Differences between 'generate_wedlm' and 'llm_engine'

Considering only single-batch inference and disregarding multi-GPU scheduling, can the `generate_wedlm` method in `WeDLMForCausalLM` class be considered a minimalist version of `wedlm`'s `llm_engine`? Are there differences in speed and accuracy between the two?