[Refactor] [MoE] Rename moe-related classes & files (#3646)

### What this PR does / why we need it?
1. Rename common_fused_moe.py to fused_moe.py.
2. Rename fused_moe_prepare_and_finalize.py / FusedMoEPrepareAndFinalize
to prepare_finalize.py / PrepareAndFinalize.
3. Rename vllm_ascend/ops/moe to vllm_ascend/ops/fused_moe.
4. Move vllm_ascend/ops/fused_moe.py to
vllm_ascend/ops/fused_moe/fused_moe.py
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
e2e & ut

- vLLM version: v0.11.0rc3
- vLLM main:
17c540a993

Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>

vllm_ascend/ops/fused_moe/comm_utils.py

@@ -0,0 +1,113 @@
+# Copyright (c) 2024; NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+import torch
+import torch.distributed
+import torch.distributed as dist
+import torch_npu
+
+COMM_STREAM = None
+
+
+def async_all_to_all(input_,
+                     output_split_sizes,
+                     input_split_sizes,
+                     group,
+                     event=None):
+    if output_split_sizes is None:
+        # Equal split (all2all)
+        a2a_out = torch.empty_like(input_)
+    else:
+        # Unequal split (all2all-v)
+        a2a_out = input_.new_empty(
+            size=[sum(output_split_sizes)] + list(input_.size()[1:]),
+            dtype=input_.dtype,
+            device=torch.npu.current_device(),
+        )
+
+    if event:
+        # multi stream wait event
+        global COMM_STREAM
+        if COMM_STREAM is None:
+            COMM_STREAM = torch_npu.npu.Stream(
+                device=torch.npu.current_device())
+        with torch_npu.npu.stream(COMM_STREAM):
+            event.wait()
+            handle = dist.all_to_all_single(
+                a2a_out,
+                input_.contiguous(),
+                output_split_sizes=output_split_sizes,
+                input_split_sizes=input_split_sizes,
+                group=group,
+                async_op=True)
+    else:
+        handle = dist.all_to_all_single(a2a_out,
+                                        input_.contiguous(),
+                                        output_split_sizes=output_split_sizes,
+                                        input_split_sizes=input_split_sizes,
+                                        group=group,
+                                        async_op=True)
+    return input_, a2a_out, handle
+
+
+def _gather_along_first_dim(input_, group, output_split_sizes=None):
+    """Gather tensors and concatenate along the first dimension.
+
+    Args:
+        input_tensor (torch.Tensor):
+            A tensor to be gathered.
+        output_split_sizes (List[int], optional):
+            A list specifying the sizes of the output splits along the first dimension.
+            If None, equal splitting is assumed. Default: None.
+
+    Returns:
+        torch.Tensor: Gathered tensor.
+    """
+    world_size = torch.distributed.get_world_size(group)
+    # Bypass the function if we are using only 1 GPU.
+    if world_size == 1:
+        return input_
+
+    dim_size = list(input_.size())
+    if output_split_sizes is None:
+        dim_size[0] = dim_size[0] * world_size
+
+        output = torch.empty(dim_size,
+                             dtype=input_.dtype,
+                             device=torch.npu.current_device())
+        torch.distributed.all_gather_into_tensor(output,
+                                                 input_.contiguous(),
+                                                 group=group)
+    else:
+        dim_size[0] = sum(output_split_sizes)
+        output = torch.empty(dim_size,
+                             dtype=input_.dtype,
+                             device=torch.npu.current_device())
+        output_tensor_list = list(
+            torch.split(output, output_split_sizes, dim=0))
+        torch.distributed.all_gather(output_tensor_list, input_, group=group)
+
+    return output
+
+
+def gather_from_sequence_parallel_region(
+    input_,
+    group,
+    output_split_sizes=None,
+):
+    """Wrapper for autograd function: forward: AG, backward: RS <first dim>"""
+    return _gather_along_first_dim(input_, group, output_split_sizes)

vllm_ascend/ops/fused_moe/experts_selector.py

@@ -0,0 +1,277 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+from typing import Callable, Optional
+
+import torch
+import torch_npu
+from vllm.forward_context import get_forward_context
+
+from vllm_ascend.ascend_config import get_ascend_config
+
+
+def select_experts(hidden_states: torch.Tensor,
+                   router_logits: torch.Tensor,
+                   top_k: int,
+                   use_grouped_topk: bool,
+                   renormalize: bool,
+                   topk_group: Optional[int] = None,
+                   num_expert_group: Optional[int] = None,
+                   custom_routing_function: Optional[Callable] = None,
+                   scoring_func: str = "softmax",
+                   routed_scaling_factor=1.0,
+                   e_score_correction_bias: Optional[torch.Tensor] = None,
+                   indices_type: Optional[torch.dtype] = None,
+                   global_num_experts: int = -1):
+    """
+    Fused experts with select experts.
+
+    Args:
+        router_logits: router logits of shape (num_tokens, hidden_size).
+        hidden_states: Hidden states of shape (num_tokens, hidden_size).
+        top_k: number of top k experts.
+        use_grouped_topk: Whether to group experts before selecting top-k.
+        renormalize: Whether to renormalize the routing weights.
+        topk_group: Number of expert groups to select from.
+        num_expert_group: Number of experts in each group.
+        custom_routing_function: Custom routing function.
+        scoring_func: Scoring function to use.
+        e_score_correction_bias: Correction bias to apply to expert scores.
+        indices_type: dtype of indices
+        global_num_experts: Global number of experts.
+
+    Returns:
+        topk_weights: router weights of shape (num_tokens, top_k).
+        topk_ids: selected expert IDs of shape (num_tokens, top_k).
+    """
+    # prefetch w1_w3_proj.weight preprocess
+    weight_prefetch_method = get_forward_context().weight_prefetch_method
+    if weight_prefetch_method:
+        weight_prefetch_method.maybe_prefetch_moe_weight_preprocess(
+            hidden_states, "gate_up")
+    topk_weights, topk_ids = _select_experts_with_fusion_ops(
+        hidden_states=hidden_states,
+        router_logits=router_logits,
+        top_k=top_k,
+        use_grouped_topk=use_grouped_topk,
+        topk_group=topk_group,
+        renormalize=renormalize,
+        e_score_correction_bias=e_score_correction_bias,
+        num_expert_group=num_expert_group,
+        custom_routing_function=custom_routing_function,
+        scoring_func=scoring_func,
+        routed_scaling_factor=routed_scaling_factor,
+        global_num_experts=global_num_experts)
+
+    if topk_weights is None:
+        topk_weights, topk_ids = _native_select_experts(
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+            top_k=top_k,
+            use_grouped_topk=use_grouped_topk,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias,
+            global_num_experts=global_num_experts,
+        )
+    return topk_weights, topk_ids
+
+
+def _native_grouped_topk(
+    topk_weights: torch.Tensor,
+    num_expert_group: Optional[int],
+    topk_group: Optional[int],
+):
+    topk_group = 0 if topk_group is None else topk_group
+    num_expert_group = 0 if num_expert_group is None else num_expert_group
+
+    num_token = topk_weights.shape[0]
+    grouped_weights = topk_weights.view(num_token, num_expert_group,
+                                        -1).max(dim=-1).values
+    topk_group_indices = torch.topk(grouped_weights.to(torch.float32),
+                                    k=topk_group,
+                                    dim=-1,
+                                    sorted=False)[1]
+    topk_group_mask = torch.zeros_like(grouped_weights)
+    topk_group_mask.scatter_(1, topk_group_indices, 1)
+    topk_weight_mask = (topk_group_mask.unsqueeze(-1).expand(
+        num_token, num_expert_group,
+        topk_weights.shape[-1] // num_expert_group).reshape(num_token, -1))
+    topk_weights = topk_weights.masked_fill(~topk_weight_mask.bool(), 0.0)
+
+    return topk_weights
+
+
+def _renormalize_topk_weights(
+    topk_weights: torch.Tensor,
+    renormalize: bool,
+):
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+    return topk_weights
+
+
+def _select_expert_use_group_topk(
+        topk_weights: torch.Tensor, topk_group: Optional[int],
+        renormalize: bool, top_k: int, num_expert_group: Optional[int],
+        e_score_correction_bias: Optional[torch.Tensor]):
+    assert topk_group is not None
+    assert num_expert_group is not None
+
+    if e_score_correction_bias is not None:
+        # Store original scores before applying correction bias. We use biased
+        # scores for expert selection but original scores for routing weights
+        original_weights = topk_weights
+        topk_weights = topk_weights + e_score_correction_bias.unsqueeze(0)
+
+    # TODO: Change to npu_group_topk when the latest CANN and NNAL is available
+    # >>> torch_npu._npu_group_topk(topk_weights, group_num=num_expert_group, k=topk_group)
+    topk_weights = _native_grouped_topk(topk_weights, num_expert_group,
+                                        topk_group)
+    # TODO bfloat16 is not supported in torch.topk with ge graph.
+    if e_score_correction_bias is not None:
+        topk_ids = torch.topk(topk_weights.to(torch.float32),
+                              k=top_k,
+                              dim=-1,
+                              sorted=False)[1]
+        # Use original unbiased scores for the routing weights
+        topk_weights = original_weights.gather(1, topk_ids)
+    else:
+        topk_weights, topk_ids = torch.topk(topk_weights.to(torch.float32),
+                                            k=top_k,
+                                            dim=-1,
+                                            sorted=False)
+    topk_ids = topk_ids.to(torch.int32)
+    topk_weights = _renormalize_topk_weights(topk_weights, renormalize)
+    return topk_weights, topk_ids
+
+
+def _select_experts_with_fusion_ops(
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        use_grouped_topk: bool,
+        renormalize: bool,
+        e_score_correction_bias: Optional[torch.Tensor],
+        topk_group: Optional[int],
+        num_expert_group: Optional[int],
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        routed_scaling_factor=1.0,
+        global_num_experts: int = -1):
+
+    topk_weights, topk_ids = None, None
+    # NOTE: now npu_moe_gating_top_k can only support 'group_count=256' pattern
+    global_redundant_expert_num = get_ascend_config().init_redundancy_expert
+    is_deepseek_v3_r1 = global_num_experts - global_redundant_expert_num == 256
+    if is_deepseek_v3_r1:
+        topk_weights, topk_ids, _ = torch_npu.npu_moe_gating_top_k(
+            router_logits,
+            k=top_k,  # topk currently 8
+            bias=e_score_correction_bias,
+            k_group=topk_group,  # fix: 4
+            group_count=num_expert_group,  # fix 8
+            group_select_mode=
+            1,  # 0: the maximum in the group; 1: topk2.sum(fix)
+            renorm=0,  # 0: softmax->topk(fix); 1: topk->softmax
+            norm_type=1,  # 0: softmax; 1: sigmoid(fix)
+            # out_flag=False, # todo new api; should the third output be output
+            # y2_flag=False, # old api; should the third output be output
+            routed_scaling_factor=1,
+            eps=float(1e-20))
+    if not use_grouped_topk and custom_routing_function is None and scoring_func == "softmax":
+        topk_weights, topk_ids, _ = torch_npu.npu_moe_gating_top_k_softmax(
+            x=router_logits, finished=None, k=top_k)
+        topk_ids = topk_ids.to(torch.int32)
+        topk_weights = _renormalize_topk_weights(topk_weights, renormalize)
+
+    return topk_weights, topk_ids
+
+
+def _native_select_experts(
+    hidden_states: torch.Tensor,
+    router_logits: torch.Tensor,
+    top_k: int,
+    use_grouped_topk: bool,
+    renormalize: bool,
+    topk_group: Optional[int] = None,
+    num_expert_group: Optional[int] = None,
+    custom_routing_function: Optional[Callable] = None,
+    scoring_func: str = "softmax",
+    e_score_correction_bias: Optional[torch.Tensor] = None,
+    global_num_experts: Optional[torch.Tensor] = None
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Select top-k experts based on router logits.
+
+    Args:
+        hidden_states: Hidden states of shape (num_tokens, hidden_size).
+        router_logits: Router logits of shape (num_tokens, num_experts).
+        top_k: Number of experts to select.
+        use_grouped_topk: Whether to group experts before selecting top-k.
+        renormalize: Whether to renormalize the routing weights.
+        topk_group: Number of expert groups to select from.
+        num_expert_group: Number of experts in each group.
+        custom_routing_function: Custom routing function.
+        scoring_func: Scoring function to use.
+        e_score_correction_bias: Correction bias to apply to expert scores.
+
+    Returns:
+        topk_weights: Routing weights of shape (num_tokens, top_k).
+        topk_ids: Selected expert IDs of shape (num_tokens, top_k).
+
+    Raises:
+        ValueError: If an unsupported scoring function is provided.
+    """
+
+    if scoring_func == "softmax":
+        topk_weights = router_logits.softmax(dim=-1)
+    elif scoring_func == "sigmoid":
+        topk_weights = router_logits.sigmoid()
+    else:
+        raise ValueError(f"Unsupported scoring function: {scoring_func}")
+
+    if use_grouped_topk:
+        return _select_expert_use_group_topk(
+            topk_weights=topk_weights,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            e_score_correction_bias=e_score_correction_bias)
+
+    if custom_routing_function is not None:
+        topk_weights, topk_ids = custom_routing_function(
+            hidden_states=hidden_states,
+            gating_output=router_logits,
+            topk=top_k,
+            renormalize=renormalize,
+            global_num_experts=global_num_experts)
+        # Required by npu_moe_init_routing
+        topk_ids = topk_ids.to(torch.int32)
+        return topk_weights, topk_ids
+
+    topk_weights, topk_ids = topk_weights.topk(top_k, dim=-1)
+    topk_weights = topk_weights.to(hidden_states.dtype)
+
+    # Required by npu_moe_init_routing
+    topk_ids = topk_ids.to(torch.int32)
+    topk_weights = _renormalize_topk_weights(topk_weights, renormalize)
+
+    return topk_weights, topk_ids

