update triton_mla.py

vllm/attention/backends/triton_mla.py

@@ -120,8 +120,9 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
             match_seq_len = int((decode_meta.seq_lens_tensor.sum()/ max(1, B)).item())
         else:
             match_seq_len = max_seq_len
-            
-        best_config = self.attn_configs[min(self.attn_configs.keys(), key=lambda x: abs(int(x) - match_seq_len))]
+        
+        if envs.VLLM_USE_TRITON_OPT_MLA:
+            best_config = self.attn_configs[min(self.attn_configs.keys(), key=lambda x: abs(int(x) - match_seq_len))]
                 
         # Run MQA
         decode_attention_fwd(q, kv_c_and_k_pe_cache, kv_c_cache, o,