voice_audio_multi/include/engine/framework/modules/weight_binding.h

#pragma once

#include "engine/framework/assets/tensor_source.h"
#include "engine/framework/core/module.h"
#include "engine/framework/modules/conditioning_modules.h"
#include "engine/framework/modules/conv_modules.h"
#include "engine/framework/modules/linear_module.h"
#include "engine/framework/modules/norm_modules.h"
#include "engine/framework/modules/recurrent_modules.h"
#include "engine/framework/modules/streaming_conv_modules.h"

#include <cstddef>
#include <cstdint>
#include <cmath>
#include <optional>
#include <stdexcept>
#include <string>
#include <utility>
#include <vector>

namespace engine::modules::binding {

template <typename Store>
ChannelAffineWeights batch_norm_eval_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    int64_t channels,
    float eps,
    assets::TensorStorageType storage_type = assets::TensorStorageType::F32) {
    const auto gamma = source.require_f32(prefix + ".weight", {channels});
    const auto beta = source.require_f32(prefix + ".bias", {channels});
    const auto mean = source.require_f32(prefix + ".running_mean", {channels});
    const auto variance = source.require_f32(prefix + ".running_var", {channels});
    std::vector<float> scale(static_cast<size_t>(channels)), bias(static_cast<size_t>(channels));
    for (size_t i = 0; i < scale.size(); ++i) {
        scale[i] = gamma[i] / std::sqrt(variance[i] + eps);
        bias[i] = beta[i] - mean[i] * scale[i];
    }
    const auto shape = core::TensorShape::from_dims({channels});
    return {store.make_from_f32(shape, storage_type, std::move(scale)),
        store.make_from_f32(shape, storage_type, std::move(bias))};
}

inline LinearConfig linear_config(
    int64_t in_features,
    int64_t out_features,
    bool use_bias) {
    return {in_features, out_features, use_bias, GGML_PREC_DEFAULT};
}

template <typename Store, typename Values>
core::TensorValue tensor(Store & store, const core::TensorShape & shape, const Values & values) {
    return store.make_f32(shape, values);
}

template <typename Store, typename TensorData>
core::TensorValue tensor_data(Store & store, const TensorData & data) {
    return store.make_f32(data.shape, data.values);
}

template <typename Store>
core::TensorValue tensor_data(Store &, const core::TensorValue & data) {
    return data;
}

template <typename Store>
core::TensorValue tensor_data(Store & store, const assets::TensorData & data) {
    return store.make_tensor(data.shape, data.type, data.bytes.data(), data.bytes.size());
}

inline std::vector<int64_t> tensor_shape_from_source(
    const assets::TensorSource & source,
    const std::string & name) {
    return source.require_metadata(name).shape;
}

template <typename Store>
core::TensorValue tensor_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & name,
    assets::TensorStorageType storage_type) {
    return store.load_tensor(source, name, storage_type, tensor_shape_from_source(source, name));
}

template <typename Store>
core::TensorValue f32_tensor_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & name) {
    return store.load_f32_tensor(source, name, tensor_shape_from_source(source, name));
}

template <typename Store>
core::TensorValue weight_norm_tensor_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    const core::TensorShape & shape) {
    if (shape.rank == 0 || shape.dims[0] <= 0) {
        throw std::runtime_error(prefix + " weight-norm tensor shape is empty");
    }
    const std::string weight_v_name = source.has_tensor(prefix + ".weight_v")
        ? prefix + ".weight_v"
        : prefix + ".parametrizations.weight.original1";
    const std::string weight_g_name = source.has_tensor(prefix + ".weight_g")
        ? prefix + ".weight_g"
        : prefix + ".parametrizations.weight.original0";

    std::vector<int64_t> weight_shape;
    weight_shape.reserve(shape.rank);
    int64_t inner_size = 1;
    for (size_t dim = 0; dim < shape.rank; ++dim) {
        if (shape.dims[dim] <= 0) {
            throw std::runtime_error(prefix + " weight-norm tensor shape contains non-positive dimension");
        }
        weight_shape.push_back(shape.dims[dim]);
        if (dim > 0) {
            inner_size *= shape.dims[dim];
        }
    }
    const int64_t leading = shape.dims[0];
    const auto weight_v = source.require_f32(weight_v_name, weight_shape);
    const auto weight_g_meta = source.require_metadata(weight_g_name);
    int64_t weight_g_elements = 1;
    for (const int64_t dim : weight_g_meta.shape) {
        weight_g_elements *= dim;
    }
    if (weight_g_elements != leading) {
        throw std::runtime_error(prefix + " weight-norm g shape mismatch");
    }
    const auto weight_g = source.require_f32(weight_g_name, weight_g_meta.shape);