vllm_ascend/ops/fused_moe/fused_moe.py

@@ -0,0 +1,492 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+import os.path
+from typing import Any, Callable, Optional
+
+import torch
+import torch_npu
+from vllm.config import get_current_vllm_config
+from vllm.distributed import (get_dp_group, get_ep_group, get_tp_group,
+                              tensor_model_parallel_all_reduce)
+from vllm.forward_context import get_forward_context
+from vllm.logger import logger
+from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig
+from vllm.model_executor.layers.fused_moe.layer import (
+    FusedMoE, UnquantizedFusedMoEMethod, determine_expert_map,
+    get_compressed_expert_map)
+
+from vllm_ascend.ascend_config import get_ascend_config
+from vllm_ascend.ascend_forward_context import MoECommType
+from vllm_ascend.distributed.parallel_state import get_mc2_group
+from vllm_ascend.eplb.core.eplb_utils import (determine_default_expert_map,
+                                              determine_default_log2phy_map)
+from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
+from vllm_ascend.ops.fused_moe.experts_selector import select_experts
+from vllm_ascend.ops.fused_moe.moe_comm_method import setup_moe_comm_method
+from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, enable_sp, is_310p,
+                               is_enable_nz, npu_stream_switch,
+                               shared_expert_dp_enabled,
+                               shared_experts_calculation_stream,
+                               vllm_version_is)
+
+if vllm_version_is("0.11.0"):
+    from vllm.config import CompilationLevel
+
+    from vllm.model_executor.layers.shared_fused_moe import SharedFusedMoE  # type: ignore # isort:skip
+else:
+    from vllm.config import CompilationMode
+    from vllm.model_executor.layers.fused_moe.shared_fused_moe import \
+        SharedFusedMoE
+
+
+class AscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
+
+    def __init__(self, moe: FusedMoEConfig = None):
+
+        super().__init__(moe=moe)
+
+        # NOTE: Currently, this self.use_aclgraph is only used in
+        # UnquantizedFusedMoEMethod.forward_oot to decide whether to use in
+        # ops/fused_moe.py:568 to circumvent torch.randint_like not supported issue.
+        # Once torch.randint_like is supported or removed, this flag can be removed.
+        vllm_config = get_current_vllm_config()
+        ascend_config = get_ascend_config()
+        self.dynamic_eplb = get_ascend_config().dynamic_eplb
+        if ascend_config.torchair_graph_config.enabled:
+            self.use_aclgraph = False
+        else:
+            if vllm_version_is("0.11.0"):
+                self.use_aclgraph = (
+                    vllm_config.compilation_config.level
+                    == CompilationLevel.PIECEWISE
+                    and not vllm_config.model_config.enforce_eager)
+            else:
+                self.use_aclgraph = (
+                    vllm_config.compilation_config.mode
+                    == CompilationMode.VLLM_COMPILE
+                    and not vllm_config.model_config.enforce_eager)
+
+        self.transpose = True
+
+    def process_weights_after_loading(self, layer):
+        super(UnquantizedFusedMoEMethod,
+              self).process_weights_after_loading(layer)
+        if self.transpose:
+            w13_data = self._maybe_pad_weight(layer.w13_weight.data).transpose(
+                1, 2).contiguous()
+            layer.w13_weight = torch.nn.Parameter(w13_data,
+                                                  requires_grad=False)
+
+            w2_data = self._maybe_pad_weight(layer.w2_weight.data).transpose(
+                1, 2).contiguous()
+            layer.w2_weight = torch.nn.Parameter(w2_data, requires_grad=False)
+
+            self.transpose = False
+        else:
+            w13_data = self._maybe_pad_weight(layer.w13_weight.data)
+            layer.w13_weight = torch.nn.Parameter(w13_data,
+                                                  requires_grad=False)
+
+            w2_data = self._maybe_pad_weight(layer.w2_weight.data)
+            layer.w2_weight = torch.nn.Parameter(w2_data, requires_grad=False)
+
+        if not is_310p() and is_enable_nz():
+            layer.w13_weight.data = torch_npu.npu_format_cast(
+                layer.w13_weight.data, ACL_FORMAT_FRACTAL_NZ)
+            layer.w2_weight.data = torch_npu.npu_format_cast(
+                layer.w2_weight.data, ACL_FORMAT_FRACTAL_NZ)
+
+    def apply(self,
+              layer: torch.nn.Module,
+              x: torch.Tensor,
+              use_grouped_topk: bool,
+              top_k: int,
+              router_logits: torch.Tensor,
+              renormalize: bool,
+              topk_group: Optional[int] = None,
+              num_expert_group: Optional[int] = None,
+              custom_routing_function: Optional[Callable] = None,
+              scoring_func: str = "softmax",
+              routed_scaling_factor: float = 1.0,
+              e_score_correction_bias: Optional[torch.Tensor] = None,
+              global_num_experts: int = -1,
+              expert_map: Optional[torch.Tensor] = None,
+              apply_router_weight_on_input: bool = False,
+              enable_force_load_balance: bool = False,
+              shared_experts: Optional[Any] = None,
+              **kwargs) -> torch.Tensor:
+
+        topk_weights, topk_ids = select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            top_k=top_k,
+            use_grouped_topk=use_grouped_topk,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            routed_scaling_factor=routed_scaling_factor,
+            e_score_correction_bias=e_score_correction_bias,
+            global_num_experts=global_num_experts)
+
+        topk_weights = topk_weights.to(x.dtype)
+        # this is a naive implementation for experts load balance so as
+        # to avoid accumulating too much tokens on a single rank.
+        # currently it is only activated when doing profile runs.
+        if enable_force_load_balance and not self.use_aclgraph:
+            topk_ids = torch.randint_like(topk_ids, 0, global_num_experts)
+
+        moe_comm_method = get_forward_context().moe_comm_method
+        return moe_comm_method.fused_experts(
+            hidden_states=x,
+            w1=layer.w13_weight,
+            w2=layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            shared_experts=shared_experts,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            dynamic_eplb=self.dynamic_eplb,
+            mc2_mask=kwargs.get("mc2_mask", None))
+
+
+class AscendFusedMoE(FusedMoE):
+    moe_counter = -1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        num_experts = kwargs["num_experts"]
+        intermediate_size = kwargs["intermediate_size"]
+
+        AscendFusedMoE.moe_counter += 1
+        self.moe_instance_id = AscendFusedMoE.moe_counter
+
+        self.global_num_experts = num_experts
+        self.expert_map = None
+        self.log2phy = None
+        self.global_redundant_expert_num = 0
+
+        if self.quant_config is None:
+            self.quant_method = AscendUnquantizedFusedMoEMethod(
+                self.moe_config)
+        else:
+            self.quant_method = self.quant_config.get_quant_method(
+                self, self.layer_name)
+
+        assert self.quant_method is not None
+
+        self.moe_config.tp_group = get_tp_group()
+        self.moe_config.dp_group = get_dp_group()
+        self.moe_config.ep_group = get_ep_group()
+        self.moe_config.mc2_group = get_mc2_group()
+        ascend_config = get_ascend_config()
+        self.dynamic_eplb = ascend_config.dynamic_eplb or ascend_config.expert_map_record_path
+        self.expert_map_path = ascend_config.expert_map_path
+        self.global_redundant_expert_num = ascend_config.init_redundancy_expert
+        self.global_num_experts = num_experts + self.global_redundant_expert_num
+        if self.custom_routing_function is None and self.e_score_correction_bias is not None:
+            vllm_config = get_current_vllm_config()
+            self.e_score_correction_bias.data = self.e_score_correction_bias.data.to(
+                dtype=vllm_config.model_config.dtype)
+        # static eplb initializing with expert_map_path
+        if self.expert_map_path and os.path.exists(
+                self.expert_map_path) and os.access(self.expert_map_path,
+                                                    os.R_OK):
+            self.expert_load_balancer = ExpertLoadBalancer(
+                self.expert_map_path, self.global_num_experts)
+            self.expert_load_balancer.check_expert_map_tensor()
+            self.global_redundant_expert_num = (
+                self.expert_load_balancer.get_global_redundant_expert_num())
+            try:
+                self.local_num_experts, self.expert_map = (
+                    self.expert_load_balancer.get_rank_placement_map(
+                        self.moe_instance_id, self.ep_rank))
+                self.log2phy = self.expert_load_balancer.get_rank_log2phy_map(
+                    self.moe_instance_id, self.ep_rank).npu()
+            except Exception as e:
+                logger.warning(
+                    f"Init expert map of mtp/eagle when using sample.{e}")
+                self.local_num_experts, self.expert_map = determine_default_expert_map(
+                    self.global_num_experts, self.ep_size, self.ep_rank,
+                    self.global_redundant_expert_num)
+                self.log2phy = determine_default_log2phy_map(
+                    self.global_num_experts, self.ep_size, self.ep_rank,
+                    self.global_redundant_expert_num).npu()
+            if self.expert_map is not None and isinstance(
+                    self.expert_map, torch.Tensor):
+                logger.info_once(
+                    "[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
+                    " number of experts: %s/%s. Experts local to global index map:"
+                    " %s.", self.ep_rank, self.ep_size, self.local_num_experts,
+                    self.global_num_experts,
+                    get_compressed_expert_map(self.expert_map))
+        else:
+            # init moe.
+            if vllm_version_is("0.11.0"):
+                self.local_num_experts, self.expert_map = determine_expert_map(
+                    self.ep_size, self.ep_rank, self.global_num_experts)
+            else:
+                self.local_num_experts, self.expert_map, _ = determine_expert_map(
+                    self.ep_size, self.ep_rank, self.global_num_experts)
+            # dynamic eplb initializing with not expert_map_path
+            if self.dynamic_eplb:
+                self.global_redundant_expert_num = ascend_config.init_redundancy_expert
+                self.local_num_experts, self.expert_map = determine_default_expert_map(
