Function effective_weight_norm_dim0

src/models/heartmula/codec.cpp:67–98

    int64_t hidden_dim = 4 * dim;
    hidden_dim = 2 * hidden_dim / 3;
    return multiple_of * ((hidden_dim + multiple_of - 1) / multiple_of);
}

std::vector<float> effective_weight_norm_dim0(
    const assets::TensorSource & source,
    const std::string & prefix,
    const std::vector<int64_t> & shape) {
    if (shape.empty()) {
        throw std::runtime_error("HeartCodec weight-norm tensor shape is empty");
    }
    const auto g = source.require_f32(prefix + ".parametrizations.weight.original0", {shape[0], 1, 1});
    const auto v = source.require_f32(prefix + ".parametrizations.weight.original1", shape);
    std::vector<float> out(v.size(), 0.0F);
    int64_t inner = 1;
    for (size_t axis = 1; axis < shape.size(); ++axis) {
        inner *= shape[axis];
    }
    for (int64_t row = 0; row < shape[0]; ++row) {
        double sum = 0.0;
        const size_t base = static_cast<size_t>(row * inner);
        for (int64_t index = 0; index < inner; ++index) {
            const float value = v[base + static_cast<size_t>(index)];
            sum += static_cast<double>(value) * static_cast<double>(value);
        }
        const double norm = std::sqrt(sum);
        if (norm == 0.0) {
            throw std::runtime_error("HeartCodec weight-norm tensor has zero norm: " + prefix);
        }
        const float scale = static_cast<float>(static_cast<double>(g[static_cast<size_t>(row)]) / norm);
        for (int64_t index = 0; index < inner; ++index) {
            out[base + static_cast<size_t>(index)] = v[base + static_cast<size_t>(index)] * scale;
        }
    }
    return out;
}