[Dist][EP] Remove ETP/EP maintained in vllm-ascend (#1681)

### What this PR does / why we need it?
Remove ETP/EP maintained in branch main. We drop this as there is no
relevant scenarios to use ETP now, and we may subsequently advocate
implementing expert tensor parallelism in vLLM to support scenarios
where the expert is needed to be sliced

This is a part of #1422 backport.

Fixes https://github.com/vllm-project/vllm-ascend/issues/1396
https://github.com/vllm-project/vllm-ascend/issues/1154

### Does this PR introduce _any_ user-facing change?
We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in
vllm instead.

### How was this patch tested?
CI passed with new added and existing test.


- vLLM version: v0.9.2
- vLLM main:
fe8a2c544a

Signed-off-by: MengqingCao <cmq0113@163.com>

vllm_ascend/models/deepseek_v2.py

@@ -39,7 +39,7 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                               tensor_model_parallel_all_gather,
                               tensor_model_parallel_all_reduce,
                               tensor_model_parallel_reduce_scatter)
-from vllm.distributed.parallel_state import get_dp_group
+from vllm.distributed.parallel_state import get_dp_group, get_ep_group
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -69,7 +69,6 @@ from vllm.model_executor.models.utils import (
 from vllm.sequence import IntermediateTensors
 
 from vllm_ascend.ascend_config import get_ascend_config
-from vllm_ascend.distributed.parallel_state import get_ep_group
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 from vllm_ascend.quantization.quant_config import AscendLinearMethod
 from vllm_ascend.quantization.w8a8_dynamic import AscendW8A8DynamicLinearMethod

vllm_ascend/models/pangu_moe.py

@@ -30,8 +30,8 @@ from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (divide, get_pp_group,
                               get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_reduce)
-from vllm.distributed.parallel_state import (get_dp_group, get_tp_group,
-                                             get_world_group)
+from vllm.distributed.parallel_state import (get_dp_group, get_ep_group,
+                                             get_tp_group, get_world_group)
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
@@ -58,7 +58,6 @@ from vllm.model_executor.utils import set_weight_attrs
 from vllm.sequence import IntermediateTensors
 
 from vllm_ascend.ascend_config import get_ascend_config
-from vllm_ascend.distributed.parallel_state import get_ep_group
 from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, is_310p
 
 logger = init_logger(__name__)
@@ -93,7 +92,7 @@ class CustomMergedColumnParallelLinear(LinearBase):
         # Divide the weight matrix along the last dimension.
         output_size = sum(output_sizes)
         self.output_sizes = output_sizes
-        self.tp_size = get_world_group().world_size
+        self.tp_size = get_tp_group().world_size
         self.input_size_per_partition = input_size
         self.output_size_per_partition = divide(output_size, self.tp_size)
         self.output_partition_sizes = [self.output_size_per_partition]
@@ -144,8 +143,8 @@ class CustomMergedColumnParallelLinear(LinearBase):
 
         assert loaded_shard_id < len(self.output_sizes)
 
-        tp_rank = get_world_group().rank_in_group
-        tp_size = get_world_group().world_size
+        tp_rank = get_tp_group().rank_in_group
+        tp_size = get_tp_group().world_size
         if output_dim is not None:
             shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
             shard_size = self.output_sizes[loaded_shard_id] // tp_size
@@ -204,7 +203,7 @@ class CustomRowParallelLinear(LinearBase):
         group=None,
     ):
         # Divide the weight matrix along the first dimension.
-        self.group = group if group is not None else get_world_group()
+        self.group = group if group is not None else get_tp_group()
         self.tp_rank = self.group.rank_in_group
         self.tp_size = self.group.world_size
         self.input_size_per_partition = divide(input_size, self.tp_size)
@@ -357,7 +356,7 @@ def topk_wrapper(num_voted_experts):
         num_tokens = scores.shape[0]
         router_scale = _ROUTER_SCALE.squeeze(  # type: ignore
         )
-
+        # TODO: support disable expert parallel
         ep_size = get_ep_group().world_size
         local_num_experts = global_num_experts // ep_size
         local_num_group = topk // ep_size
@@ -464,6 +463,7 @@ class PanguProMoESparseMoeBlock(nn.Module):
             custom_routing_function=topk_wrapper(num_voted_experts),
             prefix=f"{prefix}.experts",
         )
+        self.use_ep = self.experts.use_ep
 
         self.gate = ReplicatedLinear(
             config.hidden_size,