Skip to content

    Development documentation — this describes the current development branch. Documented behavior may not exist in the latest release.

    Compatibility

    Compatibility

    Section titled “Compatibility”

    Hardware and platforms

    Section titled “Hardware and platforms”
    TargetStatusNotes
    NVIDIA CUDA (x86_64)Supportedcu128 image; RTX 3000 Ampere → RTX 5000 Blackwell; bakes large-v3
    NVIDIA Jetson Orin (arm64)SupportedJetPack 6 image; bakes small with int8_float16
    CPUSupportedAuto-downgrades to int8; reduced throughput
    AMD (ROCm)Not currently supportedPlanned; see project status
    IntelNot currently supportedPlanned; see project status
    Apple SiliconNot supportedNo published image

    Container engines: Docker with the NVIDIA Container Toolkit, and Kubernetes with the NVIDIA device plugin. There is no supported bare-metal install path.

    API compatibility vs. OpenAI parity

    Section titled “API compatibility vs. OpenAI parity”

    The API is OpenAI-compatible, meaning existing OpenAI audio SDKs work by pointing base_url at this server. It is not a byte-for-byte reimplementation of the OpenAI service. Practical differences:

    • model is ignored. The loaded Whisper model is used regardless of the model field; /v1/models reports whisper-1 plus the loaded size.
    • temperature is accepted but ignored.
    • Response formats. text, json, verbose_json, srt, and vtt are supported. In verbose_json, word timings are a flattened top-level words[] array (matching OpenAI), not nested in segments.
    • Endpoints implemented. /v1/audio/transcriptions, /v1/audio/translations, and /v1/models. Other OpenAI endpoints (chat, embeddings, and so on) are not part of this server.
    • Authentication. Optional bearer API_KEY following the OpenAI Authorization: Bearer convention, but off by default.
    • Operational routes. /health, /ready, /system, and /api/model/* are ForgeGuard additions, not part of the OpenAI API.

    Whisper models

    Section titled “Whisper models”

    Transcription quality, language coverage, and translation behavior come from the underlying Whisper models. Available presets range from tiny to large-v3 (see Model selection). Language support follows faster-whisper’s supported language set; an unsupported language code returns 400.

    Versioning

    Section titled “Versioning”

    The server, container images, and Helm chart share a version (currently 1.1.0). See Upgrades for the tagging and release policy.