{
  "dataset": "libraries",
  "source": "https://gpucompat.com/data/libraries.json",
  "about": "https://gpucompat.com/about-the-data/",
  "count": 6,
  "entries": [
    {
      "id": "bitsandbytes",
      "name": "bitsandbytes",
      "version": "0.50.2",
      "verified": "2026-09-27",
      "order": 4,
      "nvidia": {
        "compiledFor": [
          "6.0",
          "7.0",
          "7.5",
          "8.0",
          "8.6",
          "8.9",
          "9.0",
          "10.0",
          "12.0"
        ],
        "features": [
          {
            "name": "LLM.int8()",
            "minComputeCapability": "7.5"
          }
        ]
      },
      "amd": {
        "gfxTargets": [
          "gfx908",
          "gfx90a",
          "gfx942",
          "gfx950",
          "gfx1250",
          "gfx1010",
          "gfx1011",
          "gfx1012",
          "gfx1030",
          "gfx1031",
          "gfx1032",
          "gfx1033",
          "gfx1034",
          "gfx1035",
          "gfx1036",
          "gfx1100",
          "gfx1101",
          "gfx1102",
          "gfx1103",
          "gfx1150",
          "gfx1151",
          "gfx1152",
          "gfx1153",
          "gfx1200",
          "gfx1201"
        ],
        "note": "Pre-built for ROCm 7.14.0 and 10.0.0 (Linux x86_64); older ROCm builds cover fewer RDNA targets."
      },
      "intel": {
        "supported": true,
        "note": "Intel XPU through PyTorch's XPU backend (PyTorch 2.6+)."
      },
      "sources": [
        "https://github.com/bitsandbytes-foundation/bitsandbytes/blob/main/docs/source/installation.mdx"
      ],
      "notes": "From the installation docs (2026-09-27): NVIDIA compute capability 6.0+, with LLM.int8() needing 7.5+ (8-bit optimizers and NF4/FP4 quantization need 6.0+). compiledFor is the union of the Linux x86_64 PyPI wheel targets across its CUDA builds (11.8-12.6: sm60-sm90; 12.8: sm70-sm120; 13.0-13.2: sm75-sm120). The AMD list is the ROCm 7.14.0 / 10.0.0 Linux wheel targets. The RX 6700 XT's gfx1031 and RX 6600's gfx1032 are included there, though ROCm itself doesn't officially support those cards.\n"
    },
    {
      "id": "flash-attention-2",
      "name": "FlashAttention-2",
      "version": "2.8.3.post1",
      "verified": "2026-09-27",
      "order": 2,
      "nvidia": {
        "compiledFor": [
          "8.0",
          "9.0",
          "10.0",
          "12.0"
        ]
      },
      "amd": {
        "gfxTargets": [
          "gfx90a",
          "gfx942",
          "gfx950"
        ],
        "note": "Composable Kernel backend, ROCm 6.0+. RDNA 3/4 and Strix targets are allowed on the main branch but not in a release yet."
      },
      "intel": {
        "supported": false
      },
      "sources": [
        "https://github.com/Dao-AILab/flash-attention/blob/v2.8.3.post1/setup.py",
        "https://github.com/Dao-AILab/flash-attention/blob/main/README.md"
      ],
      "notes": "compiledFor is the default FLASH_ATTN_CUDA_ARCHS in setup.py at the v2.8.3.post1 tag (sm_100/sm_120 need a CUDA 12.8+ build). The README text names Ampere, Ada and Hopper; the build also covers Blackwell. Ada (8.9) runs the sm_80 code under CUDA's same-major compatibility. Turing (T4, RTX 20xx) is not supported here; the README points to the separate flash-attention-turing project. The AMD list is setup.py's allowed ROCm archs at the same tag. The main branch (2026-09-27) also allows gfx1100-1102, gfx1150/1151 and gfx1200/1201, which the README already advertises, but no release includes them yet.\n"
    },
    {
      "id": "flash-attention-3",
      "name": "FlashAttention-3",
      "version": "main (2026-09-27)",
      "verified": "2026-09-27",
      "order": 3,
      "nvidia": {
        "compiledFor": [
          "9.0"
        ]
      },
      "amd": {
        "gfxTargets": []
      },
      "intel": {
        "supported": false
      },
      "sources": [
        "https://github.com/Dao-AILab/flash-attention/blob/main/README.md"
      ],
      "notes": "The README's stated requirement is \"H100 / H800 GPU, CUDA >= 12.3\". FlashAttention-3 is built from the repo's hopper/ directory and is Hopper-only. FlashAttention-4 (the flash-attn-4 package, 4.0.0b32 on 2026-09-27) is a beta and is not tracked here yet.\n"
    },
    {
      "id": "llama-cpp",
      "name": "llama.cpp",
      "version": "master (2026-09-27)",
      "verified": "2026-09-27",
      "order": 6,
      "nvidia": {
        "minComputeCapability": "5.0",
        "fallback": "Vulkan backend only",
        "note": "CUDA backend. Compute capability 5.0-7.0 needs a CUDA 12 build (CUDA 13 builds start at 7.5). Older cards can use the Vulkan backend."