    std::vector<float> folded(weight_v.size(), 0.0F);
    for (int64_t row = 0; row < leading; ++row) {
        const size_t base = static_cast<size_t>(row * inner_size);
        double norm_sq = 0.0;
        for (int64_t index = 0; index < inner_size; ++index) {
            const float value = weight_v[base + static_cast<size_t>(index)];
            norm_sq += static_cast<double>(value) * static_cast<double>(value);
        }
        const float scale = weight_g[static_cast<size_t>(row)] / static_cast<float>(std::sqrt(norm_sq));
        for (int64_t index = 0; index < inner_size; ++index) {
            folded[base + static_cast<size_t>(index)] = weight_v[base + static_cast<size_t>(index)] * scale;
        }
    }
    return store.make_from_f32(shape, storage_type, std::move(folded));
}

template <typename Store, typename TensorData>
LinearWeights linear_data(Store & store, const TensorData & weight) {
    return {tensor_data(store, weight), std::nullopt};
}

template <typename Store, typename WeightData, typename BiasData>
LinearWeights linear_data(Store & store, const WeightData & weight, const BiasData & bias) {
    return {tensor_data(store, weight), tensor_data(store, bias)};
}

template <typename Store, typename TensorData>
LinearWeights linear_data(Store & store, const TensorData & weight, const std::optional<TensorData> & bias) {
    return bias.has_value() ? linear_data(store, weight, *bias) : linear_data(store, weight);
}

template <typename Store, typename WeightValues>
LinearWeights linear(
    Store & store,
    int64_t in_features,
    int64_t out_features,
    const WeightValues & weight) {
    return {
        tensor(store, core::TensorShape::from_dims({out_features, in_features}), weight),
        std::nullopt,
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
LinearWeights linear(
    Store & store,
    int64_t in_features,
    int64_t out_features,
    const WeightValues & weight,
    const BiasValues & bias) {
    return {
        tensor(store, core::TensorShape::from_dims({out_features, in_features}), weight),
        tensor(store, core::TensorShape::from_dims({out_features}), bias),
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
LinearWeights linear(
    Store & store,
    int64_t in_features,
    int64_t out_features,
    bool use_bias,
    const WeightValues & weight,
    const BiasValues & bias) {
    return use_bias
        ? linear(store, in_features, out_features, weight, bias)
        : linear(store, in_features, out_features, weight);
}

template <typename Store, typename TensorData>
NormWeights norm_data(Store & store, const TensorData & weight) {
    return {tensor_data(store, weight), std::nullopt};
}

template <typename Store, typename TensorData>
NormWeights norm_data(Store & store, const TensorData & weight, const TensorData & bias) {
    return {tensor_data(store, weight), tensor_data(store, bias)};
}

template <typename Store, typename WeightValues>
NormWeights norm(
    Store & store,
    const WeightValues & weight) {
    return {
        tensor(store, core::TensorShape::from_dims({static_cast<int64_t>(weight.size())}), weight),
        std::nullopt,
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
NormWeights norm(
    Store & store,
    const WeightValues & weight,
    const BiasValues & bias) {
    return {
        tensor(store, core::TensorShape::from_dims({static_cast<int64_t>(weight.size())}), weight),
        tensor(store, core::TensorShape::from_dims({static_cast<int64_t>(bias.size())}), bias),
    };
}

template <typename Store, typename TensorData>
Conv1dWeights conv1d_data(Store & store, const TensorData & weight) {
    return {tensor_data(store, weight), std::nullopt};
}

template <typename Store, typename WeightData, typename BiasData>
Conv1dWeights conv1d_data(Store & store, const WeightData & weight, const BiasData & bias) {
    return {tensor_data(store, weight), tensor_data(store, bias)};
}

template <typename Store, typename TensorData>
Conv1dWeights conv1d_data(Store & store, const TensorData & weight, const std::optional<TensorData> & bias) {
    return bias.has_value() ? conv1d_data(store, weight, *bias) : conv1d_data(store, weight);
}

template <typename Store, typename TensorData>
Conv2dWeights conv2d_data(Store & store, const TensorData & weight) {
    return {tensor_data(store, weight), std::nullopt};
}

