Method decoder_prefill_embeddings

src/models/heartmula/mula.cpp:1980–2003

}

HeartMuLaDecoderPrefillOutput HeartMuLaWeightsRuntime::decoder_prefill_embeddings(
    const std::vector<float> & embeddings,
    int64_t batch_size,
    int64_t steps,
    int64_t codebook_index) const {
    const auto & full_config = assets_->mula_config;
    if (batch_size <= 0 || steps <= 0) {
        throw std::runtime_error("HeartMuLa decoder prefill requires positive batch and step counts");
    }
    if (steps > full_config.decoder.max_seq_len) {
        throw std::runtime_error("HeartMuLa decoder prefill exceeds model context length");
    }
    if (codebook_index < 0 || codebook_index >= full_config.audio_num_codebooks - 1) {
        throw std::runtime_error("HeartMuLa decoder prefill codebook index is out of range");
    }
    if (static_cast<int64_t>(embeddings.size()) != batch_size * steps * full_config.backbone.embed_dim) {
        throw std::runtime_error("HeartMuLa decoder prefill embedding payload size mismatch");
    }
    if (decoder_prefill_graph_ == nullptr ||
        !decoder_prefill_graph_->matches(*this, batch_size, steps)) {
        decoder_prefill_graph_ = std::make_unique<HeartMuLaDecoderPrefillGraph>(*this, batch_size, steps);
    }
    return decoder_prefill_graph_->run(embeddings, codebook_index);
}

void HeartMuLaWeightsRuntime::reset_decoder_cached_state(
    HeartMuLaDecoderCachedState & state,