      },
      "amd": {
        "gfxTargets": [],
        "rocmTargets": true,
        "fallback": "Vulkan backend only",
        "note": "HIP backend on cards ROCm supports; the Vulkan backend works on the rest."
      },
      "intel": {
        "supported": true,
        "note": "SYCL or Vulkan backend."
      },
      "sources": [
        "https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-cuda/CMakeLists.txt",
        "https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md"
      ],
      "notes": "llama.cpp has several GPU backends, so it runs on almost any card; what varies is which backend. The CUDA backend's default build targets are 50/61/70-virtual (CUDA < 13 only), 75/80-virtual, 86-real, 89-real, 90-virtual and 120a-real (CUDA 12.8+), per ggml-cuda's CMakeLists.txt.\n"
    },
    {
      "id": "tensorrt",
      "name": "TensorRT",
      "version": "11.3.0",
      "verified": "2026-09-27",
      "order": 5,
      "nvidia": {
        "minComputeCapability": "7.5"
      },
      "amd": {
        "gfxTargets": []
      },
      "intel": {
        "supported": false
      },
      "sources": [
        "https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/support-matrix.html"
      ],
      "notes": "NVIDIA's TensorRT 11.3.0 support matrix: \"TensorRT supports NVIDIA hardware with compute capability SM 7.5 or higher.\" NVIDIA-only.\n"
    },
    {
      "id": "vllm",
      "name": "vLLM",
      "version": "0.30.0",
      "verified": "2026-09-27",
      "order": 1,
      "nvidia": {
        "minComputeCapability": "7.5"
      },
      "amd": {
        "gfxTargets": [
          "gfx90a",
          "gfx942",
          "gfx950",
          "gfx1100",
          "gfx1101",
          "gfx1150",
          "gfx1151",
          "gfx1200",
          "gfx1201"
        ],
        "note": "ROCm 6.3 or newer (MI350 needs ROCm 7.0+, Ryzen AI MAX / AI 300 need 7.0.2+). Pre-built wheels for ROCm 7.0 and 7.2.1."
      },
      "intel": {
        "supported": true,
        "note": "XPU backend: Intel Data Center GPU and Arc GPUs."
      },
      "sources": [
        "https://github.com/vllm-project/vllm/blob/v0.30.0/docs/getting_started/installation/gpu.cuda.inc.md",
        "https://github.com/vllm-project/vllm/blob/v0.30.0/docs/getting_started/installation/gpu.rocm.inc.md",
        "https://github.com/vllm-project/vllm/blob/v0.30.0/docs/getting_started/installation/gpu.xpu.inc.md"
      ],
      "notes": "Requirements as stated in vLLM's own installation docs at the v0.30.0 tag. The NVIDIA minimum has been compute capability 7.5 since at least v0.20.0 (checked at v0.20.0, 0.24.0, 0.26.0-0.29.0), so the V100 (7.0) is not supported. The AMD list is the docs' \"GPU:\" requirement line: MI200s (gfx90a), MI300 (gfx942), MI350 (gfx950), Radeon RX 7900 series (gfx1100/1101), Radeon RX 9000 series (gfx1200/1201), Ryzen AI MAX / AI 300 (gfx1151/1150).\n"
    }
  ]
}