template <typename Store, typename WeightData, typename BiasData>
Conv2dWeights conv2d_data(Store & store, const WeightData & weight, const BiasData & bias) {
    return {tensor_data(store, weight), tensor_data(store, bias)};
}

template <typename Store, typename TensorData>
Conv2dWeights conv2d_data(Store & store, const TensorData & weight, const std::optional<TensorData> & bias) {
    return bias.has_value() ? conv2d_data(store, weight, *bias) : conv2d_data(store, weight);
}

template <typename Store, typename WeightValues>
Conv1dWeights conv1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    int64_t groups,
    const WeightValues & weight) {
    return {
        tensor(store, core::TensorShape::from_dims({out_channels, in_channels / groups, kernel}), weight),
        std::nullopt,
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
Conv1dWeights conv1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    int64_t groups,
    const WeightValues & weight,
    const BiasValues & bias) {
    return {
        tensor(store, core::TensorShape::from_dims({out_channels, in_channels / groups, kernel}), weight),
        tensor(store, core::TensorShape::from_dims({out_channels}), bias),
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
Conv1dWeights conv1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    int64_t groups,
    bool use_bias,
    const WeightValues & weight,
    const BiasValues & bias) {
    return use_bias
        ? conv1d(store, in_channels, out_channels, kernel, groups, weight, bias)
        : conv1d(store, in_channels, out_channels, kernel, groups, weight);
}

template <typename Store, typename WeightValues>
ConvTranspose1dWeights conv_transpose1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    const WeightValues & weight) {
    return {
        tensor(store, core::TensorShape::from_dims({in_channels, out_channels, kernel}), weight),
        std::nullopt,
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
ConvTranspose1dWeights conv_transpose1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    const WeightValues & weight,
    const BiasValues & bias) {
    return {
        tensor(store, core::TensorShape::from_dims({in_channels, out_channels, kernel}), weight),
        tensor(store, core::TensorShape::from_dims({out_channels}), bias),
    };
}

template <typename Store, typename WeightValues, typename BiasValues>
ConvTranspose1dWeights conv_transpose1d(
    Store & store,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel,
    bool use_bias,
    const WeightValues & weight,
    const BiasValues & bias) {
    return use_bias
        ? conv_transpose1d(store, in_channels, out_channels, kernel, weight, bias)
        : conv_transpose1d(store, in_channels, out_channels, kernel, weight);
}

template <typename Store, typename TensorData>
ConvTranspose1dWeights conv_transpose1d_data(Store & store, const TensorData & weight) {
    return {tensor_data(store, weight), std::nullopt};
}

template <typename Store, typename WeightData, typename BiasData>
ConvTranspose1dWeights conv_transpose1d_data(Store & store, const WeightData & weight, const BiasData & bias) {
    return {tensor_data(store, weight), tensor_data(store, bias)};
}

template <typename Store, typename TensorData>
ConvTranspose1dWeights conv_transpose1d_data(Store & store, const TensorData & weight, const std::optional<TensorData> & bias) {
    return bias.has_value() ? conv_transpose1d_data(store, weight, *bias) : conv_transpose1d_data(store, weight);
}

template <typename Store>
LinearWeights linear_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_features,
    int64_t in_features,
    bool use_bias) {
    LinearWeights weights;
    weights.weight = store.load_tensor(source, prefix + ".weight", storage_type, {out_features, in_features});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_features});
    }
    return weights;
}

template <typename Store>
LinearWeights linear_from_transposed_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name,
    assets::TensorStorageType storage_type,
    int64_t in_features,
    int64_t out_features);

template <typename Store>
LinearWeights hf_conv1d_linear_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t in_features,
    int64_t out_features,
    bool use_bias) {
    return linear_from_transposed_named_source(
        store,
        source,
        prefix + ".weight",
        use_bias ? std::optional<std::string>{prefix + ".bias"} : std::nullopt,
        storage_type,
        in_features,
        out_features);
}

template <typename Store>
LinearWeights linear_from_transposed_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name,
    assets::TensorStorageType storage_type,
    int64_t in_features,
    int64_t out_features) {
    const auto source_weight = source.require_f32(weight_name, {in_features, out_features});
    std::vector<float> transposed(static_cast<std::size_t>(out_features * in_features));
    for (int64_t in = 0; in < in_features; ++in) {
        for (int64_t out = 0; out < out_features; ++out) {
            transposed[static_cast<std::size_t>(out * in_features + in)] =
                source_weight[static_cast<std::size_t>(in * out_features + out)];
        }
    }
    LinearWeights weights;
    weights.weight = store.make_from_f32(
        core::TensorShape::from_dims({out_features, in_features}),
        storage_type,
        std::move(transposed));
    if (bias_name.has_value()) {
        weights.bias = store.load_f32_tensor(source, *bias_name, {out_features});
    }
    return weights;
}