+                    self.global_num_experts, self.ep_size, self.ep_rank,
+                    self.global_redundant_expert_num)
+                self.log2phy = determine_default_log2phy_map(
+                    self.global_num_experts, self.ep_size, self.ep_rank,
+                    self.global_redundant_expert_num).npu()
+            if self.expert_map is not None and isinstance(
+                    self.expert_map, torch.Tensor):
+                logger.info_once(
+                    "[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
+                    " number of experts: %s/%s. Experts local to global index map:"
+                    " %s.", self.ep_rank, self.ep_size, self.local_num_experts,
+                    self.global_num_experts,
+                    get_compressed_expert_map(self.expert_map))
+        local_num_experts = (torch.sum(
+            self.expert_map != -1) if self.expert_map is not None else
+                             self.global_num_experts)
+        if self.dynamic_eplb:
+            self.moe_load = torch.zeros(local_num_experts, dtype=torch.int64)
+
+        self.moe_config.num_experts = self.global_num_experts
+        self.moe_config.num_local_experts = self.local_num_experts
+        self.moe_config.original_num_experts = num_experts
+
+        moe_quant_params = {
+            "num_experts": local_num_experts,
+            "hidden_size": self.hidden_size,
+            "intermediate_size_per_partition":
+            self.intermediate_size_per_partition,
+            "params_dtype": self.params_dtype,
+            "weight_loader": self.weight_loader,
+        }
+        # need full intermediate size pre-sharding for WNA16 act order
+        if (self.quant_method.__class__.__name__
+                in ("GPTQMarlinMoEMethod", "CompressedTensorsWNA16MoEMethod")):
+            moe_quant_params["intermediate_size_full"] = intermediate_size
+        self.quant_method.create_weights(layer=self, **moe_quant_params)
+
+        self.enable_shared_expert_dp = ascend_config.enable_shared_expert_dp
+
+        setup_moe_comm_method(self.moe_config)
+
+    def update_expert_map(self, new_expert_map):
+        self.expert_map = new_expert_map
+
+    def get_map(self):
+        return self.expert_map
+
+    def get_log2phy_map(self):
+        return self.logical_to_physical_map
+
+    def clear_moe_load(self):
+        if self.moe_load is not None:
+            self.moe_load.zero_()
+
+    def maybe_all_reduce_tensor_model_parallel(
+            self, final_hidden_states: torch.Tensor):
+        """NOTE(Yizhou): This is to override the parent class method. In `mc2commimpl`,
+        and `alltoallcommimpl`, we do not need to all-reduce the final outputs since
+        the outputs are already aggregated across tensor parallel ranks in the
+        `finalize` function. In `allgathercommimpl`, we still need to all-reduce the
+        outputs since each rank only has partial outputs.
+        """
+        return torch.ops.vllm.maybe_all_reduce_tensor_model_parallel(
+            final_hidden_states)
+
+    def forward_impl(self, hidden_states: torch.Tensor,
+                     router_logits: torch.Tensor):
+        assert self.quant_method is not None
+
+        # For w8a8 dynamic we can do npu_dynamic_quant and gate in parallel.
+        quantized_x_for_share, dynamic_scale_for_share = None, None
+
+        forward_context = get_forward_context()
+
+        # Load balancing for token distribution among experts in dummy_run
+        # TODO: The community only considers load balancing when DP > 1.
+        # This approach may overlook some extreme scenarios.
+        enable_force_load_balance = forward_context.in_profile_run
+
+        forward_context = get_forward_context()
+        hidden_states, router_logits, mc2_mask, context_metadata = forward_context.moe_comm_method.prepare(
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+            replace_allreduce=forward_context.sp_enabled,
+            enable_shared_expert_dp=self.enable_shared_expert_dp)
+
+        # Matrix multiply.
+        final_hidden_states = self.quant_method.apply(
+            layer=self,
+            x=hidden_states,
+            router_logits=router_logits,
+            top_k=self.top_k,
+            renormalize=self.renormalize,
+            use_grouped_topk=self.use_grouped_topk,
+            global_num_experts=self.global_num_experts,
+            expert_map=self.expert_map,
+            topk_group=self.topk_group,
+            num_expert_group=self.num_expert_group,
+            custom_routing_function=self.custom_routing_function,
+            scoring_func=self.scoring_func,
+            e_score_correction_bias=self.e_score_correction_bias,
+            activation=self.activation,
+            apply_router_weight_on_input=self.apply_router_weight_on_input,
+            quantized_x_for_share=quantized_x_for_share,
+            dynamic_scale_for_share=dynamic_scale_for_share,
+            shared_experts=None,
+            enable_force_load_balance=enable_force_load_balance,
+            log2phy=self.log2phy,
+            global_redundant_expert_num=self.global_redundant_expert_num,
+            mc2_mask=mc2_mask)
+
+        if isinstance(final_hidden_states, tuple):
+            final_hidden_states, group_list_type, expert_tokens = final_hidden_states
+
+        if self.dynamic_eplb:
+            self.moe_load += expert_tokens if group_list_type else \
+                torch.cat([expert_tokens[:1], expert_tokens[1:] - expert_tokens[:-1]])
+
+        final_hidden_states = forward_context.moe_comm_method.finalize(
+            hidden_states=final_hidden_states,
+            reduce_results=self.reduce_results,
+            context_metadata=context_metadata)
+
+        return final_hidden_states
+
+    def transpose_weight(self, loaded_weight, expert_data, shard_dim):
+        # Ensure training and inference weight shapes match during RL weight updates
+        if (
+            loaded_weight.shape[1] != expert_data.shape[1] and \
+            loaded_weight.shape[0] != expert_data.shape[0]
+        ):
+            shard_dim = int(not shard_dim)
+            loaded_weight = loaded_weight.transpose(0, 1).contiguous()
+        return loaded_weight, shard_dim
+
+    def _load_w13(self,
+                  expert_data: torch.Tensor,
+                  shard_dim: int,
+                  shard_id: str,
+                  loaded_weight: torch.Tensor,
+                  tp_rank: int,
+                  load_full: bool = False):
+        # Index the loaded weight for tp sharding.
+        # gate_up_proj: "MergedColumnParallel", so tp sharding on output_dim
+        loaded_weight, shard_dim = self.transpose_weight(
+            loaded_weight, expert_data, shard_dim)
+        shard_size = expert_data.shape[shard_dim] // 2
+        if not load_full:
+            loaded_weight = loaded_weight.narrow(shard_dim,
+                                                 shard_size * tp_rank,
+                                                 shard_size)
+        # Narrow parameter and load.
+        # w1, gate_proj: Load into first logical weight of w13.
+        if shard_id == "w1":
+            expert_data = expert_data.narrow(shard_dim, 0, shard_size)
+        # w3, up_proj: Load into second logical weight of w13.
+        else:
+            assert shard_id == "w3"
+            expert_data = expert_data.narrow(shard_dim, shard_size, shard_size)
+        expert_data.copy_(loaded_weight)
+
+    def _load_w2(self,
+                 expert_data: torch.Tensor,
+                 shard_dim: int,
+                 loaded_weight: torch.Tensor,
+                 tp_rank: int,
+                 load_full: bool = False):
+        # Index the loaded weight for tp sharding.
+        # down_proj: "RowParallel" so tp sharding on input_dim
+        # Narrow parameter and load.
+        loaded_weight, shard_dim = self.transpose_weight(
+            loaded_weight, expert_data, shard_dim)
+        shard_size = expert_data.shape[shard_dim]
+        if not load_full:
+            loaded_weight = loaded_weight.narrow(shard_dim,
+                                                 shard_size * tp_rank,
+                                                 shard_size)
+        # w2, down_proj: Load into only logical weight of w2.
+        expert_data.copy_(loaded_weight)
+
+
+class AscendSharedFusedMoE(SharedFusedMoE, AscendFusedMoE):
+
+    def __init__(
+        self,
+        shared_experts: torch.nn.Module,
+        use_overlapped: bool = True,
+        **kwargs,
+    ):
+        AscendFusedMoE.__init__(self, **kwargs)
+        self._shared_experts = shared_experts
+        self.use_overlapped = use_overlapped
+        self.shared_expert_stream = None
+        ascend_config = get_ascend_config()
+        self.multistream_overlap_shared_expert = ascend_config.multistream_overlap_shared_expert
+        if enable_sp():
+            logger.info_once(
+                "Sequence parallelism is enabled, shared experts are replicated for best performance."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        shared_out, fused_out = AscendFusedMoE.forward(
+            self,
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+        )
+        return shared_out, fused_out
+
+    def forward_impl(self, hidden_states: torch.Tensor,
+                     router_logits: torch.Tensor):
+        # Make sure the shared experts stream begins after hidden_states are ready.
+        if self.multistream_overlap_shared_expert:
+            shared_experts_calculation_stream().wait_stream(  # type: ignore
+                torch.npu.current_stream())
+        with npu_stream_switch(shared_experts_calculation_stream(),
+                               enabled=self.multistream_overlap_shared_expert):
+            # Use a separate stream to run shared experts.
+            # Note that currently we only support calculations in separate streams with aclgraph.
+            # Communication operations in another stream might cause unknown errors.
+            shared_out = self._shared_experts(hidden_states)
+
+        fused_output = AscendFusedMoE.forward_impl(
+            self,
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+        )
+        # Make sure the default stream waits for the shared experts stream to finish.
+        if self.multistream_overlap_shared_expert:
+            torch.npu.current_stream().wait_stream(
+                shared_experts_calculation_stream())
+        # NOTE: This is exactly the opposite of `maybe_all_reduce_tensor_model_parallel`
+        forward_context = get_forward_context()
+        moe_comm_type = forward_context.moe_comm_type
+        if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2} \
+                and not shared_expert_dp_enabled():
+            shared_out = tensor_model_parallel_all_reduce(shared_out)
+        return shared_out, fused_output

vllm_ascend/ops/fused_moe/moe_comm_method.py

@@ -0,0 +1,273 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, Optional
+
+import torch
+from vllm.config import get_current_vllm_config
+from vllm.forward_context import get_forward_context
+from vllm.model_executor.layers.fused_moe import FusedMoEConfig
+
+from vllm_ascend.ascend_forward_context import MoECommType
+from vllm_ascend.ops.fused_moe.moe_mlp import unified_apply_mlp
+from vllm_ascend.ops.fused_moe.prepare_finalize import (
+    PrepareAndFinalizeWithAll2All, PrepareAndFinalizeWithAllGather,
+    PrepareAndFinalizeWithMC2, PrepareAndFinalizeWithNaiveMulticast)
+from vllm_ascend.ops.fused_moe.token_dispatcher import (
+    TokenDispatcherWithAll2AllV, TokenDispatcherWithAllGather,
+    TokenDispatcherWithMC2, TokenDispatcherWithMoge)
+
+_MoECommMethods: Dict[Optional[MoECommType], MoECommMethod] = {}
+
+
+def get_moe_comm_method(
+        moe_comm_type: Optional[MoECommType]) -> Optional[MoECommMethod]:
+    return _MoECommMethods.get(moe_comm_type)
+
+
+def setup_moe_comm_method(moe_config):
+    _MoECommMethods[MoECommType.ALLTOALL] = AlltoAllCommImpl(moe_config)
+    _MoECommMethods[MoECommType.ALLGATHER] = AllGatherCommImpl(moe_config)
+    _MoECommMethods[MoECommType.MC2] = MC2CommImpl(moe_config)
+    _MoECommMethods[MoECommType.NAIVE_MULTICAST] = NaiveMulticastCommImpl(
+        moe_config)
+
+
+class MoECommMethod(ABC):
+    """Base class for MoE communication methods."""
+
+    def __init__(self, moe_config: FusedMoEConfig):
+        self.model_type = get_current_vllm_config(
+        ).model_config.hf_config.model_type
+        self.moe_config = moe_config
+
+        self.token_dispatcher = self._get_token_dispatcher()
+        self.prepare_finalize = self._get_prepare_finalize()
+
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        hidden_states, router_logits, mc2_mask, context_metadata = self.prepare_finalize.prepare(
+            hidden_states, router_logits, enable_shared_expert_dp,
+            replace_allreduce, gate)
+        return hidden_states, router_logits, mc2_mask, context_metadata
+
+    def finalize(self,
+                 hidden_states: torch.Tensor,
+                 reduce_results: bool,
+                 context_metadata: Optional[dict] = None) -> torch.Tensor:
+        hidden_states = self.prepare_finalize.finalize(hidden_states,
+                                                       reduce_results,
+                                                       context_metadata)
+        return hidden_states
+
+    def fused_experts(
+            self,
+            hidden_states: torch.Tensor,
+            w1: torch.Tensor,
+            w2: torch.Tensor,
+            topk_weights: torch.Tensor,
+            topk_ids: torch.Tensor,
+            activation: str = "silu",
+            apply_router_weight_on_input: bool = False,
+            use_int8_w8a8: bool = False,
+            use_int4_w4a8: bool = False,
+            global_num_experts: Optional[int] = None,
+            expert_map: Optional[torch.Tensor] = None,
+            w1_scale: Optional[torch.Tensor] = None,
+            w2_scale: Optional[torch.Tensor] = None,
+            w1_scale_bias: torch.Tensor = None,
+            w2_scale_bias: torch.Tensor = None,
+            # For TorchAir graph
+            is_torchair: bool = False,
+            # For Cube/Vector parallel
+            shared_experts: Optional[Any] = None,
+            quantized_x_for_share: Optional[Any] = None,
+            dynamic_scale_for_share: Optional[Any] = None,
+            # For load balance
+            log2phy: torch.Tensor = None,
+            global_redundant_expert_num: int = 0,
+            need_trans: bool = False,
+            dynamic_eplb: bool = False,
+            mc2_mask: torch.Tensor = None):
+        # Check constraints
+        assert hidden_states.dtype in [
+            torch.float32, torch.float16, torch.bfloat16
+        ]
+
+        moe_comm_method = get_forward_context().moe_comm_method
+        assert moe_comm_method is not None, "Missing communication context"
+
+        results = self.token_dispatcher.token_dispatch(
+            hidden_states=hidden_states,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            expert_map=expert_map,
+            log2phy=log2phy,
+            global_redundant_expert_num=global_redundant_expert_num,
+            shared_experts=shared_experts,
+            quantized_x_for_share=quantized_x_for_share,
+            dynamic_scale_for_share=dynamic_scale_for_share,
+            mc2_mask=mc2_mask,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            with_quant=use_int8_w8a8 or use_int4_w4a8)
+
+        permuted_hidden_states, expert_tokens, dynamic_scale, group_list_type, topk_scales, context_metadata = \
+            results["hidden_states"], results["group_list"], results.get("dynamic_scale"), results["group_list_type"], results.get("topk_scales"), results.get("context_metadata")
+
+        mlp_output = unified_apply_mlp(hidden_states=permuted_hidden_states,
+                                       w1=w1,
+                                       w1_scale=w1_scale,
+                                       w2=w2,
+                                       w2_scale=w2_scale,
+                                       group_list=expert_tokens,
+                                       dynamic_scale=dynamic_scale,
+                                       group_list_type=group_list_type,
+                                       w1_scale_bias=w1_scale_bias,
+                                       w2_scale_bias=w2_scale_bias,
+                                       topk_scales=topk_scales,
+                                       with_quant=use_int8_w8a8
+                                       or use_int4_w4a8,
+                                       fusion=use_int8_w8a8,
+                                       need_trans=need_trans,
+                                       dynamic_eplb=dynamic_eplb)
+
+        final_hidden_states = self.token_dispatcher.token_combine(
+            hidden_states=mlp_output, context_metadata=context_metadata)
+
+        if dynamic_eplb:
+            return (final_hidden_states, group_list_type, expert_tokens)
+
+        return final_hidden_states
+
+    @abstractmethod
+    def _get_token_dispatcher(self):
+        raise NotImplementedError(
+            "_get_token_dispatcher function not implemented.")
+
+    @abstractmethod
+    def _get_prepare_finalize(self):
+        raise NotImplementedError(
+            "_get_prepare_finalize function not implemented.")
+
+
+class AllGatherCommImpl(MoECommMethod):
+    """This implementation is the same as NativeAllGatherCommImpl,
+    but uses NPU-specific ops for better performance.
+
+    This implementation should be compatible with all scenarios, and
+    thus it is the default implementation for MoE communication methods.
+    It uses `torch_npu.npu_moe_init_routing_v2` for pre-processing
+    and `torch_npu.npu_moe_token_unpermute` for post-processing
+    to handle the token-to-expert mapping and communication efficiently.
+
+    NOTE(Yizhou): TBH, it is really weird that we were supposed to use
+    `torch_npu.npu_moe_init_routing_v2` and `torch_npu.npu_moe_finalize_routing`
+    or `torch_npu.npu_moe_token_permute` and `torch_npu.npu_moe_token_unpermute`
+    for pre-processing and post-processing, respectively.
+    But `npu_moe_finalize_routing` will lead to accuracy issues so we have to
+    use `torch_npu.npu_moe_token_unpermute` instead.
+    This is a workaround and should be removed after the issue is fixed.
+    """
+
+    def _get_token_dispatcher(self):
+        if self.model_type == "PanguProMoE":
+            return TokenDispatcherWithMoge(
+                top_k=self.moe_config.experts_per_token,
+                num_experts=self.moe_config.num_experts,
+                num_local_experts=self.moe_config.num_local_experts)
+        else:
+            return TokenDispatcherWithAllGather(
+                top_k=self.moe_config.experts_per_token,
+                num_experts=self.moe_config.num_experts,
+                num_local_experts=self.moe_config.num_local_experts)
+
+    def _get_prepare_finalize(self):
+        return PrepareAndFinalizeWithAllGather(self.moe_config)
+
+
+class MC2CommImpl(MoECommMethod):
+    """This implementation is for the scenarios listed below:
+    1. `enable_expert_parallel=True`.
+    2. `npu_moe_distribute_dispatch` and `npu_moe_distribute_combine` are available.
+    3. `enable_expert_parallel=False` is not supported.
+    
+    This implementation uses the MC2 communication method, which is optimized for
+    Communication and Computation parallelism on Ascend devices.
+    """
+
+    def _get_token_dispatcher(self):
+        return TokenDispatcherWithMC2()
+
+    def _get_prepare_finalize(self):
+        return PrepareAndFinalizeWithMC2(self.moe_config)
+
+
+class AlltoAllCommImpl(MoECommMethod):
+    """This implementation is for the scenarios listed below:
+    1. `enable_expert_parallel=True`.
+    2. `npu_grouped_matmul` is available.
+
+    This implementation uses all-to-all communication to exchange tokens
+    between data parallel ranks before and after the MLP computation. It should
+    have better performance than AllGatherCommImpl when DP size > 1.
+    """
+
+    def _get_token_dispatcher(self):
+        return TokenDispatcherWithAll2AllV(
+            top_k=self.moe_config.experts_per_token,
+            num_experts=self.moe_config.num_experts,
+            num_local_experts=self.moe_config.num_local_experts)
+
+    def _get_prepare_finalize(self):
+        return PrepareAndFinalizeWithAll2All(self.moe_config)
+
+
+class NaiveMulticastCommImpl(MoECommMethod):
+    """This implementation is the same as NativeAllGatherCommImpl,
+    but uses NPU-specific ops for better performance.
+
+    This implementation should be compatible with all scenarios, and
+    thus it is the default implementation for MoE communication methods.
+    It uses `torch_npu.npu_moe_init_routing_v2` for pre-processing
+    and `torch_npu.npu_moe_token_unpermute` for post-processing
+    to handle the token-to-expert mapping and communication efficiently.
+
