{
  "dataset": "pytorch",
  "source": "https://gpucompat.com/data/pytorch.json",
  "about": "https://gpucompat.com/about-the-data/",
  "count": 18,
  "entries": [
    {
      "id": "1-11-0",
      "version": "1.11.0",
      "releaseDate": "2022-03-10",
      "cudaVersions": [
        "10.2",
        "11.3"
      ],
      "rocmVersions": [
        "4.5.2"
      ],
      "pythonVersions": [
        "3.7",
        "3.8",
        "3.9",
        "3.10"
      ],
      "windowsCudaVersions": [
        "11.3"
      ],
      "cudaArchs": {
        "10.2": [
          "3.7",
          "5.0",
          "6.0",
          "7.0"
        ],
        "11.3": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ]
      },
      "notes": "Added 2026-09 at the user's request to extend coverage back past the existing 1.13.0 floor. CUDA/ROCm build matrix and Python range from PyTorch's official previous-versions page (pytorch.org/get-started/previous-versions/#v1110) and confirmed via PyPI release metadata (uploaded 2022-03-10). This is the release the rocm/4-5-2.yaml entry's ROCm wheel pairs with -- that entry's pytorchVersions has been updated to link here now that this page exists. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: Python as 3.7, 3.8, 3.9. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. No defaultPipCuda: this release's PyPI wheel bundles CUDA without machine-readable version metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "1-12-0",
      "version": "1.12.0",
      "releaseDate": "2022-06-28",
      "cudaVersions": [
        "10.2",
        "11.3",
        "11.6"
      ],
      "rocmVersions": [
        "5.0",
        "5.1.1"
      ],
      "pythonVersions": [
        "3.7",
        "3.8",
        "3.9",
        "3.10"
      ],
      "windowsCudaVersions": [
        "11.3",
        "11.6"
      ],
      "cudaArchs": {
        "10.2": [
          "3.7",
          "5.0",
          "6.0",
          "7.0"
        ],
        "11.3": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ],
        "11.6": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ]
      },
      "notes": "Added 2026-09 alongside 1.11.0 at the user's request. CUDA/ROCm build matrix and Python range from PyTorch's official previous-versions page (pytorch.org/get-started/previous-versions/#v1120) and confirmed via PyPI release metadata (uploaded 2022-06-28). PyTorch also shipped a 1.12.1 patch release (2022-08-05) with the same CUDA/ROCm/Python matrix; omitted here since this collection tracks one entry per minor version elsewhere (no 2.x.y patch releases are listed either). Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 5.1.1; Python as 3.7, 3.8, 3.9. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. No defaultPipCuda: this release's PyPI wheel bundles CUDA without machine-readable version metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "1-13-0",
      "version": "1.13.0",
      "releaseDate": "2022-10-28",
      "cudaVersions": [
        "11.6",
        "11.7"
      ],
      "rocmVersions": [
        "5.1.1",
        "5.2"
      ],
      "pythonVersions": [
        "3.7",
        "3.8",
        "3.9",
        "3.10",
        "3.11"
      ],
      "windowsCudaVersions": [
        "11.6",
        "11.7"
      ],
      "defaultPipCuda": "11.7",
      "cudaArchs": {
        "11.6": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ],
        "11.7": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ]
      },
      "notes": "Verified 2026-08 against PyTorch's official previous-versions page and PyPI release history. This is the last PyTorch release to support CUDA 10.2 — dropped starting with 2.0.0. Python-version floor/ceiling is best-effort, not independently confirmed per-release, same caveat as the rest of this collection. Added to fix dead cross-links: ROCm 5.7.0 and 6.1.0's pages listed this version as a compatible build before this entry existed. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 10.2, 11.6, 11.7; ROCm builds as 5.2; Python as 3.7, 3.8, 3.9, 3.10. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-0-0",
      "version": "2.0.0",
      "releaseDate": "2023-03-15",
      "cudaVersions": [
        "11.7",
        "11.8"
      ],
      "rocmVersions": [
        "5.3",
        "5.4.2"
      ],
      "pythonVersions": [
        "3.8",
        "3.9",
        "3.10",
        "3.11"
      ],
      "windowsCudaVersions": [
        "11.7",
        "11.8"
      ],
      "defaultPipCuda": "11.7",
      "cudaArchs": {
        "11.7": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6"
        ],
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "Verified 2026-08 against PyTorch's official previous-versions page and PyPI release history. Python-version range is best-effort, not independently confirmed per-release. Added to fix dead cross-links: ROCm 6.1.0's page listed this version as a compatible build before this entry existed. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 5.3, 5.4. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-1-0",
      "version": "2.1.0",
      "releaseDate": "2023-10-04",
