[Misc]: Why not AscendMLAImpl just inherit MLACommonImpl for common functions?

[learning-chip]

class AscendMLAImpl inherits the abstract MLAAttentionImpl:
https://github.com/vllm-project/vllm-ascend/blob/v0.8.5rc1/vllm_ascend/attention/mla_v1.py#L263

However most functions like _v_up_proj_and_o_proj and _q_proj_and_k_up_proj are already defined in MLACommonImpl, and they look identical to the re-definition in AscendMLAImpl:
https://github.com/vllm-project/vllm/blob/v0.8.5/vllm/attention/backends/mla/common.py#L1009

For example, TritonMLAImpl just reuses those common functions, but overwrites _forward_decode:
https://github.com/vllm-project/vllm/blob/v0.8.5/vllm/attention/backends/triton_mla.py#L26

AscendMLAImpl can also just inherits MLACommonImpl and overwrite _forward_prefill and _forward_decode to use NPU operators?

[Yikun]

@ganyi1996ppo Please take a look

[ganyi1996ppo]

Inherit MLACommonImpl will import triton which we are not support yet when we were developing mla_v1, so we just wrote a simple impl to make sure its functionality. However I see that the common.py have already guard the triton import with HAS_TRITON variable now, so maybe its safe to inherit it in vllm-ascend. You are welcome to file to PR to vllm-ascend on this.