+    NOTE(Yizhou): TBH, it is really weird that we were supposed to use
+    `torch_npu.npu_moe_init_routing_v2` and `torch_npu.npu_moe_finalize_routing`
+    or `torch_npu.npu_moe_token_permute` and `torch_npu.npu_moe_token_unpermute`
+    for pre-processing and post-processing, respectively.
+    But `npu_moe_finalize_routing` will lead to accuracy issues so we have to
+    use `torch_npu.npu_moe_token_unpermute` instead.
+    This is a workaround and should be removed after the issue is fixed.
+    """
+
+    def _get_token_dispatcher(self):
+        return TokenDispatcherWithAllGather(
+            top_k=self.moe_config.experts_per_token,
+            num_experts=self.moe_config.num_experts,
+            num_local_experts=self.moe_config.num_local_experts)
+
+    def _get_prepare_finalize(self):
+        return PrepareAndFinalizeWithNaiveMulticast(self.moe_config)

vllm_ascend/ops/fused_moe/moe_mlp.py

@@ -0,0 +1,258 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+
+from typing import Optional
+
+import torch
+import torch_npu
+from torch.nn.functional import pad
+from vllm.forward_context import get_forward_context
+
+from vllm_ascend.ascend_forward_context import MoECommType
+from vllm_ascend.utils import dispose_tensor, is_310p
+
+
+def cumsum_group_list(group_list: torch.Tensor,
+                      group_list_type: int,
+                      active_num: int = 0,
+                      expert_num: int = 0) -> torch.Tensor:
+    if group_list_type not in [0, 1, 2]:
+        raise ValueError(
+            f"group_list_type should be in [0, 1, 2], but received {group_list_type}"
+        )
+
+    if group_list_type == 0:
+        return group_list
+    if group_list_type == 1:
+        return group_list.cumsum(dim=0)
+
+    experts = pad(group_list[:, 0], (1, 0))
+    tokens = pad(group_list[:, 1].cumsum(dim=0), (1, 0))
+    cumsum_group_list = torch.full(size=(expert_num, ),
+                                   fill_value=active_num,
+                                   dtype=group_list.dtype,
+                                   device=group_list.device)
+
+    for i, (start, end) in enumerate(zip(experts[:-1], experts[1:])):
+        if end > start:
+            cumsum_group_list[start:end] = tokens[i]
+
+    return cumsum_group_list
+
+
+def quant_apply_mlp(hidden_states: torch.Tensor,
+                    w1: torch.Tensor,
+                    w1_scale: torch.Tensor,
+                    w2: torch.Tensor,
+                    w2_scale: torch.Tensor,
+                    group_list: torch.Tensor,
+                    group_list_type: int = 1,
+                    dynamic_scale: torch.Tensor = None,
+                    w1_scale_bias: torch.Tensor = None,
+                    w2_scale_bias: torch.Tensor = None,
+                    fusion: bool = False,
+                    dynamic_eplb: bool = False) -> torch.Tensor:
+    if dynamic_scale is None:
+        unquantized_hidden_states = hidden_states
+        hidden_states, pertoken_scale = torch_npu.npu_dynamic_quant(
+            hidden_states)
+        # Dispose the original unquantized hidden states
+        # to save npu memory because they're no longer used.
+        dispose_tensor(unquantized_hidden_states)
+    else:
+        pertoken_scale = dynamic_scale
+
+    bias1, bias2 = None, None
+    _output_dtype = w2_scale.dtype
+
+    weight_prefetch_method = get_forward_context().weight_prefetch_method
+    if weight_prefetch_method:
+        weight_prefetch_method.maybe_prefetch_moe_weight_postprocess(
+            hidden_states)
+    is_mc2 = get_forward_context().moe_comm_type == MoECommType.MC2
+    if w1_scale_bias is None and is_mc2:
+        if fusion and not dynamic_eplb:
+            # gmm1: gate_up_proj & act_fn: swiglu
+            hidden_states, swiglu_out_scale, _ = torch_npu.npu_grouped_matmul_swiglu_quant(
+                x=hidden_states,
+                weight=w1,
+                group_list=cumsum_group_list(group_list, group_list_type),
+                weight_scale=w1_scale,
+                x_scale=pertoken_scale)
+        else:
+            if w1_scale.dtype != torch.float32:
+                w1_scale = w1_scale.to(torch.float32)
+            # gmm1: gate_up_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[w1],
+                split_item=3,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=torch.int32)[0]
+            # act_fn: swiglu
+            hidden_states, swiglu_out_scale = torch_npu.npu_dequant_swiglu_quant(
+                x=hidden_states,
+                weight_scale=w1_scale,
+                activation_scale=pertoken_scale,
+                bias=None,
+                quant_scale=None,
+                quant_offset=None,
+                group_index=group_list,
+                activate_left=True,
+                quant_mode=1,
+            )
+        # gmm2: down_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w2],
+            scale=[w2_scale],
+            per_token_scale=[swiglu_out_scale],
+            split_item=2,
+            group_list_type=group_list_type,
+            group_type=0,
+            group_list=group_list,
+            output_dtype=w2_scale.dtype)[0]
+    else:
+        if w1_scale_bias is not None:
+            if group_list_type == 0:
+                group_list = torch.cat(
+                    [group_list[:1],
+                     torch.diff(group_list, dim=0)])
+                group_list_type = 1
+            bias1 = [w1_scale_bias] if not fusion else w1_scale_bias
+            bias2 = [w2_scale_bias]
+            # TODO w4a8 scene: dynamic acquisition of dtype in the future
+            _output_dtype = torch.bfloat16
+
+        if fusion and not dynamic_eplb:
+            # gmm1: gate_up_proj & act_fn: swiglu
+            hidden_states, swiglu_out_scale, _ = torch_npu.npu_grouped_matmul_swiglu_quant(
+                x=hidden_states,
+                weight=w1,
+                bias=bias1,
+                group_list=cumsum_group_list(group_list, group_list_type),
+                weight_scale=w1_scale,
+                x_scale=pertoken_scale)
+        else:
+            # gmm1: gate_up_proj
+            hidden_states = torch_npu.npu_grouped_matmul(
+                x=[hidden_states],
+                weight=[w1],
+                scale=[w1_scale.to(w2_scale.dtype)],
+                bias=bias1,
+                per_token_scale=[pertoken_scale],
+                split_item=2,
+                group_list_type=group_list_type,
+                group_type=0,
+                group_list=group_list,
+                output_dtype=_output_dtype)[0]
+            # act_fn: swiglu
+            hidden_states = torch_npu.npu_swiglu(hidden_states)
+            hidden_states, swiglu_out_scale = torch_npu.npu_dynamic_quant(
+                hidden_states)
+        # gmm2: down_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w2],
+            scale=[w2_scale],
+            bias=bias2,
+            per_token_scale=[swiglu_out_scale],
+            split_item=2,
+            group_list_type=group_list_type,
+            group_type=0,
+            group_list=group_list,
+            output_dtype=_output_dtype)[0]
+    return hidden_states
+
+
+def unquant_apply_mlp(hidden_states: torch.Tensor,
+                      w1: torch.Tensor,
+                      w2: torch.Tensor,
+                      group_list: torch.Tensor,
+                      group_list_type: int = 1,
+                      topk_scales: Optional[torch.Tensor] = None,
+                      need_trans: bool = True) -> torch.Tensor:
+
+    if need_trans:
+        w1 = w1.transpose(1, 2)
+        w2 = w2.transpose(1, 2)
+
+    gate_up_out = torch_npu.npu_grouped_matmul(
+        x=[hidden_states],
+        weight=[w1],
+        split_item=2,
+        group_list_type=group_list_type,
+        group_type=0,
+        group_list=group_list,
+    )[0]
+    if is_310p():
+        gate_up_out = torch_npu.npu_swiglu(gate_up_out.to(torch.float32)).to(
+            torch.float16)
+    else:
+        gate_up_out = torch_npu.npu_swiglu(gate_up_out)
+
+    if topk_scales is not None:
+        gate_up_out *= topk_scales
+
+    hidden_states = torch_npu.npu_grouped_matmul(
+        x=[gate_up_out],
+        weight=[w2],
+        split_item=2,
+        group_list_type=group_list_type,
+        group_type=0,
+        group_list=group_list,
+    )[0]
+    return hidden_states
+
+
+def unified_apply_mlp(hidden_states: torch.Tensor,
+                      w1: torch.Tensor,
+                      w1_scale: torch.Tensor,
+                      w2: torch.Tensor,
+                      w2_scale: torch.Tensor,
+                      group_list: torch.Tensor,
+                      dynamic_scale: torch.Tensor = None,
+                      group_list_type: int = 1,
+                      w1_scale_bias: torch.Tensor = None,
+                      w2_scale_bias: torch.Tensor = None,
+                      topk_scales: Optional[torch.Tensor] = None,
+                      with_quant: bool = False,
+                      fusion: bool = False,
+                      need_trans: bool = True,
+                      dynamic_eplb: bool = False) -> torch.Tensor:
+    if with_quant:
+        return quant_apply_mlp(hidden_states=hidden_states,
+                               w1=w1,
+                               w1_scale=w1_scale,
+                               w2=w2,
+                               w2_scale=w2_scale,
+                               group_list=group_list,
+                               dynamic_scale=dynamic_scale,
+                               group_list_type=group_list_type,
+                               w1_scale_bias=w1_scale_bias,
+                               w2_scale_bias=w2_scale_bias,
+                               fusion=fusion,
+                               dynamic_eplb=dynamic_eplb)
+    else:
+        return unquant_apply_mlp(hidden_states=hidden_states,
+                                 w1=w1,
+                                 w2=w2,
+                                 group_list=group_list,
+                                 group_list_type=group_list_type,
+                                 topk_scales=topk_scales,
+                                 need_trans=need_trans)