      "cudaVersions": [
        "11.8",
        "12.1"
      ],
      "rocmVersions": [
        "5.5",
        "5.6"
      ],
      "pythonVersions": [
        "3.8",
        "3.9",
        "3.10",
        "3.11"
      ],
      "windowsCudaVersions": [
        "11.8",
        "12.1"
      ],
      "defaultPipCuda": "12.1",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.1": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "Verified 2026-08, with a caveat: one automated pull of PyTorch's official previous-versions page returned CUDA 13.0/13.2 and ROCm 7.2 for this entry — those TLDs didn't exist until 2026 and cannot be correct for a package released in October 2023 (pytorch.org's previous-versions page appears to get its install-command index rewritten over time to whatever wheels are still hosted, rather than staying a fixed historical record). Discarded that anomalous data rather than publish it; CUDA 11.8/12.1 and ROCm 5.6 used instead, cross-referenced against the surrounding 2.0.x/2.1.1 entries' progression and PyTorch's own 2.1.0 release notes. Added to fix dead cross-links: ROCm 5.7.0's page listed this version as a compatible build before this entry existed. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 5.6. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-10-0",
      "version": "2.10.0",
      "releaseDate": "2026-01-21",
      "cudaVersions": [
        "12.6",
        "12.8",
        "12.9",
        "13.0"
      ],
      "rocmVersions": [
        "7.0",
        "7.1"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14"
      ],
      "rocmArchs": {
        "7.0": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ],
        "7.1": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "12.8",
        "13.0"
      ],
      "defaultPipCuda": "12.8",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.8": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "12.9": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Release date verified against PyPI release history. Python 3.14 gains full (non-preview) support this release. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 12.1, 12.6, 12.8, 13.0; ROCm builds as 7.1. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-11-0",
      "version": "2.11.0",
      "releaseDate": "2026-03-23",
      "cudaVersions": [
        "12.6",
        "12.8",
        "12.9",
        "13.0"
      ],
      "rocmVersions": [
        "7.1",
        "7.2"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14"
      ],
      "rocmArchs": {
        "7.1": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ],
        "7.2": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "12.8",
        "13.0"
      ],
      "defaultPipCuda": "13.0",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.8": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "12.9": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix and release date verified against official PyTorch sources, 2026-08. Python range (3.10-3.14) confirmed 2026-09 directly against the Requires-Python metadata and classifiers on pypi.org/project/torch/2.11.0/ — matches the value carried forward from 2.10.0. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 12.1, 12.6, 12.8, 13.0; ROCm builds as 7.2. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-12-0",
      "version": "2.12.0",
      "releaseDate": "2026-05-13",
      "cudaVersions": [
        "12.6",
        "13.0",
        "13.2"
      ],
      "rocmVersions": [
        "7.1",
        "7.2"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14"
      ],
      "rocmArchs": {
        "7.1": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ],
        "7.2": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "13.0",
        "13.2"
      ],
      "defaultPipCuda": "13.0",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.2": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix and release date verified against official PyTorch sources, 2026-08. Python range (3.10-3.14) confirmed 2026-09 directly against the Requires-Python metadata and classifiers on pypi.org/project/torch/2.12.0/. Patch release 2.12.1 (Jun 17, 2026) has the same CUDA/ROCm matrix. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 12.1, 12.6, 13.0, 13.2; ROCm builds as 7.2. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-13-0",
      "version": "2.13.0",
      "releaseDate": "2026-07-08",
      "cudaVersions": [
        "12.6",
        "12.9",
        "13.0",
        "13.2"
      ],
      "rocmVersions": [
        "7.1",
        "7.2"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14",
        "3.15"
      ],
      "rocmArchs": {
        "7.1": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1103",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ],
        "7.2": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1103",
          "gfx1200",
          "gfx1201",
          "gfx950",
          "gfx1150",
          "gfx1151"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "13.0",
        "13.2"
      ],
