Function audio_frame_embedding_inputs

src/models/heartmula/generator.cpp:198–224

}

HeartMuLaFrameEmbeddingInputs audio_frame_embedding_inputs(
    const std::vector<int32_t> & frame_tokens,
    int64_t batch,
    int64_t codebook,
    const HeartMuLaConfig & config) {
    if (static_cast<int64_t>(frame_tokens.size()) != batch) {
        throw std::runtime_error("HeartMuLa audio embedding token batch mismatch");
    }
    HeartMuLaFrameEmbeddingInputs inputs;
    inputs.batch_size = batch;
    inputs.steps = 1;
    const size_t codebooks = static_cast<size_t>(config.audio_num_codebooks);
    inputs.audio_token_ids.assign(static_cast<size_t>(batch) * codebooks, 0);
    inputs.audio_mask.assign(static_cast<size_t>(batch) * codebooks, 0.0F);
    inputs.text_token_ids.assign(static_cast<size_t>(batch), 0);
    inputs.text_cond_mask.assign(static_cast<size_t>(batch), 0.0F);
    inputs.text_uncond_mask.assign(static_cast<size_t>(batch), 0.0F);
    for (int64_t b = 0; b < batch; ++b) {
        for (int64_t q = 0; q < config.audio_num_codebooks; ++q) {
            const size_t index = static_cast<size_t>(b * config.audio_num_codebooks + q);
            const int32_t raw = q == codebook ? frame_tokens[static_cast<size_t>(b)] : 0;
            inputs.audio_token_ids[index] = raw + static_cast<int32_t>(q * config.audio_vocab_size);
            inputs.audio_mask[index] = q == codebook ? 1.0F : 0.0F;
        }
    }
    return inputs;
}

HeartMuLaFrameEmbeddingInputs next_frame_embedding_inputs(
    const std::vector<int32_t> & frame_tokens,