Function load_weights

src/models/qwen3_tts/tokenizer_speech_decoder.cpp:349–487

    };
}

std::shared_ptr<const Qwen3SpeechTokenizerDecoderWeights> load_weights(
    const Qwen3TTSAssets & assets,
    ggml_backend_t backend,
    core::BackendType backend_type,
    assets::TensorStorageType linear_weight_storage_type,
    assets::TensorStorageType conv_weight_storage_type) {
    const auto & source = *assets.speech_tokenizer_weights;
    auto weights = std::make_shared<Qwen3SpeechTokenizerDecoderWeights>();
    weights->store = std::make_shared<core::BackendWeightStore>(
        backend,
        backend_type,
        "qwen3_tts.speech_tokenizer_decoder.weights",
        32ull * 1024ull * 1024ull);
    weights->config = load_decoder_config(assets);
    const auto & config = weights->config;
    const int64_t split_dim = config.codebook_dim / 2;

    for (int64_t layer = 0; layer < config.num_semantic_quantizers; ++layer) {
        const std::string prefix = "decoder.quantizer.rvq_first.vq.layers." + std::to_string(layer) + "._codebook.";
        weights->semantic_codebooks.push_back({normalized_codebook(source, prefix, config.codebook_size, split_dim)});
    }
    for (int64_t layer = 0; layer < config.num_quantizers - config.num_semantic_quantizers; ++layer) {
        const std::string prefix = "decoder.quantizer.rvq_rest.vq.layers." + std::to_string(layer) + "._codebook.";
        weights->acoustic_codebooks.push_back({normalized_codebook(source, prefix, config.codebook_size, split_dim)});
    }
    weights->semantic_output_proj = load_conv1x1_as_linear(
        *weights->store,
        source,
        "decoder.quantizer.rvq_first.output_proj",
        linear_weight_storage_type,
        split_dim,
        config.hidden_size);
    weights->acoustic_output_proj = load_conv1x1_as_linear(
        *weights->store,
        source,
        "decoder.quantizer.rvq_rest.output_proj",
        linear_weight_storage_type,
        split_dim,
        config.hidden_size);
    weights->pre_conv = load_conv(*weights->store, source, "decoder.pre_conv.conv", conv_weight_storage_type, config.hidden_size, config.latent_dim, 3);
    weights->transformer_input_proj = load_linear(
        *weights->store,
        source,
        "decoder.pre_transformer.input_proj",
        linear_weight_storage_type,
        config.latent_dim,
        config.hidden_size);
    for (int64_t layer = 0; layer < config.num_layers; ++layer) {
        const std::string prefix = "decoder.pre_transformer.layers." + std::to_string(layer);
        TransformerLayerWeights block;
        block.input_norm = load_rms_norm(source, prefix + ".input_layernorm", config.hidden_size, config.rms_norm_eps);
        block.post_norm = load_rms_norm(source, prefix + ".post_attention_layernorm", config.hidden_size, config.rms_norm_eps);
        block.attention.q = load_linear(*weights->store, source, prefix + ".self_attn.q_proj", linear_weight_storage_type, config.hidden_size, config.num_heads * config.head_dim, false);
        block.attention.k = load_linear(*weights->store, source, prefix + ".self_attn.k_proj", linear_weight_storage_type, config.hidden_size, config.num_kv_heads * config.head_dim, false);
        block.attention.v = load_linear(*weights->store, source, prefix + ".self_attn.v_proj", linear_weight_storage_type, config.hidden_size, config.num_kv_heads * config.head_dim, false);
        block.attention.o = load_linear(*weights->store, source, prefix + ".self_attn.o_proj", linear_weight_storage_type, config.num_heads * config.head_dim, config.hidden_size, false);
        block.mlp.gate = load_linear(*weights->store, source, prefix + ".mlp.gate_proj", linear_weight_storage_type, config.hidden_size, config.intermediate_size, false);