add contiguous+rmsnorm to replace triton_

vllm/model_executor/models/deepseek_v2.py

@@ -596,6 +596,9 @@ class DeepseekV2MLAAttention(nn.Module):
                 q = self.q_proj(hidden_states)[0]
             kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
                 [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+            if envs.VLLM_USE_LIGHTOP:
+                kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
+            else:
                 kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
 
             q = q.view(-1, self.num_local_heads, self.qk_head_dim)