vllm_ascend/ops/fused_moe/prepare_finalize.py

@@ -0,0 +1,530 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+
+from abc import ABC, abstractmethod
+from typing import Optional
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from vllm.distributed import tensor_model_parallel_all_reduce
+from vllm.distributed.parallel_state import (
+    get_dp_group, get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size)
+from vllm.forward_context import get_forward_context
+from vllm.model_executor.layers.fused_moe import FusedMoEConfig
+
+from vllm_ascend.utils import enable_sp, get_rm_router_logits_state
+
+
+class PrepareAndFinalize(ABC):
+    """
+    Abstract base class for MoE (Mixture-of-Experts) tensor preparation and finalization
+    in distributed environments. Subclasses implement specific communication strategies
+    (e.g., AllGather, All2All, MC2, Naive Multicast) to handle tensor padding, slicing,
+    broadcasting, and reduction across TP/DP/EP groups.
+
+    Attributes:
+        moe_config (FusedMoEConfig): Configuration object containing TP/DP/EP group info,
+                                     sizes, ranks, and communication settings.
+    """
+
+    def __init__(self, moe_config: FusedMoEConfig):
+        self.moe_config = moe_config
+        is_deepseek_v3_r1 = self.moe_config.original_num_experts == 256
+        self.rm_router_logits = get_rm_router_logits_state(
+            self.moe_config.ep_size, self.moe_config.dp_size,
+            is_deepseek_v3_r1)
+
+    @abstractmethod
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Prepare tensors before MoE computation. May involve:
+          - Padding to align communication boundaries
+          - Slicing across tensor-parallel ranks
+          - Broadcasting across data-parallel ranks
+          - Recomputing router logits if needed
+
+        Args:
+            hidden_states (torch.Tensor): Input features, shape [num_tokens, hidden_size]
+            router_logits (torch.Tensor): Router outputs, shape [num_tokens, num_experts]
+            enable_shared_expert_dp (bool): Skip DP communication for shared experts
+            replace_allreduce (bool): Bypass default all-reduce behavior
+            gate (nn.Module, optional): Gate network to recompute router_logits if needed
+
+        Returns:
+            Tuple of:
+                - processed hidden_states (may be padded/sliced/broadcasted)
+                - processed router_logits (may be recomputed or broadcasted)
+                - optional communication mask (e.g., mc2_mask for sparse ops)
+                - optional context metadata (e.g., saved split_hidden_states for finalization)
+        """
+        raise NotImplementedError("Prepare not implemented.")
+
+    def finalize(self,
+                 hidden_states: torch.Tensor,
+                 reduce_results: bool,
+                 context_metadata: Optional[dict] = None) -> torch.Tensor:
+        """
+        Finalize MoE output. May involve:
+          - Gathering sliced tensors across TP ranks
+          - Reducing or scattering across DP ranks
+          - Unpadding to original token count
+          - Applying all-reduce across TP/EP if requested
+
+        Args:
+            hidden_states (torch.Tensor): MoE layer output, possibly padded or sliced
+            reduce_results (bool): Whether to apply all-reduce across TP/EP groups
+
+        Returns:
+            torch.Tensor: Final output with shape [original_num_tokens, hidden_size]
+        """
+        raise NotImplementedError("Finalize function not implemented.")
+
+
+class PrepareAndFinalizeWithAll2All(PrepareAndFinalize):
+    """
+    MoE communication strategy using All-to-All style slicing.
+    Similar to MC2 but does not use mc2_mask; instead pads to TP size for uniform slicing.
+    Will be used when num_tokens exceed mc2's limitation (512 tokens/rank).
+    """
+
+    def __init__(self, moe_config: FusedMoEConfig):
+        super().__init__(moe_config)
+        self._restore_tp_across_dp()
+
+    def _restore_tp_across_dp(self):
+        """Restore original TP configuration (same as MC2)."""
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Preparation steps:
+          1. Pad hidden_states and router_logits to next multiple of TP size.
+          2. If TP > 1, split along token dim and select current TP rank's slice.
+          3. Save splits for later all-gather in finalize.
+
+        Skips if `enable_shared_expert_dp` or `replace_allreduce` is True.
+
+        Returns:
+            Tuple of (hidden_states, router_logits, None, context_metadata) — no mask used in All2All.
+        """
+        self.replace_allreduce = replace_allreduce
+        self.enable_shared_expert_dp = enable_shared_expert_dp
+        split_hidden_states = None
+
+        if not (self.replace_allreduce or self.enable_shared_expert_dp):
+            self.num_tokens, _ = hidden_states.shape
+            pad_size = self.tp_size - self.num_tokens  # Pad to TP size (cyclic)
+
+            if pad_size > 0:
+                hidden_states = nn.functional.pad(hidden_states,
+                                                  (0, 0, 0, pad_size))
+                router_logits = nn.functional.pad(router_logits,
+                                                  (0, 0, 0, pad_size))
+
+            if self.tp_size > 1:
+                split_hidden_states = torch.tensor_split(hidden_states,
+                                                         self.tp_size,
+                                                         dim=0)
+                split_router_logits = torch.tensor_split(router_logits,
+                                                         self.tp_size,
+                                                         dim=0)
+
+                hidden_states = split_hidden_states[self.tp_rank]
+                router_logits = split_router_logits[self.tp_rank]
+
+        context_metadata = {"split_hidden_states": split_hidden_states}
+
+        return hidden_states, router_logits, None, context_metadata
+
+    def finalize(self,
+                 hidden_states: torch.Tensor,
+                 reduce_results: bool,
+                 context_metadata: Optional[dict] = None) -> torch.Tensor:
+        """
+        Finalization steps:
+          1. If TP > 1, all-gather slices to reconstruct full tensor.
+          2. Unpad to original token count.
+          3. Return [original_num_tokens, hidden_size] tensor.
+
+        Skips if `enable_shared_expert_dp` or `replace_allreduce` is True.
+        """
+        assert context_metadata is not None
+
+        split_hidden_states = context_metadata["split_hidden_states"]
+        if not (self.enable_shared_expert_dp or self.replace_allreduce):
+            if self.tp_size > 1:
+                dist.all_gather(list(split_hidden_states), hidden_states,
+                                self.moe_config.tp_group.device_group)
+                hidden_states = torch.cat(split_hidden_states, dim=0)
+
+            if self.num_tokens < hidden_states.shape[0]:
+                hidden_states = hidden_states[:self.num_tokens]
+
+        return hidden_states
+
+
+class PrepareAndFinalizeWithMC2(PrepareAndFinalizeWithAll2All):
+    """
+    MoE communication strategy using MC2, which is based on All2All. Hence, it inherits
+    All2All and share the same finalize method.
+    Designed for Ascend or environments requiring explicit padding and slicing control.
+    Relies on `mc2_mask` and `padded_num_tokens` from forward_context for alignment.
+    """
+
+    def __init__(self, moe_config: FusedMoEConfig):
+        super().__init__(moe_config)
+        self._restore_tp_across_dp()
+
+    def _restore_tp_across_dp(self):
+        """
+        Restore original TP configuration.
+        vLLM flattens TP and DP into a single dimension; this method recovers
+        the true TP world size and rank for correct tensor slicing.
+        """
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Preparation steps:
+          1. Fetch `mc2_mask` and target padding length from forward context.
+          2. Pad `hidden_states` and `router_logits` to target length if needed.
+          3. If TP > 1, split tensors along token dimension and select current TP rank's slice.
+          4. Split and return corresponding `mc2_mask`.
+
+        Skips padding/slicing if `enable_shared_expert_dp` or `replace_allreduce` is True.
+
+        Returns:
+            Tuple of (hidden_states, router_logits, mc2_mask, context_metadata), possibly sliced/padded.
+        """
+        self.replace_allreduce = replace_allreduce
+        self.enable_shared_expert_dp = enable_shared_expert_dp
+        split_hidden_states = None
+        forward_context = get_forward_context()
+        mc2_mask = forward_context.mc2_mask
+        if self.tp_size > 1:
+            # Also slice mc2_mask
+            split_mc2_mask = torch.tensor_split(mc2_mask, self.tp_size, dim=0)
+            mc2_mask = split_mc2_mask[self.tp_rank]
+
+        if not self.replace_allreduce:
+            self.num_tokens, _ = hidden_states.shape
+            target_pad_length = forward_context.padded_num_tokens
+            pad_size = target_pad_length - self.num_tokens
+
+            # Pad if necessary (unless shared expert DP is enabled)
+            if pad_size > 0 and not self.enable_shared_expert_dp:
+                hidden_states = nn.functional.pad(hidden_states,
+                                                  (0, 0, 0, pad_size))
+                router_logits = nn.functional.pad(router_logits,
+                                                  (0, 0, 0, pad_size))
+
+            # Slice across TP ranks
+            if self.tp_size > 1 and not self.enable_shared_expert_dp:
+                split_hidden_states = torch.tensor_split(hidden_states,
+                                                         self.tp_size,
+                                                         dim=0)
+                split_router_logits = torch.tensor_split(router_logits,
+                                                         self.tp_size,
+                                                         dim=0)
+                hidden_states = split_hidden_states[self.tp_rank]
+                router_logits = split_router_logits[self.tp_rank]
+
+        context_metadata = {"split_hidden_states": split_hidden_states}
+
+        return hidden_states, router_logits, mc2_mask, context_metadata
+
+
+class PrepareAndFinalizeWithAllGather(PrepareAndFinalize):
+    """
+    MoE communication strategy using All-Gather + Reduce-Scatter on EP group.
+    There are two sets of prepare and finalize:
+    1. _prepare_with_dp_group/_finalize_with_dp_group: When sequence parallelism is not enabled,
+    we gather inputs across DP ranks before MoE, scatter outputs after.
+    The communication and calculation process is as follows (AG, AR and RS
+    are abbreviations for All-Gather, All-Reduce and Reduce-Scatter, respectively):
+
+    Attn → TP AR → DP AG → MoE → DP RS → TP AR
+
+    2. _prepare_with_ep_group/_finalize_with_ep_group: When sequence parallelism is enabled,
+    the above process becomes:
+
+    TP AG → Attn → TP RS → TP AG → DP AG → MoE → DP RS → TP RS
+
+    This strategy further combines TP AG + DP AG into EP All-Gather and TP RS + DP RS
+    into EP Reduce-Scatter to improve communication performance. The optimized process is as follows:
+
+    TP AG → Attn → TP RS → EP AG → MoE → EP RS
+    """
+
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Preparation steps:
+          AllGather hidden_states and router_logits to form global tensors.
+
+        Returns:
+            Tuple of (global_hidden_states, global_router_logits, None)
+        """
+        if enable_sp():
+            return self._prepare_with_ep_group(hidden_states, router_logits)
+
+        return self._prepare_with_dp_group(hidden_states, router_logits,
+                                           enable_shared_expert_dp,
+                                           replace_allreduce, gate)
+
+    def _prepare_with_ep_group(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
+            hidden_states, True, True)
+        router_logits = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
+            router_logits, True, True)
+
+        return hidden_states, router_logits, None, None
+
+    def _prepare_with_dp_group(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Preparation steps:
+          1. Fetch max token count across DP group from forward context.
+          2. Pad local tensors to that size.
+          3. All-gather across DP group to form global input tensor.
+
+        Returns:
+            Tuple of (global_hidden_states, global_router_logits, None, None)
+        """
+        self.enable_shared_expert_dp = enable_shared_expert_dp
+        if self.moe_config.dp_size > 1:
+            forward_context = get_forward_context()
+            max_tokens_across_dp = forward_context.max_tokens_across_dp
+
+            self.num_tokens = hidden_states.shape[0]
+            pad_size = max_tokens_across_dp - self.num_tokens
+            if pad_size > 0:
+                hidden_states = nn.functional.pad(hidden_states,
+                                                  (0, 0, 0, pad_size))
+                if not self.rm_router_logits:
+                    router_logits = nn.functional.pad(router_logits,
+                                                      (0, 0, 0, pad_size))
+
+            # All-gather across DP group
+            hidden_states = self.moe_config.dp_group.all_gather(
+                hidden_states, 0)
+            if self.rm_router_logits:
+                router_logits, _ = gate(hidden_states)  # Recompute globally
+            else:
+                router_logits = self.moe_config.dp_group.all_gather(
+                    router_logits, 0)
+        return hidden_states, router_logits, None, None
+
+    def finalize(self,
+                 hidden_states: torch.Tensor,
+                 reduce_results: bool,
+                 context_metadata: Optional[dict] = None) -> torch.Tensor:
+        """
+        Finalization steps:
+          Reduce Scatter hidden states.
+
+        Returns:
+            Tensor with shape [local_num_tokens, hidden_size]
+        """
+        if enable_sp():
+            return self._finalize_with_ep_group(hidden_states)
+
+        return self._finalize_with_dp_group(hidden_states, reduce_results)
+
+    def _finalize_with_ep_group(self,
+                                hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Argument `reduce_results` is not needed in this func. Given sequence parallelism is enabled:
+        1. Reduce_results is False usually happens when models have shared experts and need to
+        allreduce hidden states after results of shared experts and routed experts are added in FusedMoe.
+        We do reduce scatter for hidden states here, then skip allreudce in FusedMoe and add it to the
+        result of shared experts.
+        2 Reduce_results is True usually happens when model has no shared experts. We still do reduce scatter
+        here, then skip allreudce in FusedMoe.
+        """
+        hidden_states = torch.ops.vllm.maybe_pad_and_reduce(
+            hidden_states, True)
+
+        return hidden_states
+
+    def _finalize_with_dp_group(self, hidden_states: torch.Tensor,
+                                reduce_results: bool) -> torch.Tensor:
+        """
+        Finalization steps:
+          1. If DP > 1 and not shared expert, reduce-scatter output across DP group.
+          2. Slice to original local token count.
+          3. If `reduce_results=True` and TP/EP > 1, apply tensor_model_parallel_all_reduce.
+
+        Returns:
+            Tensor with shape [original_local_num_tokens, hidden_size]
+        """
+        if self.moe_config.dp_size > 1 and not self.enable_shared_expert_dp:
+            hidden_states = get_dp_group().reduce_scatter(hidden_states, 0)
+            hidden_states = hidden_states[:self.num_tokens]
+
+        if reduce_results and (self.moe_config.tp_size > 1
+                               or self.moe_config.ep_size > 1):
+            hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+
+        return hidden_states
+
+
+class PrepareAndFinalizeWithNaiveMulticast(PrepareAndFinalize):
+    """
+    MoE communication strategy using Naive Multicast (point-to-point broadcast).
+    Will be used in prefill when using allgather in decode. Each DP rank broadcasts its slice to all others.
+    Uses `cu_tokens_across_dp_cpu` (cumulative tokens) to locate slice boundaries.
+    """
+
+    def _naive_multicast(self, x: torch.Tensor,
+                         cu_tokens_across_dp_cpu: torch.Tensor):
+        """
+        Naive multicast implementation:
+          1. Create global buffer sized by total tokens across DP.
+          2. Current rank copies its slice into its designated buffer region.
+          3. Each rank broadcasts its slice to all others via P2P.
+
+        Args:
+            x (torch.Tensor): Local tensor [local_tokens, hidden_size]
+            cu_tokens_across_dp_cpu (torch.Tensor): Cumulative token counts per DP rank
+
+        Returns:
+            torch.Tensor: Global tensor [total_tokens, hidden_size]
+        """
+        assert len(x.shape) == 2, "Input must be 2D [tokens, features]"
+        buffer = torch.empty((cu_tokens_across_dp_cpu[-1], x.size(1)),
+                             device=x.device,
+                             dtype=x.dtype)
+
+        # Copy local slice into buffer
+        start = 0 if self.moe_config.dp_rank == 0 else cu_tokens_across_dp_cpu[
+            self.moe_config.dp_rank - 1]
+        end = cu_tokens_across_dp_cpu[self.moe_config.dp_rank]
+        buffer[start:end, :].copy_(x)
+
+        # Broadcast each slice to all ranks
+        for idx in range(self.moe_config.dp_size):
+            start = 0 if idx == 0 else cu_tokens_across_dp_cpu[idx - 1]
+            end = cu_tokens_across_dp_cpu[idx]
+            get_dp_group().broadcast(buffer[start:end, :], idx)
+        return buffer
+
+    def prepare(
+        self,
+        hidden_states: torch.Tensor,
+        router_logits: torch.Tensor,
+        enable_shared_expert_dp: bool = False,
+        replace_allreduce: bool = False,
+        gate=None
+    ) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor],
+               Optional[torch.Tensor]]:
+        """
+        Preparation steps:
+          1. Fetch cumulative token boundaries from forward context.
+          2. Multicast hidden_states and router_logits to form global tensors.
+
+        Returns:
+            Tuple of (global_hidden_states, global_router_logits, None, None)
+        """
+        self.enable_shared_expert_dp = enable_shared_expert_dp
+
+        if self.moe_config.dp_size > 1:
+            self.cu_tokens_across_dp_cpu = get_forward_context(
+            ).dp_metadata.cu_tokens_across_sp(1)
+            hidden_states = self._naive_multicast(hidden_states,
+                                                  self.cu_tokens_across_dp_cpu)
+            if self.rm_router_logits:
+                router_logits, _ = gate(hidden_states)
+            else:
+                router_logits = self._naive_multicast(
+                    router_logits, self.cu_tokens_across_dp_cpu)
+
+        return hidden_states, router_logits, None, None
+
+    def finalize(self,
+                 hidden_states: torch.Tensor,
+                 reduce_results: bool,
+                 context_metadata: Optional[dict] = None) -> torch.Tensor:
+        """
+        Finalization steps:
+          1. If DP > 1 and not shared expert:
+               - All-reduce across DP
+               - Slice to current rank's token range using cu_tokens_across_dp_cpu
+          2. If `reduce_results=True` and TP/EP > 1, apply tensor_model_parallel_all_reduce.
+
+        Returns:
+            Tensor with shape [local_num_tokens, hidden_size]
+        """
+        if self.moe_config.dp_size > 1 and not self.enable_shared_expert_dp:
+            start = 0 if self.moe_config.dp_rank == 0 else self.cu_tokens_across_dp_cpu[
+                self.moe_config.dp_rank - 1]
+            end = self.cu_tokens_across_dp_cpu[self.moe_config.dp_rank]
+            hidden_states = get_dp_group().all_reduce(
+                hidden_states)  # Sum across DP
+            hidden_states = hidden_states[start:end, :]
+
+        if reduce_results and (self.moe_config.tp_size > 1
+                               or self.moe_config.ep_size > 1):
+            hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+
+        return hidden_states

