fix: fix bug http://hpczentao.sugon.com/bug-view-118388.html

vllm/model_executor/layers/attention/mla_attention.py

@@ -1249,7 +1249,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
         # `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
         from vllm.model_executor.layers.linear import UnquantizedLinearMethod
-        if (envs.VLLM_USE_NN or self.use_llama_nn) and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod):
+        if (envs.VLLM_USE_NN or 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, out_dtype=act_dtype
             )