Method build

src/framework/modules/text_encoders/t5_base_encoder.cpp:221–246

}

core::TensorValue T5BaseEncoderModule::build(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input_ids,
    const core::TensorValue & relative_position_buckets,
    const core::TensorValue & additive_attention_mask,
    const T5BaseEncoderWeights & weights) const {
    core::validate_rank_between(input_ids, 2, 2, "T5BaseEncoder input_ids");
    const int64_t batch = input_ids.shape.dims[0];
    const int64_t tokens = input_ids.shape.dims[1];
    core::validate_shape(relative_position_buckets, core::TensorShape::from_dims({tokens, tokens}), "T5BaseEncoder relative_position_buckets");
    core::validate_shape(
        additive_attention_mask,
        core::TensorShape::from_dims({batch, config_.attention_heads, tokens, tokens}),
        "T5BaseEncoder additive_attention_mask");
    core::validate_shape(weights.relative_attention_bias, core::TensorShape::from_dims(
        {config_.relative_attention_num_buckets, config_.attention_heads}), "T5BaseEncoder relative_attention_bias");
    if (static_cast<int64_t>(weights.layers.size()) != config_.layers) {
        throw std::runtime_error("T5BaseEncoder layer count mismatch");
    }
    auto hidden = EmbeddingModule({config_.vocab_size, config_.hidden_size}).build(ctx, input_ids, weights.embed_tokens);
    const auto position_bias = relative_position_bias(ctx, relative_position_buckets, weights, config_, batch);
    for (int64_t i = 0; i < config_.layers; ++i) {
        hidden = encoder_layer(ctx, hidden, position_bias, additive_attention_mask, weights.layers[static_cast<size_t>(i)], config_);
    }
    return t5_layer_norm(ctx, hidden, weights.final_layer_norm, config_);
}

}  // namespace engine::modules