[feat] enable hierarchical mc2 ops on A2 by default (#5300)

### What this PR does / why we need it?
Previously, it was necessary to set the environment variables
HCCL_INTRA_PCIE_ENABLE=1 and HCCL_INTRA_ROCE_ENABLE=0. This PR enables
hierarchical MC2 operations on A2 by default.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c

Signed-off-by: hwhaokun <haokun0405@163.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -31,8 +31,7 @@ from vllm.distributed.parallel_state import get_ep_group
 from vllm_ascend.distributed.parallel_state import get_mc2_group
 from vllm_ascend.ops.fused_moe.comm_utils import (
     async_all_to_all, gather_from_sequence_parallel_region)
-from vllm_ascend.utils import (AscendDeviceType, get_ascend_device_type,
-                               is_hierarchical_communication_enabled)
+from vllm_ascend.utils import AscendDeviceType, get_ascend_device_type
 
 
 class MoETokenDispatcher(ABC):
@@ -101,10 +100,6 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
         self.need_extra_args = (
             get_ascend_device_type() == AscendDeviceType.A3)
 
-        # NOTE: When in A2, setting the environment variables HCCL_INTRA_PCIE_ENABLE=1 and
-        # HCCL_INTRA_ROCE_ENABLE=0 can reduce cross-machine communication traffic and significantly
-        # improve communication performance.
-        self.need_expert_scale = is_hierarchical_communication_enabled()
         self.with_quant = False
 
         # Here we need to calculate the global_bs = max_bs_per_rank * ep_world_size to execute
@@ -142,6 +137,7 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
         else:
             quant_mode = 0
             moe_expert_num = len(expert_map)
+
         kwargs_mc2 = {
             "x": hidden_states,
             "expert_ids": topk_ids,
@@ -150,8 +146,12 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
             "moe_expert_num": moe_expert_num,
             "global_bs": self.global_bs,
             "expert_token_nums_type": 0,
+            "expert_scales": topk_weights.to(torch.float32),
         }
 
+        if get_ascend_device_type() == AscendDeviceType.A2:
+            kwargs_mc2["comm_alg"] = "hierarchy"
+
         stage1_kwargs = {
             "scales": None,
             "quant_mode": quant_mode,
@@ -165,11 +165,6 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
                 "tp_world_size": 1,
                 "tp_rank_id": 0,
             })
-        if self.need_expert_scale:
-            stage1_kwargs.update({
-                "expert_scales":
-                topk_weights.to(torch.float32),
-            })
 
         kwargs_mc2.update(stage1_kwargs)
         return kwargs_mc2
@@ -269,8 +264,12 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
             "shared_expert_rank_num": 0,
             "moe_expert_num": moe_expert_num,
             "global_bs": self.global_bs,
+            "expand_scales": expand_scales,
         }
 
+        if get_ascend_device_type() == AscendDeviceType.A2:
+            kwargs_mc2["comm_alg"] = "hierarchy"
+
         if self.with_quant:
             tp_recv_counts = torch.empty(1,
                                          dtype=torch.int32,
@@ -281,7 +280,6 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
             "group_ep": self.moe_all_to_all_group_name,
             "ep_world_size": self.ep_world_size,
             "ep_rank_id": self.ep_rank_id,
-            "expand_scales": expand_scales,
         }
 
         if self.enable_dispatch_v2: