Function prompt_embedding_inputs

src/models/heartmula/generator.cpp:150–196

}

HeartMuLaFrameEmbeddingInputs prompt_embedding_inputs(
    const HeartMuLaPromptEncoding & encoding,
    const HeartMuLaConfig & config,
    bool use_cfg) {
    HeartMuLaFrameEmbeddingInputs inputs;
    inputs.batch_size = encoding.batch_size;
    inputs.steps = encoding.prompt_len;
    const size_t batch = static_cast<size_t>(encoding.batch_size);
    const size_t steps = static_cast<size_t>(encoding.prompt_len);
    const size_t lanes = static_cast<size_t>(encoding.parallel_number);
    const size_t codebooks = static_cast<size_t>(config.audio_num_codebooks);
    inputs.audio_token_ids.assign(batch * steps * codebooks, 0);
    inputs.text_token_ids.assign(batch * steps, 0);
    inputs.audio_mask.assign(batch * steps * codebooks, 0.0F);
    inputs.text_cond_mask.assign(batch * steps, 0.0F);
    inputs.text_uncond_mask.assign(batch * steps, 0.0F);
    const size_t actual_batch = use_cfg ? batch / 2 : batch;
    for (size_t b = 0; b < batch; ++b) {
        const bool uncond = use_cfg && b >= actual_batch;
        for (size_t t = 0; t < steps; ++t) {
            for (size_t q = 0; q < codebooks; ++q) {
                const size_t src = prompt_flat(b, t, q, steps, lanes);
                const size_t dst = (b * steps + t) * codebooks + q;
                inputs.audio_token_ids[dst] = static_cast<int32_t>(
                    encoding.tokens[src] + static_cast<int64_t>(q) * config.audio_vocab_size);
                inputs.audio_mask[dst] = encoding.tokens_mask[src] != 0U ? 1.0F : 0.0F;
            }
            const size_t text_src = prompt_flat(b, t, lanes - 1, steps, lanes);
            inputs.text_token_ids[b * steps + t] = static_cast<int32_t>(encoding.tokens[text_src]);
            const float text_mask = encoding.tokens_mask[text_src] != 0U ? 1.0F : 0.0F;
            inputs.text_cond_mask[b * steps + t] = uncond ? 0.0F : text_mask;
            inputs.text_uncond_mask[b * steps + t] = uncond ? text_mask : 0.0F;
        }
    }
    inputs.apply_muq = true;
    inputs.muq_row = encoding.muq_idx.empty() ? 0 : encoding.muq_idx.front();
    inputs.muq_embed = encoding.muq_embed;
    inputs.muq_cond_mask.assign(batch, 1.0F);
    inputs.muq_uncond_mask.assign(batch, 0.0F);
    if (use_cfg) {
        for (size_t b = actual_batch; b < batch; ++b) {
            inputs.muq_cond_mask[b] = 0.0F;
            inputs.muq_uncond_mask[b] = 1.0F;
        }
    }
    return inputs;
}

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