diff --git a/.ci/manywheel/build_cuda.sh b/.ci/manywheel/build_cuda.sh
index bf2622cc55b..f6b2cb61d7b 100644
--- a/.ci/manywheel/build_cuda.sh
+++ b/.ci/manywheel/build_cuda.sh
@@ -54,7 +54,7 @@ cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
 TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6"
 case ${CUDA_VERSION} in
     12.8)
-        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0;10.0;12.0+PTX" #Ripping out 5.0 and 6.0 due to ld error
+        TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0;10.0;12.0+PTX" #Ripping out 5.0 and 6.0 due to ld error
         EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
     12.6)
diff --git a/.github/scripts/generate_binary_build_matrix.py b/.github/scripts/generate_binary_build_matrix.py
index 5b50917fc70..67160bda12c 100644
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@@ -246,9 +246,6 @@ def generate_libtorch_matrix(
         if os == "linux":
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
-            # skip CUDA 12.8 builds for libtorch
-            if "12.8" in arches:
-                arches.remove("12.8")
         elif os == "windows":
             arches += CUDA_ARCHES
     if libtorch_variants is None:
diff --git a/.github/templates/linux_binary_build_workflow.yml.j2 b/.github/templates/linux_binary_build_workflow.yml.j2
index 5b1acca98fe..e0fda97e324 100644
--- a/.github/templates/linux_binary_build_workflow.yml.j2
+++ b/.github/templates/linux_binary_build_workflow.yml.j2
@@ -111,7 +111,10 @@ jobs:
       ALPINE_IMAGE: "docker.io/s390x/alpine"
       {%- elif config["gpu_arch_type"] == "rocm" %}
       runs_on: linux.rocm.gpu
-      {%- elif config["gpu_arch_type"] == "cuda" %}
+      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] == "12.8" %}
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
+      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] != "12.8"%}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.4xlarge.nvidia.gpu
       {%- else %}
diff --git a/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml b/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml
index 415e0a0645e..4114ab11f05 100644
--- a/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml
+++ b/.github/workflows/generated-linux-binary-libtorch-cxx11-abi-nightly.yml
@@ -301,6 +301,71 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-cuda12_8-shared-with-deps-cxx11-abi-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: 12.8
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/libtorch-cxx11-builder:cuda12.8-main
+      LIBTORCH_VARIANT: shared-with-deps
+      DESIRED_DEVTOOLSET: cxx11-abi
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: libtorch-cuda12_8-shared-with-deps-cxx11-abi
+      build_environment: linux-binary-libtorch-cxx11-abi
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_8-shared-with-deps-cxx11-abi-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_8-shared-with-deps-cxx11-abi-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu128
+      GPU_ARCH_VERSION: 12.8
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/libtorch-cxx11-builder:cuda12.8-main
+      LIBTORCH_VARIANT: shared-with-deps
+      DESIRED_DEVTOOLSET: cxx11-abi
+      build_name: libtorch-cuda12_8-shared-with-deps-cxx11-abi
+      build_environment: linux-binary-libtorch-cxx11-abi
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-cuda12_8-shared-with-deps-cxx11-abi-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_8-shared-with-deps-cxx11-abi-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: cu128
+      GPU_ARCH_VERSION: 12.8
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: pytorch/libtorch-cxx11-builder:cuda12.8-main
+      LIBTORCH_VARIANT: shared-with-deps
+      DESIRED_DEVTOOLSET: cxx11-abi
+      build_name: libtorch-cuda12_8-shared-with-deps-cxx11-abi
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-rocm6_2_4-shared-with-deps-cxx11-abi-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
diff --git a/.github/workflows/generated-linux-binary-manywheel-main.yml b/.github/workflows/generated-linux-binary-manywheel-main.yml
index 3a8530cce7a..e619743227e 100644
--- a/.github/workflows/generated-linux-binary-manywheel-main.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-main.yml
@@ -223,6 +223,6 @@ jobs:
       build_name: manywheel-py3_9-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
diff --git a/.github/workflows/generated-linux-binary-manywheel-nightly.yml b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
index cfc4dbd7b73..72c26af7bda 100644
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
@@ -424,7 +424,7 @@ jobs:
       build_name: manywheel-py3_9-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_9-cuda12_8-upload:  # Uploading
@@ -1122,7 +1122,7 @@ jobs:
       build_name: manywheel-py3_10-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_8-upload:  # Uploading
@@ -1885,7 +1885,7 @@ jobs:
       build_name: manywheel-py3_11-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_8-upload:  # Uploading
@@ -2583,7 +2583,7 @@ jobs:
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-upload:  # Uploading
@@ -3281,7 +3281,7 @@ jobs:
       build_name: manywheel-py3_13-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_8-upload:  # Uploading
@@ -3979,7 +3979,7 @@ jobs:
       build_name: manywheel-py3_13t-cuda12_8
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 build needs sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_8-upload:  # Uploading