import type { ResolvedDevice } from "../../device.js";
/**
 * The four ONNX sessions a Chatterbox model is split into, named by their
 * *file* names: Transformers.js resolves per-session `dtype` entries against
 * the ONNX file name (`language_model`), not the session key (`model`) — a
 * mismatched key is silently ignored and the device default (fp32 on WebGPU,
 * 2 GB for the language model) loads instead.
 */
export type ChatterboxSession = "embed_tokens" | "speech_encoder" | "language_model" | "conditional_decoder";
/** Per-session quantization, as accepted by `from_pretrained({ dtype })`. */
export type DtypeConfig = Record<ChatterboxSession, string>;
/** One attempt at loading the model: a device plus a quantization choice. */
export interface LoadPlan {
    readonly device: ResolvedDevice;
    readonly dtype: DtypeConfig;
}
/**
 * Build the ordered list of load attempts for a device.
 *
 * fp16 support varies by GPU, so a WebGPU request degrades to integer-only
 * weights before giving up on the GPU entirely and landing on WASM. A WASM
 * request has nothing to fall back to and yields a single plan.
 *
 * Pass `fp16: false` when the adapter lacks `shader-f16`: an f16 plan on such
 * a device *loads* fine and only fails at the first inference, which the
 * load-time fallback can no longer catch.
 */
export declare function buildLoadPlans(device: ResolvedDevice, overrides?: Partial<DtypeConfig>, fp16?: boolean): LoadPlan[];
//# sourceMappingURL=dtype-plan.d.ts.map