template <typename Store>
NormWeights norm_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    int64_t hidden_size) {
    return {
        store.load_f32_tensor(source, prefix + ".weight", {hidden_size}),
        store.load_f32_tensor(source, prefix + ".bias", {hidden_size}),
    };
}

template <typename Store>
NormWeights norm_weight_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    int64_t hidden_size) {
    return {
        store.load_f32_tensor(source, prefix + ".weight", {hidden_size}),
        std::nullopt,
    };
}

template <typename Store>
Conv1dWeights conv1d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_size,
    bool use_bias) {
    Conv1dWeights weights;
    weights.weight = store.load_tensor(source, prefix + ".weight", storage_type, {out_channels, in_channels, kernel_size});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
Conv1dWeights weight_norm_conv1d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_size,
    bool use_bias) {
    Conv1dWeights weights;
    weights.weight = weight_norm_tensor_from_source(
        store,
        source,
        prefix,
        storage_type,
        core::TensorShape::from_dims({out_channels, in_channels, kernel_size}));
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
Conv1dWeights conv1d_from_source_resolving_weight_norm(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_size,
    bool use_bias) {
    if (source.has_tensor(prefix + ".weight")) {
        return conv1d_from_source(
            store,
            source,
            prefix,
            storage_type,
            out_channels,
            in_channels,
            kernel_size,
            use_bias);
    }
    return weight_norm_conv1d_from_source(
        store,
        source,
        prefix,
        storage_type,
        out_channels,
        in_channels,
        kernel_size,
        use_bias);
}

template <typename Store>
Conv2dWeights conv2d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_height,
    int64_t kernel_width,
    bool use_bias) {
    Conv2dWeights weights;
    weights.weight = store.load_tensor(source, prefix + ".weight", storage_type, {out_channels, in_channels, kernel_height, kernel_width});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
Conv3dWeights conv3d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_depth,
    int64_t kernel_height,
    int64_t kernel_width,
    bool use_bias) {
    Conv3dWeights weights;
    weights.weight = store.load_tensor(
        source,
        prefix + ".weight",
        storage_type,
        {out_channels * in_channels, kernel_depth, kernel_height, kernel_width});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
DepthwiseConv1dWeights depthwise_conv1d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t channels,
    int64_t kernel_size,
    bool use_bias) {
    DepthwiseConv1dWeights weights;
    weights.weight = store.load_tensor(source, prefix + ".weight", storage_type, {channels, 1, kernel_size});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {channels});
    }
    return weights;
}

template <typename Store>
ConvTranspose1dWeights conv_transpose1d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel_size,
    bool use_bias) {
    ConvTranspose1dWeights weights;
    weights.weight = store.load_tensor(source, prefix + ".weight", storage_type, {in_channels, out_channels, kernel_size});
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
ConvTranspose1dWeights weight_norm_conv_transpose1d_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel_size,
    bool use_bias) {
    ConvTranspose1dWeights weights;
    weights.weight = weight_norm_tensor_from_source(
        store,
        source,
        prefix,
        storage_type,
        core::TensorShape::from_dims({in_channels, out_channels, kernel_size}));
    if (use_bias) {
        weights.bias = store.load_f32_tensor(source, prefix + ".bias", {out_channels});
    }
    return weights;
}

template <typename Store>
ConvTranspose1dWeights conv_transpose1d_from_source_resolving_weight_norm(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    assets::TensorStorageType storage_type,
    int64_t in_channels,
    int64_t out_channels,
    int64_t kernel_size,
    bool use_bias) {
    if (source.has_tensor(prefix + ".weight")) {
        return conv_transpose1d_from_source(
            store,
            source,
            prefix,
            storage_type,
            in_channels,
            out_channels,
            kernel_size,
            use_bias);
    }
    return weight_norm_conv_transpose1d_from_source(
        store,
        source,
        prefix,
        storage_type,
        in_channels,
        out_channels,
        kernel_size,
        use_bias);
}

