/**
 * export/onnx.ts — export a trained {@link EvermindLM} to a runnable ONNX graph.
 *
 * Produces a standard causal-LM ONNX `ModelProto`:  input `input_ids` [batch,seq]
 * (int64) → output `logits` [batch,seq,vocab] (float). No KV cache — the model is
 * stateless and recomputes the full context each step, so a host (onnxruntime-web,
 * onnxruntime-node, or transformers.js driving the generation loop) re-runs the
 * whole sequence per token. The graph reproduces the CPU reference forward
 * EXACTLY:
 *
 *   • embedding  → Gather(emb, input_ids)
 *   • RMSNorm    → Mul/ReduceMean/Add/Sqrt/Div/Mul (eps = 1e-5)
 *   • causal     → Transpose → Conv(group=D, pads=[K-1,0], kernel reversed) → Transpose
 *     depthwise conv (so out[t] = Σⱼ ker[j]·x[t−j], matching the reference tap order)
 *   • MoE        → router MatMul → TopK → Softmax → ScatterElements (dense top-k mask)
 *                  → shared FFN + Σₑ combineₑ · expertₑ(x)   (numerically identical to
 *                  the sparse reference: non-selected experts get combine 0)
 *   • tied head  → MatMul(x, embᵀ)
 *
 * Opset 18 / IR v8. Weights are emitted as raw little-endian f32 initializers that
 * are byte-identical to {@link namedTensors} (matmul weights are transposed by a
 * graph node at runtime, not duplicated on disk).
 *
 * Zero dependencies — the protobuf is hand-encoded by {@link ./protobuf}.
 */
import type { EvermindLM } from "../lm/evermind_lm.js";
/** Build the ONNX bytes for a trained LM. */
export declare function exportOnnx(lm: EvermindLM, opts?: {
    producerName?: string;
    producerVersion?: string;
}): Uint8Array;
//# sourceMappingURL=onnx.d.ts.map