Function apply_modulated_rms_norm

src/models/ace_step/diffusion.cpp:228–241

    auto embedding = modules::ConcatModule({1}).build(ctx, cos_part, sin_part);
    auto temb = modules::LinearModule({256, hidden_size, true, GGML_PREC_F32}).build(ctx, embedding, weights.fc1);
    temb = modules::SiluModule{}.build(ctx, temb);
    temb = modules::LinearModule({hidden_size, hidden_size, true, GGML_PREC_F32}).build(ctx, temb, weights.fc2);
    auto proj_in = modules::SiluModule{}.build(ctx, temb);
    auto timestep_proj =
        modules::LinearModule({hidden_size, hidden_size * 6, true, GGML_PREC_F32}).build(ctx, proj_in, weights.time_proj);
    timestep_proj = core::reshape_tensor(
        ctx,
        timestep_proj,
        core::TensorShape::from_dims({timestep_proj.shape.dims[0], int64_t{6}, hidden_size}));
    return TimeEmbeddingOutputs{temb, timestep_proj};
}

core::TensorValue build_mlp(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input,
    const modules::LinearWeights & gate_proj,
    const modules::LinearWeights & up_proj,