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -0,0 +1,741 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright (c) 2024; NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from abc import ABC, abstractmethod
+from typing import Any, Optional
+
+import torch
+import torch_npu
+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 (AscendSocVersion, get_ascend_soc_version,
+                               is_hierarchical_communication_enabled)
+
+
+class MoETokenDispatcher(ABC):
+
+    def __init__(self, **kwargs) -> None:
+        """
+        Initialize the MoE Token Dispatcher.
+        """
+        self.top_k = kwargs.get("top_k", 0)
+        self.num_experts = kwargs.get("num_experts", 0)
+
+    @property
+    def ep_group(self):
+        """Get expert model parallel group."""
+        return get_ep_group().device_group
+
+    @property
+    def ep_rank(self):
+        return get_ep_group().rank_in_group
+
+    @property
+    def ep_size(self):
+        return get_ep_group().world_size
+
+    @abstractmethod
+    def token_dispatch(self,
+                       hidden_states: torch.Tensor,
+                       topk_weights: torch.Tensor,
+                       topk_ids: torch.Tensor,
+                       expert_map: Optional[torch.Tensor] = None,
+                       log2phy: Optional[torch.Tensor] = None,
+                       global_redundant_expert_num: int = 0,
+                       shared_experts: Optional[Any] = None,
+                       quantized_x_for_share: Optional[Any] = None,
+                       dynamic_scale_for_share: Optional[Any] = None,
+                       mc2_mask: Optional[torch.Tensor] = None,
+                       apply_router_weight_on_input: bool = False,
+                       with_quant: bool = False):
+        raise NotImplementedError("Dispatch function not implemented.")
+
+    @abstractmethod
+    def token_combine(self,
+                      hidden_states: torch.Tensor,
+                      context_metadata: dict,
+                      bias: torch.Tensor = None):
+        raise NotImplementedError("Combine function not implemented.")
+
+
+class TokenDispatcherWithMC2(MoETokenDispatcher):
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        device_group = get_mc2_group().device_group
+        # TODO: Try local_rank = ep_group.rank_in_group
+        local_rank = torch.distributed.get_rank(group=device_group)
+        backend = device_group._get_backend(torch.device("npu"))
+        self.moe_all_to_all_group_name = backend.get_hccl_comm_name(local_rank)
+        self.ep_rank_id = get_mc2_group().rank_in_group
+        self.ep_world_size = get_mc2_group().world_size
+        self.enable_dispatch_v2 = hasattr(torch_npu,
+                                          "npu_moe_distribute_dispatch_v2")
+        self.need_extra_args = (
+            get_ascend_soc_version() == AscendSocVersion.A3)
+
+        # NOTE: Currently, when in A3, we need to pass in some extra param into dispatch & combine
+        self.a3_need_extra_args = \
+            get_ascend_soc_version() == AscendSocVersion.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
+
+    def get_dispatch_mc2_kwargs(
+        self,
+        hidden_states: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        expert_map: torch.Tensor,
+        mc2_mask: torch.Tensor,
+        global_redundant_expert_num: int = 0,
+    ):
+        if self.with_quant:
+            quant_mode = 2
+            moe_expert_num = len(expert_map)
+        else:
+            quant_mode = 0
+            moe_expert_num = len(expert_map)
+        kwargs_mc2 = {
+            "x": hidden_states,
+            "expert_ids": topk_ids,
+            "expert_shard_type": 0,
+            "shared_expert_rank_num": 0,
+            "moe_expert_num": moe_expert_num,
+            "global_bs": 0,
+            "expert_token_nums_type": 0,
+        }
+
+        stage1_kwargs = {
+            "scales": None,
+            "quant_mode": quant_mode,
+            "group_ep": self.moe_all_to_all_group_name,
+            "ep_world_size": self.ep_world_size,
+            "ep_rank_id": self.ep_rank_id,
+        }
+        if self.need_extra_args:
+            stage1_kwargs.update({
+                "group_tp": self.moe_all_to_all_group_name,
+                "tp_world_size": 1,
+                "tp_rank_id": 0,
+            })
+        if self.a3_need_extra_args and self.enable_dispatch_v2:
+            stage1_kwargs.update({
+                "x_active_mask": mc2_mask,
+            })
+        if self.need_expert_scale:
+            stage1_kwargs.update({
+                "expert_scales":
+                topk_weights.to(torch.float32),
+            })
+
+        kwargs_mc2.update(stage1_kwargs)
+        return kwargs_mc2
+
+    def token_dispatch(
+        self,
+        hidden_states: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        expert_map: Optional[torch.Tensor] = None,
+        log2phy: Optional[torch.Tensor] = None,
+        global_redundant_expert_num: int = 0,
+        shared_experts: Optional[Any] = None,
+        quantized_x_for_share: Optional[Any] = None,
+        dynamic_scale_for_share: Optional[Any] = None,
+        mc2_mask: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        with_quant: bool = False,
+    ):
+        self.with_quant = with_quant
+
+        # Apply log2phy if needed
+        if log2phy is not None:
+            topk_ids = log2phy[topk_ids]
+
+        kwargs_mc2 = self.get_dispatch_mc2_kwargs(hidden_states, topk_weights,
+                                                  topk_ids, expert_map,
+                                                  mc2_mask,
+                                                  global_redundant_expert_num)
+        output = torch_npu.npu_moe_distribute_dispatch_v2(
+            **kwargs_mc2
+        ) if self.enable_dispatch_v2 else torch_npu.npu_moe_distribute_dispatch(
+            **kwargs_mc2)
+        # comm_stream.wait_stream(torch.npu.current_stream())
+        expand_x, dynamic_scale, assist_info_for_combine, expert_token_nums, \
+            ep_recv_counts, tp_recv_counts, expand_scales = output[0:7]
+
+        # Handle shared experts (store intermediate results in local vars, not self)
+        shared_act = None
+        swiglu_out_scale = None
+        if with_quant:
+            if shared_experts is not None:
+                share_up_out, _ = shared_experts.gate_up_proj(
+                    (quantized_x_for_share, dynamic_scale_for_share))
+                shared_gate_up, shared_dequant_scale = share_up_out[
+                    0], share_up_out[1]
+                shared_act_out = shared_experts.act_fn(
+                    (shared_gate_up, shared_dequant_scale))
+                shared_act, swiglu_out_scale = shared_act_out[
+                    0], shared_act_out[1]
+        else:
+            if shared_experts is not None:
+                shared_gate_up, _ = shared_experts.gate_up_proj(hidden_states)
+                shared_act = shared_experts.act_fn(shared_gate_up)
+
+        context_metadata = {
+            "topk_ids": topk_ids,
+            "topk_weights": topk_weights,
+            "mc2_mask": mc2_mask,
+            "expert_map": expert_map,
+            "ep_recv_counts": ep_recv_counts,
+            "tp_recv_counts": tp_recv_counts,
+            "assist_info_for_combine": assist_info_for_combine,
+            "shared_experts": shared_experts,
+            "shared_act": shared_act,
+            "swiglu_out_scale": swiglu_out_scale,
+            "expand_scales": expand_scales
+        }
+
+        return {
+            "group_list_type": 0,
+            "hidden_states": expand_x,
+            "group_list": expert_token_nums,
+            "dynamic_scale": dynamic_scale,
+            "context_metadata": context_metadata
+        }
+
+    def get_combine_mc_kwargs(self, hidden_states: torch.Tensor,
+                              context_metadata: dict):
+        expert_map = context_metadata["expert_map"]
+        topk_ids = context_metadata["topk_ids"]
+        topk_weights = context_metadata["topk_weights"]
+        ep_recv_counts = context_metadata["ep_recv_counts"]
+        tp_recv_counts = context_metadata["tp_recv_counts"]
+        assist_info_for_combine = context_metadata["assist_info_for_combine"]
+        mc2_mask = context_metadata["mc2_mask"]
+        expand_scales = context_metadata["expand_scales"]
+
+        assert expert_map is not None
+        moe_expert_num = len(expert_map)
+
+        kwargs_mc2 = {
+            "expand_x": hidden_states,
+            "expert_ids": topk_ids,
+            "expert_scales": topk_weights.to(torch.float32),
+            "expert_shard_type": 0,
+            "shared_expert_rank_num": 0,
+            "moe_expert_num": moe_expert_num,
+            "global_bs": 0,
+        }
+
+        if self.with_quant:
+            tp_recv_counts = torch.empty(1,
+                                         dtype=torch.int32,
+                                         device=hidden_states.device)
+
+        stage3_kwargs = {
+            "ep_send_counts": ep_recv_counts,
+            "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:
+            stage3_kwargs["assist_info_for_combine"] = assist_info_for_combine
+        else:
+            stage3_kwargs["expand_idx"] = assist_info_for_combine
+
+        if self.need_extra_args:
+            stage3_kwargs.update({
+                "tp_send_counts": tp_recv_counts,
+                "group_tp": self.moe_all_to_all_group_name,
+                "tp_world_size": 1,
+                "tp_rank_id": 0,
+            })
+
+        if self.a3_need_extra_args and self.enable_dispatch_v2:
+            stage3_kwargs["x_active_mask"] = mc2_mask
+
+        kwargs_mc2.update(stage3_kwargs)
+        return kwargs_mc2
+
+    def token_combine(
+        self,
+        hidden_states: torch.Tensor,
+        context_metadata: dict,
+        bias: torch.Tensor = None,
+    ):
+        assert bias is None, "Bias is not supported in MoEAlltoAllvTokenDispatcher."
+
+        kwargs_mc2 = self.get_combine_mc_kwargs(hidden_states,
+                                                context_metadata)
+        combined_output = torch_npu.npu_moe_distribute_combine_v2(**kwargs_mc2) \
+            if self.enable_dispatch_v2 else torch_npu.npu_moe_distribute_combine(**kwargs_mc2)
+
+        # Handle shared experts from metadata
+        shared_experts = context_metadata["shared_experts"]
+        if shared_experts is None:
+            return combined_output
+
+        shared_act = context_metadata["shared_act"]
+        if self.with_quant:
+            swiglu_out_scale = context_metadata["swiglu_out_scale"]
+            shared_hidden_states, _ = shared_experts.down_proj(
+                (shared_act, swiglu_out_scale))
+        else:
+            shared_hidden_states, _ = shared_experts.down_proj(shared_act)
+
+        return combined_output, shared_hidden_states
+
+
+class TokenDispatcherWithAllGather(MoETokenDispatcher):
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.apply_router_weight_on_input = False
+        self.max_num_tokens = kwargs.get("max_num_tokens")
+        self.num_experts_local = kwargs.get("num_local_experts", 0)
+        self.original_shape = None
+        self.with_quant = False
+
+    def token_dispatch(self,
+                       hidden_states: torch.Tensor,
+                       topk_weights: torch.Tensor,
+                       topk_ids: torch.Tensor,
+                       expert_map: Optional[torch.Tensor] = None,
+                       log2phy: Optional[torch.Tensor] = None,
+                       global_redundant_expert_num: int = 0,
+                       shared_experts: Optional[Any] = None,
+                       quantized_x_for_share: Optional[Any] = None,
+                       dynamic_scale_for_share: Optional[Any] = None,
+                       mc2_mask: Optional[torch.Tensor] = None,
+                       apply_router_weight_on_input: bool = False,
+                       with_quant: bool = False):
+        self.with_quant = with_quant
+        self.original_shape = hidden_states.shape
+
+        num_tokens = hidden_states.shape[:-1].numel()
+        self.apply_router_weight_on_input = apply_router_weight_on_input
+        if self.apply_router_weight_on_input:
+            assert (topk_weights.dim() == 2
+                    ), "`topk_weights` should be in shape (num_tokens, topk)"
+            _, topk = topk_weights.shape
+            assert (
+                topk == 1
+            ), "Only support topk=1 when `apply_router_weight_on_input` is True"
+            hidden_states = hidden_states * \
+                topk_weights.to(hidden_states.dtype)
+        if expert_map is not None:
+            global_num_experts = len(expert_map)
+            mask = (expert_map[topk_ids] != -1)
+            topk_weights = topk_weights * mask
+            first_expert_idx = get_ep_group(
+            ).rank_in_group * self.num_experts_local
+            last_expert_idx = first_expert_idx + self.num_experts_local
+        else:
+            first_expert_idx = 0
+            last_expert_idx = self.num_experts_local
+            global_num_experts = self.num_experts_local
+
+        sorted_hidden_states, expanded_row_idx, expert_tokens, pertoken_scale = (
+            torch_npu.npu_moe_init_routing_v2(
+                hidden_states,
+                topk_ids,
+                active_num=num_tokens * self.top_k,
+                expert_num=global_num_experts,
+                expert_tokens_num_type=1,
+                expert_tokens_num_flag=True,
+                active_expert_range=[first_expert_idx, last_expert_idx],
+                quant_mode=1 if self.with_quant else -1,
+            ))
+        expert_tokens = expert_tokens.to(torch.int64)
+        group_list_type = 1  # `count` mode
+        context_metadata = {
+            "topk_weights": topk_weights,
+            "expanded_row_idx": expanded_row_idx
+        }
+        return {
+            "group_list_type": group_list_type,
+            "hidden_states": sorted_hidden_states,
+            "group_list": expert_tokens,
+            "dynamic_scale": pertoken_scale if self.with_quant else None,
+            "context_metadata": context_metadata
+        }
+
+    def token_combine(self,
+                      hidden_states: torch.Tensor,
+                      context_metadata: dict,
+                      bias: torch.Tensor = None):
+        assert self.original_shape is not None
+        final_hidden_states = torch_npu.npu_moe_token_unpermute(
+            permuted_tokens=hidden_states,
+            sorted_indices=torch.abs(context_metadata["expanded_row_idx"]),