      "defaultPipCuda": "13.0",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.9": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.2": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against download.pytorch.org wheel indexes (cu126/cu130/cu132/rocm7.2 present for 2.13.0; cu128/cu121/cu131 absent), 2026-08. Release date from PyPI upload metadata. Adds Python 3.15 wheels (with free-threaded 3.14t/3.15t variants); no longer ships cp313t (free-threaded 3.13) wheels — users on that variant should move to 3.14t/3.15t. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 12.6, 13.0, 13.2; ROCm builds as 7.2. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-14-0",
      "version": "2.14.0",
      "releaseDate": "2026-09-02",
      "cudaVersions": [
        "12.6",
        "13.0",
        "13.2"
      ],
      "rocmVersions": [
        "7.2",
        "7.14"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14",
        "3.15"
      ],
      "rocmArchs": {
        "7.2": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx950",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1103",
          "gfx1200",
          "gfx1201",
          "gfx1150",
          "gfx1151"
        ],
        "7.14": [
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx950",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1103",
          "gfx1200",
          "gfx1201",
          "gfx1150",
          "gfx1151"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "13.0",
        "13.2"
      ],
      "defaultPipCuda": "13.0",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.2": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "Added 2026-09-27. Release date from PyPI upload metadata. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch); the ROCm 7.14 build uses the TheRock-based image and drops gfx900/gfx906, the 7.2 legacy build keeps them. xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-2-0",
      "version": "2.2.0",
      "releaseDate": "2024-01-30",
      "cudaVersions": [
        "11.8",
        "12.1"
      ],
      "rocmVersions": [
        "5.6",
        "5.7"
      ],
      "pythonVersions": [
        "3.8",
        "3.9",
        "3.10",
        "3.11",
        "3.12"
      ],
      "windowsCudaVersions": [
        "11.8",
        "12.1"
      ],
      "defaultPipCuda": "12.1",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.1": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "Verified 2026-08 against PyTorch's official previous-versions page and PyPI release history. Python-version range is best-effort, not independently confirmed per-release. Added to fix dead cross-links: ROCm 6.2.1's page listed this version as a compatible build before this entry existed. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 6.0; Python as 3.9, 3.10, 3.11, 3.12. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-3-0",
      "version": "2.3.0",
      "releaseDate": "2024-04-24",
      "cudaVersions": [
        "11.8",
        "12.1"
      ],
      "rocmVersions": [
        "5.7",
        "6.0"
      ],
      "pythonVersions": [
        "3.8",
        "3.9",
        "3.10",
        "3.11",
        "3.12"
      ],
      "windowsCudaVersions": [
        "11.8",
        "12.1"
      ],
      "defaultPipCuda": "12.1",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.1": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "Verified 2026-08 against PyTorch's official previous-versions page and PyPI release history. Python-version range is best-effort, not independently confirmed per-release. Added to fix dead cross-links: ROCm 6.2.1's and 6.4.3's pages listed this version as a compatible build before this entry existed. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 6.0. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-4-0",
      "version": "2.4.0",
      "releaseDate": "2024-07",
      "cudaVersions": [
        "11.8",
        "12.1",
        "12.4"
      ],
      "rocmVersions": [
        "6.0",
        "6.1"
      ],
      "pythonVersions": [
        "3.8",
        "3.9",
        "3.10",
        "3.11",
        "3.12"
      ],
      "windowsCudaVersions": [
        "11.8",
        "12.1",
        "12.4"
      ],
      "defaultPipCuda": "12.1",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.1": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.4": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: ROCm builds as 6.1. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-5-0",
      "version": "2.5.0",
      "releaseDate": "2024-10",
      "cudaVersions": [
        "11.8",
        "12.1",
        "12.4"
      ],
      "rocmVersions": [
        "6.1",
        "6.2"
      ],
      "pythonVersions": [
        "3.9",
        "3.10",
        "3.11",
        "3.12",
        "3.13"
      ],
      "windowsCudaVersions": [
        "11.8",
        "12.1",
        "12.4"
      ],
      "defaultPipCuda": "12.4",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.1": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.4": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: Python as 3.9, 3.10, 3.11, 3.12. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases."
