[1/N][Feat] Support MoE models with ACL Graph and refactor MoE communication logic (#2125)

### What this PR does / why we need it?
This PR refactors the MoE (Mixture of Experts) communication logic by
introducing a strategy pattern. It defines an abstract base class,
`MoECommMethod`, which encapsulates different communication strategies
for MoE layers. By decoupling the MoE implementation from any single
communication method, this change makes it simpler to add, replace, or
optimize communication strategies in the future.

Plan / Roadmap

1. Introduce `MoECommMethod`, implement `AllGatherImpl`, and adapt ACL
Graph handling to cover all scenarios (this PR).
2. Implement `MC2CommImpl` and `AllToAllCommImpl` to optimize
performance in specific scenarios.
3. Enable W8A8 / Int8 models to use `unified_fused_experts`.

Other notes

* Data-parallel (DP) communication currently does not work with vLLM's
dispatch/combine mechanisms; an alternative approach is required to
resolve this incompatibility.

- vLLM version: v0.10.0
- vLLM main:
f7ad6a1eb3

---------

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

vllm_ascend/ops/common_fused_moe.py

@@ -19,12 +19,13 @@ from typing import Callable, Optional
 
 import torch
 from vllm.config import CompilationLevel, get_current_vllm_config
+from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.fused_moe.layer import \
     UnquantizedFusedMoEMethod
 
 from vllm_ascend.ascend_config import get_ascend_config
-from vllm_ascend.ops.fused_moe import (fused_experts, fused_experts_moge,
-                                       select_experts)
+from vllm_ascend.ops.fused_moe import (fused_experts_moge, select_experts,
+                                       unified_fused_experts)
 from vllm_ascend.utils import is_310p
 
 original_unquantized_fused_moe_init_func = UnquantizedFusedMoEMethod.__init__
@@ -95,20 +96,18 @@ def forward_oot(
             expert_map=expert_map,
             apply_router_weight_on_input=apply_router_weight_on_input)
 
-    # If use aclgraph, we need to set max_num_tokens to make
-    # the input shape of `npu_moe_init_routing` fixed
-    max_num_tokens = self.max_num_batched_tokens if self.use_aclgraph else None
+    moe_comm_method = get_forward_context().moe_comm_method
 
-    return fused_experts(
+    return unified_fused_experts(
         hidden_states=x,
         w1=layer.w13_weight,
         w2=layer.w2_weight,
         topk_weights=topk_weights,
         topk_ids=topk_ids,
-        top_k=top_k,
+        global_num_experts=global_num_experts,
         expert_map=expert_map,
-        apply_router_weight_on_input=apply_router_weight_on_input,
-        max_num_tokens=max_num_tokens)
+        moe_comm_method=moe_comm_method,
+    )
 
 
 UnquantizedFusedMoEMethod.__init__ = unquantized_fused_moe_init_func

vllm_ascend/ops/fused_moe.py

@@ -43,6 +43,7 @@ from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.ascend_forward_context import FusedMoEState
 from vllm_ascend.distributed.communication_op import \
     data_parallel_reduce_scatter
+from vllm_ascend.distributed.moe_comm_method import MoECommMethod
 from vllm_ascend.distributed.parallel_state import get_mc2_group
 from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
 from vllm_ascend.ops.moe_dispatcher.token_dispatcher import (
@@ -57,6 +58,62 @@ from vllm_ascend.utils import (AscendSocVersion, dispose_tensor,
 MOE_ALL2ALL_BUFFER: bool = envs_ascend.MOE_ALL2ALL_BUFFER
 
 
+def unified_fused_experts(
+    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,
+    moe_comm_method: Optional[MoECommMethod] = 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,
+) -> torch.Tensor:
+    # Check constraints
+    assert hidden_states.shape[1] == w1.shape[2], (
+        f"Hidden size mismatch {hidden_states.shape[1]} != {w1.shape[2]}")
+
+    assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
+    assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
+    assert w1.stride(-1) == 1, "Stride of last dimension must be 1"
+    assert w2.stride(-1) == 1, "Stride of last dimension must be 1"
+    assert hidden_states.dtype in [
+        torch.float32, torch.float16, torch.bfloat16
+    ]
+    assert moe_comm_method is not None, "Missing communication context"
+
+    num_experts = w1.shape[0]
+
+    permuted_hidden_states, expert_tokens, group_list_type = torch.ops.vllm.moe_comm_pre_process(
+        hidden_states, topk_ids, topk_weights, expert_map, num_experts)
+    mlp_output = apply_mlp(
+        permuted_hidden_states,
+        w1,
+        w2,
+        expert_tokens,
+        group_list_type=group_list_type,
+    )
+    torch.ops.vllm.moe_comm_post_process(mlp_output, hidden_states)
+
+    return hidden_states
+
+
 def process_topk_ids(topk_ids: torch.Tensor, expert_num: int, ep_size: int,
                      max_row_per_ep_rank: int, num_tokens: int,
                      top_k: int) -> tuple[torch.Tensor, torch.Tensor]: