Method backbone_prefill_embeddings

src/models/heartmula/mula.cpp:1893–1912

}

HeartMuLaBackbonePrefillOutput HeartMuLaWeightsRuntime::backbone_prefill_embeddings(
    const std::vector<float> & embeddings,
    int64_t batch_size,
    int64_t steps) const {
    const auto & config = assets_->mula_config.backbone;
    if (batch_size <= 0 || steps <= 0) {
        throw std::runtime_error("HeartMuLa backbone prefill requires positive batch and step counts");
    }
    if (steps > config.max_seq_len) {
        throw std::runtime_error("HeartMuLa backbone prefill exceeds model context length");
    }
    if (static_cast<int64_t>(embeddings.size()) != batch_size * steps * config.embed_dim) {
        throw std::runtime_error("HeartMuLa backbone prefill embedding payload size mismatch");
    }
    if (backbone_prefill_graph_ == nullptr ||
        !backbone_prefill_graph_->matches(*this, batch_size, steps)) {
        backbone_prefill_graph_ = std::make_unique<HeartMuLaBackbonePrefillGraph>(*this, batch_size, steps);
    }
    return backbone_prefill_graph_->run(embeddings);
}

void HeartMuLaWeightsRuntime::reset_backbone_cached_state(
    HeartMuLaBackboneCachedState & state,