    },
    {
      "id": "2-6-0",
      "version": "2.6.0",
      "releaseDate": "2025-01-29",
      "cudaVersions": [
        "11.8",
        "12.4",
        "12.6"
      ],
      "rocmVersions": [
        "6.1",
        "6.2.4"
      ],
      "pythonVersions": [
        "3.9",
        "3.10",
        "3.11",
        "3.12",
        "3.13"
      ],
      "rocmArchs": {
        "6.1": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101"
        ],
        "6.2.4": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "11.8",
        "12.4",
        "12.6"
      ],
      "defaultPipCuda": "12.4",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.4": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Release date verified against PyPI release history. Python 3.13 support added this release. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 11.8, 12.1, 12.4, 12.6. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-7-0",
      "version": "2.7.0",
      "releaseDate": "2025-04-23",
      "cudaVersions": [
        "11.8",
        "12.6",
        "12.8"
      ],
      "rocmVersions": [
        "6.2.4",
        "6.3"
      ],
      "pythonVersions": [
        "3.9",
        "3.10",
        "3.11",
        "3.12",
        "3.13"
      ],
      "rocmArchs": {
        "6.2.4": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ],
        "6.3": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "11.8",
        "12.6",
        "12.8"
      ],
      "defaultPipCuda": "12.6",
      "cudaArchs": {
        "11.8": [
          "3.7",
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.8": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Release date verified against PyPI release history. Patch release 2.7.1 (Jun 4, 2025) has the same CUDA/ROCm matrix. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 11.8, 12.1, 12.6, 12.8; ROCm builds as 6.3. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-8-0",
      "version": "2.8.0",
      "releaseDate": "2025-08-06",
      "cudaVersions": [
        "12.6",
        "12.8",
        "12.9"
      ],
      "rocmVersions": [
        "6.3",
        "6.4"
      ],
      "pythonVersions": [
        "3.9",
        "3.10",
        "3.11",
        "3.12",
        "3.13"
      ],
      "rocmArchs": {
        "6.3": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ],
        "6.4": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "12.8",
        "12.9"
      ],
      "defaultPipCuda": "12.8",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.8": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "12.9": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Release date verified against PyPI release history. Python floor (3.9) confirmed 2026-09 directly against the Requires-Python metadata and classifiers on pypi.org/project/torch/2.8.0/. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 11.8, 12.1, 12.6, 12.8, 12.9; ROCm builds as 6.4. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    },
    {
      "id": "2-9-0",
      "version": "2.9.0",
      "releaseDate": "2025-10-15",
      "cudaVersions": [
        "12.6",
        "12.8",
        "12.9",
        "13.0"
      ],
      "rocmVersions": [
        "6.3",
        "6.4"
      ],
      "pythonVersions": [
        "3.10",
        "3.11",
        "3.12",
        "3.13",
        "3.14 (preview)"
      ],
      "rocmArchs": {
        "6.3": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ],
        "6.4": [
          "gfx900",
          "gfx906",
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx1030",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1200",
          "gfx1201"
        ]
      },
      "xpuBuild": true,
      "windowsCudaVersions": [
        "12.6",
        "12.8",
        "12.9",
        "13.0"
      ],
      "defaultPipCuda": "12.8",
      "cudaArchs": {
        "12.6": [
          "5.0",
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0"
        ],
        "12.8": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "12.9": [
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ],
        "13.0": [
          "7.5",
          "8.0",
          "8.6",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "notes": "CUDA/ROCm matrix verified against the official PyTorch previous-versions page, 2026-08. Release date verified against PyPI release history. Minimum Python version confirmed as 3.10 (dropped 3.9) via PyTorch dev-discuss; Python 3.14 available as preview this release. CUDA 11.8 wheels dropped starting this version. Build variants re-verified 2026-09-27 against the files actually published: download.pytorch.org/whl/<variant>/torch/ (x86_64 Linux) for the CUDA and ROCm builds, and PyPI plus that index for Python versions. Corrected then: CUDA builds were listed as 12.1, 12.6, 12.8, 13.0; ROCm builds as 6.4. cudaArchs is each CUDA build's TORCH_CUDA_ARCH_LIST for x86_64 Linux wheels, taken by running the arch-selection code in PyTorch's own build scripts at this release (pytorch/builder release branch manywheel/build_cuda.sh up to 2.5, pytorch .ci/manywheel/build_cuda.sh for 2.6-2.12, .ci/manywheel/build_env_setup.py from 2.13); +PTX suffixes are omitted. defaultPipCuda: the CUDA version a plain `pip install torch` gets on Linux x86_64, from the PyPI wheel's own dependency metadata. windowsCudaVersions: the CUDA builds with win_amd64 wheels on the same index (2026-09-27); Windows wheels are assumed to use the Linux architecture lists, as PyTorch's RELEASE.md states for recent releases. rocmArchs: PYTORCH_ROCM_ARCH per ROCm build, from pytorch .ci/docker/manywheel/build.sh at this tag (manylinux2_28-builder:rocm* branch). xpuBuild: an Intel GPU (XPU) build of this release is on download.pytorch.org/whl/xpu (checked 2026-09-27)."
    }
  ]
}