Edgewise API
    Preparing search index...

    Interface Variant

    interface Variant {
        bytes?: number;
        devices: Device[];
        dtype: Dtype | Record<string, Dtype>;
        files?: {
            bucket: string;
            names: string[];
            path: string;
            sha256?: Record<string, string>;
        };
        memory?: number
        | Partial<Record<Device, number>>;
        shaderF16?: boolean;
    }
    Index
    bytes?: number

    Approximate download size in bytes.

    devices: Device[]

    Devices this variant is known to run correctly on.

    dtype: Dtype | Record<string, Dtype>

    One dtype for the whole model, or one per ONNX component.

    files?: {
        bucket: string;
        names: string[];
        path: string;
        sha256?: Record<string, string>;
    }

    ONNX files for this variant that come from an Edgewise bucket instead of the model's repo, for example a re-export that runs on ONNX Runtime Web's WebAssembly build. Other files (config, tokenizer) still come from the repo.

    memory?: number | Partial<Record<Device, number>>

    Measured memory in bytes once loaded, for all devices or per device. Without it Edgewise estimates from bytes, the dtype and the device. Used by the memory budget and low-memory variant choice.

    shaderF16?: boolean

    Requires the WebGPU shader-f16 feature.