CustomOp: Unify aiter impl into GroupedTopk

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -35,6 +35,9 @@
 from vllm.model_executor.layers.fused_moe.prepare_finalize import (
     MoEPrepareAndFinalizeNoEP,
 )
+from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa: E501
+    rocm_aiter_grouped_topk,
+)
 from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceNoOP,
 )
@@ -1295,6 +1298,7 @@ def __init__(
         topk_group: int = 0,
         scoring_func: str = "softmax",
         routed_scaling_factor: float = 1.0,
+        num_fused_shared_experts: int = 0,
     ) -> None:
         super().__init__()
         self.native_impl = grouped_topk
@@ -1304,6 +1308,7 @@ def __init__(
         self.topk_group = topk_group
         self.scoring_func = scoring_func
         self.routed_scaling_factor = routed_scaling_factor
+        self.num_fused_shared_experts = num_fused_shared_experts
 
     def forward_native(
         self,
@@ -1333,6 +1338,32 @@ def forward_cuda(
             hidden_states, gating_output, e_score_correction_bias
         )
 
+    def forward_hip(
+        self,
+        hidden_states: torch.Tensor,
+        gating_output: torch.Tensor,
+        e_score_correction_bias: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if rocm_aiter_ops.is_fused_moe_enabled():
+            if not rocm_aiter_ops.is_fusion_moe_shared_experts_enabled():
+                assert self.num_fused_shared_experts == 0
+            return rocm_aiter_grouped_topk(
+                hidden_states,
+                gating_output,
+                self.topk,
+                self.renormalize,
+                self.num_expert_group,
+                self.topk_group,
+                self.scoring_func,
+                self.routed_scaling_factor,
+                e_score_correction_bias,
+                self.num_fused_shared_experts,
+            )
+        else:
+            return self.forward_native(
+                hidden_states, gating_output, e_score_correction_bias
+            )
+
 
 @torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
 def eplb_map_to_physical_and_record(

vllm/model_executor/layers/fused_moe/layer.py

@@ -4,7 +4,6 @@
 from collections.abc import Callable, Iterable
 from contextlib import nullcontext
 from enum import Enum
-from functools import partial
 from typing import Literal, cast, get_args, overload
 
 import torch
@@ -67,9 +66,6 @@ def _eplb_map_to_physical_and_record(
 
     eplb_map_to_physical_and_record = _eplb_map_to_physical_and_record
 from vllm.model_executor.layers.fused_moe.fused_moe import GroupedTopk
-from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa: E501
-    rocm_aiter_grouped_topk,
-)
 
 if current_platform.is_tpu():
     from .moe_pallas import fused_moe as fused_moe_pallas
@@ -1583,28 +1579,15 @@ def valid_grouping() -> bool:
         elif self.use_grouped_topk and valid_grouping():
             assert self.topk_group is not None
             assert self.num_expert_group is not None
-            if rocm_aiter_ops.is_fused_moe_enabled():
-                if not rocm_aiter_ops.is_fusion_moe_shared_experts_enabled():
-                    assert self.num_fused_shared_experts == 0
-                grouped_topk_impl = partial(
-                    rocm_aiter_grouped_topk,
-                    num_fused_shared_experts=self.num_fused_shared_experts,
-                    topk=self.top_k,
-                    renormalize=self.renormalize,
-                    num_expert_group=self.num_expert_group,
-                    topk_group=self.topk_group,
-                    scoring_func=self.scoring_func,
-                    routed_scaling_factor=self.routed_scaling_factor,
-                )
-            else:
-                grouped_topk_impl = GroupedTopk(
-                    topk=self.top_k,
-                    renormalize=self.renormalize,
-                    num_expert_group=self.num_expert_group,
-                    topk_group=self.topk_group,
-                    scoring_func=self.scoring_func,
-                    routed_scaling_factor=self.routed_scaling_factor,
-                )
+            grouped_topk_impl = GroupedTopk(
+                topk=self.top_k,
+                renormalize=self.renormalize,
+                num_expert_group=self.num_expert_group,
+                topk_group=self.topk_group,
+                scoring_func=self.scoring_func,
+                routed_scaling_factor=self.routed_scaling_factor,
+                num_fused_shared_experts=self.num_fused_shared_experts,
+            )
 
             topk_weights, topk_ids = grouped_topk_impl(
                 hidden_states=hidden_states,