q_lora_rank in vLLM with DeepSeek MLA Model

[erictanjn]

I'm using vLLM to run the DeepSeek model with MLA architecture.
I'd like to know - ignoring matrix absorption - in the current vLLM MLA prefill implementation, whether the model parameters are reduced compared to MHA. It seems they're not, only the KV is reduced. If in this way MLA doesn't reduce model parameter size and only decreases KV cache memory usage, does it become a paradox and unnecessary in non-long-sequence scenarios?
If I've made any mistakes, I hope everyone can point them out for me. Thanks