+            probs=context_metadata["topk_weights"])
+        if len(self.original_shape) == 3:
+            final_hidden_states = final_hidden_states.view(self.original_shape)
+
+        # these values are no longer used, so they need to be set to None for memory release.
+        return final_hidden_states
+
+
+# mypy: disable-error-code="override"
+class TokenDispatcherWithMoge(MoETokenDispatcher):
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.apply_router_weight_on_input = False
+        self.local_num_experts = self.num_experts // self.ep_size
+        self.local_num_group = self.top_k // self.ep_size
+        self.bsz = None
+
+    def token_dispatch(self,
+                       hidden_states: torch.Tensor,
+                       topk_weights: torch.Tensor,
+                       topk_ids: torch.Tensor,
+                       expert_map: Optional[torch.Tensor] = None,
+                       log2phy: Optional[torch.Tensor] = None,
+                       global_redundant_expert_num: int = 0,
+                       shared_experts: Optional[Any] = None,
+                       quantized_x_for_share: Optional[Any] = None,
+                       dynamic_scale_for_share: Optional[Any] = None,
+                       mc2_mask: Optional[torch.Tensor] = None,
+                       apply_router_weight_on_input: bool = False,
+                       with_quant: bool = False):
+        self.bsz, _ = hidden_states.shape
+        flatten_topk_ids = topk_ids.view(-1)
+        self.sorted_topk_ids = torch.argsort(flatten_topk_ids.float())
+        self.sorted_topk_ids = self.sorted_topk_ids.to(torch.int32)
+        sorted_hidden_states = hidden_states.index_select(
+            0, self.sorted_topk_ids // self.local_num_group)
+
+        experts_id = torch.arange(0,
+                                  self.local_num_experts,
+                                  dtype=topk_ids.dtype,
+                                  device=topk_ids.device)
+        num_tokens_per_expert = (
+            flatten_topk_ids.unsqueeze(-1) == experts_id).to(
+                torch.float32).sum(0)
+        topk_scales = topk_weights.view(-1).index_select(
+            0, self.sorted_topk_ids).unsqueeze(-1)
+        group_list = num_tokens_per_expert.cumsum(dim=0).to(torch.int64)
+        group_list_type = 0
+        return {
+            "group_list_type": group_list_type,
+            "hidden_states": sorted_hidden_states,
+            "group_list": group_list,
+            "topk_scales": topk_scales
+        }
+
+    def token_combine(self,
+                      hidden_states: torch.Tensor,
+                      context_metadata: dict,
+                      bias: torch.Tensor = None):
+        unsorted_topk_ids = torch.argsort(self.sorted_topk_ids.float()).to(
+            torch.int32)
+        unsorted_hidden_states = hidden_states.index_select(
+            0, unsorted_topk_ids)
+        final_hidden_states = unsorted_hidden_states.reshape(
+            self.bsz, self.top_k // self.ep_size, -1).sum(1)
+        return final_hidden_states
+
+
+class TokenDispatcherWithAll2AllV(MoETokenDispatcher):
+    """
+    The implementation of the AlltoAll-based token dispatcher, which handles token
+    dispatching on the sequence level instead of token level. The core of this implementation
+    lies in each device dispatching on the entire sequence, with the hidden state being partitioned.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.with_quant = False
+        self.num_local_experts = kwargs.get("num_local_experts", 0)
+
+        self.hidden_shape = None
+        self.hidden_shape_before_permute = None
+
+        assert self.num_local_experts > 0, "Expected at least one expert"
+        if self.num_local_experts > 1:
+            self.expert_ids_per_ep_rank = torch.tensor(
+                [i % self.num_local_experts for i in range(self.num_experts)],
+                dtype=torch.int32,
+                device=torch.npu.current_device(),
+            )
+
+        local_expert_indices_offset = (self.ep_rank * self.num_local_experts)
+
+        self.local_expert_indices = [
+            local_expert_indices_offset + i
+            for i in range(self.num_local_experts)
+        ]
+        assert (len(self.local_expert_indices) == self.num_local_experts
+                ), "Invalid local expert indices"
+        for i in range(len(self.local_expert_indices) - 1):
+            assert (self.local_expert_indices[i] ==
+                    self.local_expert_indices[i + 1] -
+                    1), "local_expert_indices must be continuous"
+
+    def token_dispatch(
+        self,
+        hidden_states: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        expert_map: Optional[torch.Tensor] = None,
+        log2phy: Optional[torch.Tensor] = None,
+        global_redundant_expert_num: int = 0,
+        shared_experts: Optional[Any] = None,
+        quantized_x_for_share: Optional[Any] = None,
+        dynamic_scale_for_share: Optional[Any] = None,
+        mc2_mask: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        with_quant: bool = False,
+    ):
+        self.with_quant = with_quant
+        self.hidden_shape = hidden_states.shape
+
+        if log2phy is not None:
+            topk_ids = log2phy[topk_ids]
+
+        (
+            permutated_local_input_tokens,
+            reversed_local_input_permutation_mapping,
+            tokens_per_expert,
+            input_splits,
+            output_splits,
+            num_global_tokens_per_local_expert,
+            global_input_tokens_local_experts_indices,
+        ) = self._dispatch_preprocess(hidden_states, topk_ids)
+
+        dynamic_scale_after_all2all = None
+        if self.with_quant:
+            permutated_local_input_tokens, dynamic_scale = torch_npu.npu_dynamic_quant(
+                permutated_local_input_tokens)
+            _, dynamic_scale_after_all2all, permute2_ep_all_to_all_handle = async_all_to_all(
+                dynamic_scale, output_splits, input_splits, self.ep_group)
+            permute2_ep_all_to_all_handle.wait()
+            dynamic_scale.untyped_storage().resize_(0)
+
+        _, global_input_tokens, permute1_ep_all_to_all_handle = async_all_to_all(
+            permutated_local_input_tokens, output_splits, input_splits,
+            self.ep_group)
+        permute1_ep_all_to_all_handle.wait()
+        permutated_local_input_tokens.untyped_storage().resize_(0)
+
+        # Postprocess
+        global_input_tokens, dynamic_scale_final, reversed_global_input_permutation_mapping = self._dispatch_postprocess(
+            global_input_tokens, dynamic_scale_after_all2all,
+            global_input_tokens_local_experts_indices)
+
+        context_metadata = {
+            "input_splits":
+            input_splits,
+            "output_splits":
+            output_splits,
+            "topk_weights":
+            topk_weights,
+            "reversed_local_input_permutation_mapping":
+            reversed_local_input_permutation_mapping,
+            "reversed_global_input_permutation_mapping":
+            reversed_global_input_permutation_mapping
+        }
+
+        return {
+            "hidden_states": global_input_tokens,
+            "group_list": tokens_per_expert,
+            "group_list_type": 1,
+            "dynamic_scale": dynamic_scale_final,
+            "context_metadata": context_metadata,
+        }
+
+    def token_combine(
+        self,
+        hidden_states: torch.Tensor,
+        context_metadata: dict,
+        bias: torch.Tensor = None,
+    ):
+        assert bias is None, "Bias is not supported in MoEAlltoAllvTokenDispatcher."
+
+        # 1. Preprocess using metadata
+        hidden_states = self._combine_preprocess(hidden_states,
+                                                 context_metadata)
+
+        # 2. AllToAll
+        _, permutated_local_input_tokens, handle = async_all_to_all(
+            hidden_states,
+            context_metadata["input_splits"],
+            context_metadata["output_splits"],
+            self.ep_group,
+        )
+        handle.wait()
+        hidden_states.untyped_storage().resize_(0)
+
+        # 3. Postprocess using metadata
+        output = self._combine_postprocess(permutated_local_input_tokens,
+                                           context_metadata)
+
+        return output
+
+    def _dispatch_preprocess(self, hidden_states, topk_ids):
+        assert self.hidden_shape is not None
+        hidden_states = hidden_states.view(-1, hidden_states.size(-1))
+        (
+            tokens_per_expert,
+            input_splits,
+            output_splits,
+            num_global_tokens_per_local_expert,
+            global_input_tokens_local_experts_indices,
+        ) = self._preprocess(topk_ids)
+
+        self.hidden_shape_before_permute = hidden_states.shape
+
+        permutated_local_input_tokens, reversed_local_input_permutation_mapping = torch_npu.npu_moe_token_permute(
+            tokens=hidden_states,
+            indices=topk_ids,
+            num_out_tokens=self.num_out_tokens,
+        )
+
+        return (
+            permutated_local_input_tokens,
+            reversed_local_input_permutation_mapping,
+            tokens_per_expert,
+            input_splits,
+            output_splits,
+            num_global_tokens_per_local_expert,
+            global_input_tokens_local_experts_indices,
+        )
+
+    def _preprocess(self, topk_ids: torch.Tensor):
+        num_local_tokens_per_expert = torch.histc(topk_ids,
+                                                  bins=self.num_experts,
+                                                  min=0,
+                                                  max=self.num_experts)
+
+        ep_size = self.ep_size
+        self.num_out_tokens = topk_ids.numel()
+
+        input_splits = (num_local_tokens_per_expert.reshape(
+            ep_size,
+            self.num_local_experts).sum(axis=1).to(torch.device("cpu"),
+                                                   non_blocking=True).numpy())
+
+        num_global_tokens_per_expert = gather_from_sequence_parallel_region(
+            num_local_tokens_per_expert,
+            group=self.ep_group).reshape(ep_size, self.num_experts)
+        num_global_tokens_per_local_expert = num_global_tokens_per_expert[:, self.local_expert_indices[
+            0]:self.local_expert_indices[-1] + 1]
+        if num_global_tokens_per_local_expert is None:
+            raise ValueError(
+                "num_global_tokens_per_local_expert must be set before sum.")
+
+        output_splits = (num_global_tokens_per_local_expert.sum(axis=-1).to(
+            torch.device("cpu"), non_blocking=True).numpy())
+        num_tokens_per_local_expert = num_global_tokens_per_local_expert.sum(
+            axis=0)
+
+        global_input_tokens_local_experts_indices = None
+        if self.num_local_experts > 1:
+            if num_global_tokens_per_local_expert is None:
+                raise ValueError(
+                    "num_global_tokens_per_local_expert must be set before operations."
+                )
+            global_input_tokens_local_experts_indices = torch.repeat_interleave(
+                self.expert_ids_per_ep_rank,
+                num_global_tokens_per_local_expert.ravel())
+        else:
+            torch.npu.synchronize()
+
+        return (
+            num_tokens_per_local_expert,
+            input_splits,
+            output_splits,
+            num_global_tokens_per_local_expert,
+            global_input_tokens_local_experts_indices,
+        )
+
+    def _dispatch_postprocess(self, global_input_tokens,
+                              dynamic_scale_after_all2all,
+                              global_input_tokens_local_experts_indices):
+        # Early return if no local experts or no tokens
+        if self.num_local_experts <= 1:
+            return global_input_tokens, dynamic_scale_after_all2all, None
+
+        # Handle quantized case
+        if self.with_quant:
+            assert global_input_tokens_local_experts_indices is not None, \
+                "global_input_tokens_local_experts_indices must be provided"
+            expert_idx_2d = global_input_tokens_local_experts_indices.unsqueeze(
+                -1)
+            active_num = global_input_tokens_local_experts_indices.numel()
+
+            if active_num <= 0:
+                reversed_global_input_permutation_mapping = global_input_tokens_local_experts_indices
+                return global_input_tokens, dynamic_scale_after_all2all, reversed_global_input_permutation_mapping
+
+            global_input_tokens, reversed_global_input_permutation_mapping, _, expanded_scale = torch_npu.npu_moe_init_routing_v2(
+                global_input_tokens,
+                expert_idx_2d,
+                scale=dynamic_scale_after_all2all,
+                active_num=active_num,
+                expert_capacity=0,
+                expert_num=self.num_local_experts,
+                expert_tokens_num_type=1,
+                expert_tokens_num_flag=True,
+                active_expert_range=[0, self.num_local_experts],
+                quant_mode=-1,
+                row_idx_type=0,
+            )
+            return global_input_tokens, expanded_scale, reversed_global_input_permutation_mapping
+
+        # Non-quantized case
+        global_input_tokens, reversed_global_input_permutation_mapping = torch_npu.npu_moe_token_permute(
+            global_input_tokens, global_input_tokens_local_experts_indices)
+        return global_input_tokens, None, reversed_global_input_permutation_mapping
+
+    def _combine_preprocess(self, hidden_states: torch.Tensor,
+                            context_metadata: dict) -> torch.Tensor:
+        # Unpermutation 2: expert output to AlltoAll input
+        if hidden_states.shape[0] > 0 and self.num_local_experts > 1:
+            rev_global = context_metadata[
+                "reversed_global_input_permutation_mapping"]
+            hidden_states = torch_npu.npu_moe_token_unpermute(
+                hidden_states, rev_global)
+        return hidden_states
+
+    def _combine_postprocess(self, permutated_local_input_tokens: torch.Tensor,
+                             context_metadata: dict) -> torch.Tensor:
+        # Unpermutation 1: AlltoAll output to output
+        output = torch_npu.npu_moe_token_unpermute(
+            permuted_tokens=permutated_local_input_tokens,
+            sorted_indices=context_metadata[
+                "reversed_local_input_permutation_mapping"].to(torch.int32),
+            probs=context_metadata["topk_weights"],
+            restore_shape=self.hidden_shape_before_permute,
+        )
+        output = output.view(self.hidden_shape)
+        return output