[Paper Bug] The paper missed an important detail

[Kamichanw]

The paper does not mention always refreshing the first layer, which is critical to reproduce your work. The related code is pasted below:

dLLM-cache/dllm_cache/hooks/cache_hook_LLaDA.py

Lines 156 to 157 in e289751

	refresh_gen = feature_cache.refresh_gen(layer_id=self.layer_id) or self.layer_id == 0
	refresh_prompt = feature_cache.refresh_prompt(layer_id=self.layer_id) or self.layer_id == 0