[Feature] Skip useless computation

### Motivation

I found that the logit in lmdeploy was directly from the last lm_head. Can we skip useless computation from the last attn to lm_head for unwanted logits? llama.cpp applied this for example https://github.com/ggml-org/llama.cpp/blob/5d46babdc2d4675d96ebcf23cac098a02f0d30cc/src/llama-model.cpp#L6888
@lvhan028

### Related resources

_No response_

### Additional context

_No response_