Function compute_s3_flow_decoder_forward

src/models/chatterbox/s3gen_flow.cpp:1875–1894

        load_decoder_conv1d(*weights->store, source, "flow.decoder.estimator.down_blocks.0.2", 256, 256, 3, 1, true, weight_storage_type);

    weights->mid_blocks.resize(12);
    for (int block_index = 0; block_index < 12; ++block_index) {
        weights->mid_blocks[static_cast<size_t>(block_index)].resnet =
            load_resnet("flow.decoder.estimator.mid_blocks." + std::to_string(block_index) + ".0", 256, 256);
        for (int i = 0; i < 4; ++i) {
            weights->mid_blocks[static_cast<size_t>(block_index)].transformers.push_back(
                load_transformer("flow.decoder.estimator.mid_blocks." + std::to_string(block_index) + ".1." + std::to_string(i)));
        }
    }

    weights->up_blocks.resize(1);
    weights->up_blocks[0].resnet = load_resnet("flow.decoder.estimator.up_blocks.0.0", 512, 256);
    for (int i = 0; i < 4; ++i) {
        weights->up_blocks[0].transformers.push_back(load_transformer("flow.decoder.estimator.up_blocks.0.1." + std::to_string(i)));
    }
    weights->up_blocks[0].upsample =
        load_decoder_conv1d(*weights->store, source, "flow.decoder.estimator.up_blocks.0.2", 256, 256, 3, 1, true, weight_storage_type);

    weights->final_block = load_causal_block("flow.decoder.estimator.final_block", 256, 256);
    weights->final_proj = load_decoder_conv1d(*weights->store, source, "flow.decoder.estimator.final_proj", 80, 256, 1, 1, true, weight_storage_type);

    if (source.has_tensor("flow.decoder.estimator.time_embed_mixer.weight")) {
        weights->meanflow = true;