remove two_batch_overlap of moe and update use_mla

vllm/model_executor/layers/fused_moe/layer.py

@@ -1200,9 +1200,6 @@ class FusedMoE(CustomOp):
 
         self.quant_method.create_weights(layer=self, **moe_quant_params)
 
-        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
-        self.tbo_all_reduce = tbo_all_reduce
-
         # moe_fused_gate kernel ensure that num_experts/num_expert_group does not exceed MAX_VPT=32 now. And when kernel can handle MAX_VPT > 32, we can remove this assertion.
         self.use_fused_gate = envs.VLLM_ENABLE_MOE_FUSED_GATE \
             and self.e_score_correction_bias is not None \

vllm/v1/worker/gpu_model_runner.py

@@ -3854,7 +3854,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                               kv_cache_spec.page_size_bytes)
                 if isinstance(kv_cache_spec, AttentionSpec):
                     has_attn = True
-                    if envs.VLLM_USE_FLASH_ATTN_PA and not kv_cache_spec.use_mla:
+                    if envs.VLLM_USE_FLASH_ATTN_PA and not self.vllm_config.model_config.use_mla:
                         key_cache_shape, value_cache_shape = attn_backend.get_kv_cache_shape(
                             num_blocks,
                             kv_cache_spec.block_size,