template <typename Store>
LSTMCellWeights lstm_cell_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    int64_t layer,
    int64_t input_size,
    int64_t hidden_size,
    assets::TensorStorageType storage_type) {
    const std::string suffix = "_l" + std::to_string(layer);
    return {
        store.load_tensor(source, prefix + ".weight_ih" + suffix, storage_type, {4 * hidden_size, input_size}),
        store.load_tensor(source, prefix + ".weight_hh" + suffix, storage_type, {4 * hidden_size, hidden_size}),
        store.load_f32_tensor(source, prefix + ".bias_ih" + suffix, {4 * hidden_size}),
        store.load_f32_tensor(source, prefix + ".bias_hh" + suffix, {4 * hidden_size}),
    };
}

template <typename Store>
LSTMStackWeights lstm_stack_from_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & prefix,
    int64_t layers,
    int64_t input_size,
    int64_t hidden_size,
    assets::TensorStorageType storage_type) {
    LSTMStackWeights weights;
    weights.layers.reserve(static_cast<size_t>(layers));
    for (int64_t layer = 0; layer < layers; ++layer) {
        weights.layers.push_back(lstm_cell_from_source(
            store,
            source,
            prefix,
            layer,
            layer == 0 ? input_size : hidden_size,
            hidden_size,
            storage_type));
    }
    return weights;
}

template <typename Store>
LinearWeights linear_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name,
    assets::TensorStorageType storage_type) {
    LinearWeights weights;
    weights.weight = tensor_from_named_source(store, source, weight_name, storage_type);
    if (bias_name.has_value()) {
        weights.bias = f32_tensor_from_named_source(store, source, *bias_name);
    }
    return weights;
}

template <typename Store>
NormWeights norm_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name) {
    NormWeights weights;
    weights.weight = f32_tensor_from_named_source(store, source, weight_name);
    if (bias_name.has_value()) {
        weights.bias = f32_tensor_from_named_source(store, source, *bias_name);
    }
    return weights;
}

template <typename Store>
Conv1dWeights conv1d_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name,
    assets::TensorStorageType storage_type) {
    Conv1dWeights weights;
    weights.weight = tensor_from_named_source(store, source, weight_name, storage_type);
    if (bias_name.has_value()) {
        weights.bias = f32_tensor_from_named_source(store, source, *bias_name);
    }
    return weights;
}

template <typename Store>
ConvTranspose1dWeights conv_transpose1d_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & weight_name,
    const std::optional<std::string> & bias_name,
    assets::TensorStorageType storage_type,
    bool prepack_col2im = false) {
    ConvTranspose1dWeights weights;
    weights.weight = tensor_from_named_source(store, source, weight_name, storage_type);
    if (bias_name.has_value()) {
        weights.bias = f32_tensor_from_named_source(store, source, *bias_name);
    }
    if (prepack_col2im) {
        core::validate_rank_between(weights.weight, 3, 3, "conv_transpose1d weight");
        const auto & shape = weights.weight.shape;
        const int64_t inputs = shape.dims[0], outputs = shape.dims[1], kernel = shape.dims[2];
        const auto values = assets::tensor_data_to_f32(weight_name,
            source.require_tensor(weight_name, storage_type, {inputs, outputs, kernel}));
        std::vector<float> matrix(values.size());
        for (int64_t output = 0; output < outputs; ++output) {
            for (int64_t tap = 0; tap < kernel; ++tap) {
                for (int64_t input = 0; input < inputs; ++input) {
                    matrix[static_cast<size_t>((output * kernel + tap) * inputs + input)] =
                        values[static_cast<size_t>((input * outputs + output) * kernel + tap)];
                }
            }
        }
        weights.col2im_matrix = store.make_f32(core::TensorShape::from_dims({outputs * kernel, inputs}), std::move(matrix));
    }
    return weights;
}

template <typename Store>
LayerScaleWeights layer_scale_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & name) {
    return {f32_tensor_from_named_source(store, source, name)};
}

template <typename Store>
std::optional<LayerScaleWeights> optional_layer_scale_from_named_source(
    Store & store,
    const assets::TensorSource & source,
    const std::string & name) {
    if (!source.has_tensor(name)) {
        return std::nullopt;
    }
    return layer_scale_from_named_source(store, source, name);
}

template <typename Store, typename TensorData>
LayerScaleWeights layer_scale_data(Store & store, const TensorData & scale) {
    return {tensor_data(store, scale)};
}

}  // namespace engine::modules::binding