Function load_decoder_config

src/models/qwen3_tts/tokenizer_speech_decoder.cpp:216–238

    for (int64_t code = 0; code < size; ++code) {
        const float denom = std::max(cluster_usage[static_cast<size_t>(code)], kCodebookEps);
        for (int64_t col = 0; col < dim; ++col) {
            const size_t offset = static_cast<size_t>(code * dim + col);
            embedding[offset] = embedding_sum[offset] / denom;
        }
    }
    return embedding;
}

DecoderConfig load_decoder_config(const Qwen3TTSAssets & assets) {
    const auto root = assets.resources.parse_json("speech_tokenizer_config");
    const auto & decoder = root.require("decoder_config");
    DecoderConfig config;
    config.codebook_size = json::require_i64(decoder, "codebook_size");
    config.codebook_dim = json::require_i64(decoder, "codebook_dim");
    config.latent_dim = json::require_i64(decoder, "latent_dim");
    config.hidden_size = json::require_i64(decoder, "hidden_size");
    config.intermediate_size = json::require_i64(decoder, "intermediate_size");
    config.decoder_dim = json::require_i64(decoder, "decoder_dim");
    config.num_heads = json::require_i64(decoder, "num_attention_heads");
    config.num_kv_heads = json::require_i64(decoder, "num_key_value_heads");
    config.num_layers = json::require_i64(decoder, "num_hidden_layers");
    config.num_quantizers = json::require_i64(decoder, "num_quantizers");
    config.num_semantic_quantizers = json::require_i64(decoder, "num_semantic_quantizers");
    config.sliding_window = json::require_i64(decoder, "sliding_window");
    config.head_dim = json::require_i64(decoder, "head_dim");
    config.rope_theta = json::optional_f32(decoder, "rope_theta", config.rope_theta);