Method build

src/framework/modules/text_encoders/t5_gemma_encoder.cpp:187–213

T5GemmaEncoderModule::T5GemmaEncoderModule(T5GemmaEncoderConfig config) : config_(config) {
    validate_config(config_);
}

const T5GemmaEncoderConfig & T5GemmaEncoderModule::config() const noexcept {
    return config_;
}

core::TensorValue T5GemmaEncoderModule::build(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input_ids,
    const core::TensorValue & positions,
    const core::TensorValue & additive_attention_mask,
    const T5GemmaEncoderWeights & weights) const {
    core::validate_rank_between(input_ids, 2, 2, "T5GemmaEncoder input_ids");
    core::validate_shape(positions, core::TensorShape::from_dims({input_ids.shape.dims[1]}), "T5GemmaEncoder positions");
    core::validate_shape(
        additive_attention_mask,
        core::TensorShape::from_dims({input_ids.shape.dims[0], config_.attention_heads, input_ids.shape.dims[1], input_ids.shape.dims[1]}),
        "T5GemmaEncoder additive_attention_mask");
    if (static_cast<int64_t>(weights.layers.size()) != config_.layers) {
        throw std::runtime_error("T5GemmaEncoder layer count mismatch");
    }
    auto hidden = EmbeddingModule({config_.vocab_size, config_.hidden_size}).build(ctx, input_ids, weights.embed_tokens);
    if (config_.scale_embeddings) {
        hidden = core::wrap_tensor(
            ggml_scale(ctx.ggml, ensure_contiguous(ctx, hidden).tensor, std::sqrt(static_cast<float>(config_.hidden_size))),
            hidden.shape,
            GGML_TYPE_F32);
    }
    for (int64_t i = 0; i < config_.layers; ++i) {
        hidden = layer(ctx, hidden, positions, additive_attention_mask, weights.layers[static_cast<size_t>(i)], config_);