@ inline void get_strides(int64_t* strides, ArrayRef<OperandInfo> operands, int64_
   }
 }
 
-static OptionalTensorRef make_otr(const TensorBase &tensor) {
+OptionalTensorRef make_otr(const TensorBase &tensor) {
   if (tensor.defined()) {
     return OptionalTensorRef(tensor);
   } else {

aten/src/ATen/TensorIterator.h

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

aten/src/ATen/Utils.h

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

aten/src/ATen/Version.cpp

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

aten/src/ATen/autocast_mode.cpp

@@ -36,7 +36,7 @@ namespace {
 using weakref_type = c10::weak_intrusive_ptr<TensorImpl, UndefinedTensorImpl>;
 using val_type = std::tuple<weakref_type, Tensor>;
 
-static ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
+ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
   static ska::flat_hash_map<TensorImpl*, val_type> cached_casts;
   return cached_casts;
 }
@@ -148,7 +148,7 @@ Tensor cached_cast(at::ScalarType to_type, const Tensor& arg, DeviceType device_
 Banned functions
 *******************************/
 
-static Tensor binary_cross_entropy_banned(const Tensor &, const Tensor &, const std::optional<Tensor>&, int64_t) {
+static Tensor binary_cross_entropy_banned(const Tensor & /*unused*/, const Tensor & /*unused*/, const std::optional<Tensor>& /*unused*/, int64_t /*unused*/) {
   TORCH_CHECK(false, "torch.nn.functional.binary_cross_entropy and torch.nn.BCELoss are unsafe to autocast.\n"
            "Many models use a sigmoid layer right before the binary cross entropy layer.\n"
            "In this case, combine the two layers using torch.nn.functional.binary_cross_entropy_with_logits\n"

aten/src/ATen/core/CachingHostAllocator.cpp

@@ -6,9 +6,9 @@ namespace at {
 
 namespace {
 
-static std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_array{};
-static std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
+std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
     allocator_priority{};
 
 } // anonymous namespace

aten/src/ATen/core/NamedTensor.h

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

aten/src/ATen/core/PhiloxRNGEngine.h

@@ -229,10 +229,10 @@ private:
   }
 
 
-  static const uint32_t kPhilox10A = 0x9E3779B9;
-  static const uint32_t kPhilox10B = 0xBB67AE85;
-  static const uint32_t kPhiloxSA = 0xD2511F53;
-  static const uint32_t kPhiloxSB = 0xCD9E8D57;
+  static constexpr uint32_t kPhilox10A = 0x9E3779B9;
+  static constexpr uint32_t kPhilox10B = 0xBB67AE85;
+  static constexpr uint32_t kPhiloxSA = 0xD2511F53;
+  static constexpr uint32_t kPhiloxSB = 0xCD9E8D57;
 };
 
 typedef philox_engine Philox4_32;