Function parse_codec_config

src/models/heartmula/assets.cpp:168–246

    engine::io::require_all_positive(config.downsample_factors, "downsample_factors");
    engine::io::require_all_positive(config.downsample_kernel_sizes, "downsample_kernel_sizes");
    engine::io::require_all_positive(config.upsample_factors, "upsample_factors");
    engine::io::require_all_positive(config.upsample_kernel_sizes, "upsample_kernel_sizes");
    if (config.downsample_factors.size() != config.downsample_kernel_sizes.size() ||
        config.upsample_factors.size() != config.upsample_kernel_sizes.size()) {
        throw std::runtime_error("HeartCodec config sampling factor/kernel counts must match");
    }
    if (config.norm_type != "ada_norm_single") {
        throw std::runtime_error("HeartCodec config norm_type must be ada_norm_single");
    }
    if (!config.causal) {
        throw std::runtime_error("HeartCodec config must be causal");
    }
    return config;
}

void validate_mula_weight_anchors(const HeartMuLaAssets & assets) {
    const auto & config = assets.mula_config;
    const auto & weights = *assets.mula_weights;
    const auto & backbone = config.backbone;
    const auto & decoder = config.decoder;
    assets::require_tensor_shape(weights, "text_embeddings.weight", {config.text_vocab_size, backbone.embed_dim});
    assets::require_tensor_shape(
        weights,
        "audio_embeddings.weight",
        {config.audio_vocab_size * config.audio_num_codebooks, backbone.embed_dim});
    assets::require_tensor_shape(weights, "unconditional_text_embedding.weight", {1, backbone.embed_dim});
    assets::require_tensor_shape(weights, "projection.weight", {decoder.embed_dim, backbone.embed_dim});
    assets::require_tensor_shape(weights, "codebook0_head.weight", {config.audio_vocab_size, backbone.embed_dim});
    assets::require_tensor_shape(
        weights,
        "audio_head",
        {config.audio_num_codebooks - 1, decoder.embed_dim, config.audio_vocab_size});
    assets::require_tensor_shape(weights, "muq_linear.weight", {backbone.embed_dim, config.muq_dim});
    assets::require_tensor_shape(weights, "muq_linear.bias", {backbone.embed_dim});
    assets::require_tensor_shape(weights, "backbone.norm.scale", {backbone.embed_dim});
    assets::require_tensor_shape(weights, "decoder.norm.scale", {decoder.embed_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.attn.q_proj.weight", {backbone.embed_dim, backbone.embed_dim});
    assets::require_tensor_shape(
        weights,
        "backbone.layers.0.attn.k_proj.weight",
        {backbone.num_kv_heads * backbone.head_dim, backbone.embed_dim});
    assets::require_tensor_shape(
        weights,
        "backbone.layers.0.attn.v_proj.weight",
        {backbone.num_kv_heads * backbone.head_dim, backbone.embed_dim});
    assets::require_tensor_shape(
        weights,
        "backbone.layers.0.attn.output_proj.weight",
        {backbone.embed_dim, backbone.embed_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.mlp.w1.weight", {backbone.intermediate_dim, backbone.embed_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.mlp.w2.weight", {backbone.embed_dim, backbone.intermediate_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.mlp.w3.weight", {backbone.intermediate_dim, backbone.embed_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.sa_norm.scale", {backbone.embed_dim});
    assets::require_tensor_shape(weights, "backbone.layers.0.mlp_norm.scale", {backbone.embed_dim});
    assets::require_tensor_shape(weights, "decoder.layers.0.attn.q_proj.weight", {decoder.embed_dim, decoder.embed_dim});
    assets::require_tensor_shape(
        weights,
        "decoder.layers.0.attn.k_proj.weight",