Revert "update kv_b_proj_weight"

This reverts commit 3b121add.

vllm/v1/attention/backends/mla/common.py

@@ -1102,9 +1102,6 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
         # we currently do not have quantized bmm's which are needed for
         # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform
         # the bmm's in 16-bit, the extra memory overhead of this is fairly low
-        if self.use_llama_nn and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod):
-            kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj)
-        else:
         kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
         assert kv_b_proj_weight.shape == (
             self.kv_lora_rank,