/** * export/gguf.ts — export weights to a GGUF v3 container (the llama.cpp format). * * Emits a spec-valid GGUF file: magic + version, KV metadata (architecture, dims, * expert counts…), aligned tensor-info table, then aligned tensor data. ggml * stores dimensions innermost-first, so row-major shapes are reversed here. * * NOTE: GGUF is a *container*. llama.cpp can only RUN a file whose * `general.architecture` it has a graph for; "evermind" is custom, so this file * is a valid, inspectable artifact (gguf tooling reads it) but won't execute in * stock llama.cpp until the architecture is upstreamed. It exists so the same * weights can travel through GGUF-native tooling. For an executable export use * {@link ./onnx} or {@link ./safetensors}. */ import type { EvermindLM } from "../lm/evermind_lm.js"; export interface GgufOptions { name?: string; /** Store tensors as float16 (ggml type F16). Default false (F32). */ fp16?: boolean; } /** Export a trained LM to GGUF bytes. */ export declare function exportGguf(lm: EvermindLM, opts?: GgufOptions): Uint8Array; //# sourceMappingURL=gguf.d.ts.map