Method build

src/framework/modules/linear_module.cpp:73–108

}

core::TensorValue LinearModule::build(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input,
    const LinearWeights & weights) const {
    if (ctx.ggml == nullptr) {
        throw std::runtime_error("ModuleBuildContext.ggml is null");
    }

    core::validate_rank_between(input, 1, core::kMaxTensorRank, "input");
    core::validate_last_dim(input, config_.in_features, "input");
    validate_weight_shape(config_, weights);

    const core::TensorValue contiguous_input = core::ensure_backend_addressable_layout(ctx, input);

    const core::TensorShape matrix_input_shape = flatten_to_matrix_shape(contiguous_input.shape);
    core::TensorValue matrix_input = core::reshape_tensor(ctx, contiguous_input, matrix_input_shape);

    ggml_tensor * projected_raw = ggml_mul_mat(ctx.ggml, weights.weight.tensor, matrix_input.tensor);
    if (config_.precision != GGML_PREC_DEFAULT) {
        ggml_mul_mat_set_prec(projected_raw, config_.precision);
    }
    if (config_.cuda_nvfp4_f16_activation_lowering) {
        ggml_mul_mat_set_lowering(projected_raw, GGML_MUL_MAT_LOWERING_CUDA_NVFP4_F16_ACTIVATION);
    }
    core::TensorValue projected = core::wrap_tensor(
        projected_raw,
        core::TensorShape::from_dims({matrix_input_shape.at(0), config_.out_features}),
        GGML_TYPE_F32);

    if (config_.use_bias) {
        ggml_tensor * biased_raw = ggml_add(ctx.ggml, projected.tensor, weights.bias->tensor);
        projected = core::wrap_tensor(
            biased_raw,
            core::TensorShape::from_dims({matrix_input_shape.at(0), config_.out_features}),
            GGML_TYPE_F32);
    }

    return core::reshape_tensor(ctx, projected, input.shape.with_last_dim(config_.out_features));
}