Skip to content

    Hardware profiles

    The server is published as two images. Match the image to your hardware.

    ghcr.io/forgeguard-ai/kokoro-server:latest (alias kokoro-server-cu128).

    • Supported. Built on CUDA 12.8 with cuDNN. The cu128 PyTorch wheels carry kernels for compute capabilities sm_86 through sm_120 — that is NVIDIA RTX 3000 (Ampere) through RTX 5000 (Blackwell) in a single image, plus datacenter GPUs in that range.
    • Runs as a container with --gpus all. Also runs on CPU with USE_GPU=false.

    ghcr.io/forgeguard-ai/kokoro-server-jetson:latest.

    • Supported on Jetson Orin (compute capability sm_87) running JetPack 6 / L4T r36 (CUDA 12.6, cuDNN 9.3). Confirm your L4T version with cat /etc/nv_tegra_release.

    • Launch with the NVIDIA runtime:

      Terminal window
      docker run -d --name kokoro --runtime nvidia -p 8880:8880 \
      ghcr.io/forgeguard-ai/kokoro-server-jetson:latest
    • The Jetson image pins its CUDA math libraries to the JetPack 6 toolkit to avoid a cuBLAS version conflict; use the Jetson image (not the cu128 image) on these devices.

    • Supported, reduced throughput. The cu128 image runs on CPU with USE_GPU=false. Expect substantially slower synthesis than on a GPU. Useful for CI, development, and low-volume use.
    • Planned, not currently supported. Tracked on the roadmap; do not treat these as available today.
    TargetStatusNotes
    NVIDIA RTX 3000 → 5000 (x86_64, CUDA cu128)SupportedSingle cu128 image (sm_86–sm_120).
    NVIDIA datacenter GPUs in that rangeSupportedSame cu128 image.
    NVIDIA Jetson Orin (arm64)SupportedJetPack 6 / L4T r36; use the Jetson image.
    CPU (x86_64)SupportedUSE_GPU=false; reduced throughput.
    Apple Silicon (MPS)ExperimentalAuto-detected for local runs; not a published container target.
    AMD (ROCm), IntelPlannedNot currently supported.

    See Compatibility for API and version details.