[BugFix] Fix the bug that qwen3 moe doesn't work with aclgraph (#2183)

What's the PR does:
1. Move AscendSparseMoeBlock to qwen3 model, since it's only used by
qwen3 model.
2. Disable AscendSparseMoeBlock if aclgraph is enabled,
AscendSparseMoeBlock doesn't work with aclgraph currently.

- vLLM version: v0.10.0
- vLLM main:
vllm-project/vllm@cdfd687

Signed-off-by: wangxiyuan <[email protected]>

Author: wangxiyuan

tests/e2e/multicard/test_qwen3_moe.py

@@ -0,0 +1,55 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+"""Compare the short outputs of HF and vLLM when using greedy sampling.
+
+Run `pytest tests/test_offline_inference.py`.
+"""
+
+from tests.e2e.conftest import VllmRunner
+
+
+def test_models_distributed_Qwen3_MOE_TP2():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    dtype = "half"
+    max_tokens = 5
+    with VllmRunner(
+            "Qwen/Qwen3-30B-A3B",
+            dtype=dtype,
+            tensor_parallel_size=2,
+            distributed_executor_backend="mp",
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+def test_models_distributed_Qwen3_MOE_TP2_WITH_EP():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    dtype = "half"
+    max_tokens = 5
+    with VllmRunner(
+            "Qwen/Qwen3-30B-A3B",
+            dtype=dtype,
+            tensor_parallel_size=2,
+            enable_expert_parallel=True,
+            distributed_executor_backend="mp",
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

vllm_ascend/models/qwen3_moe.py

@@ -17,25 +17,102 @@
 # This file is a part of the vllm-ascend project.
 from typing import Optional
 
+import torch
 from torch import nn
 from transformers import PretrainedConfig
 from vllm.compilation.decorators import support_torch_compile
-from vllm.config import CacheConfig
+from vllm.config import CacheConfig, CompilationLevel, VllmConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.distributed.parallel_state import (get_dp_group, get_ep_group,
+                                             get_tp_group)
+from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import ReplicatedLinear
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.models.qwen3_moe import (Qwen3MoeAttention,
                                                   Qwen3MoeDecoderLayer,
                                                   Qwen3MoeForCausalLM,
-                                                  Qwen3MoeMLP, Qwen3MoeModel)
+                                                  Qwen3MoeMLP, Qwen3MoeModel,
+                                                  Qwen3MoeSparseMoeBlock)
 from vllm.model_executor.models.utils import (
     extract_layer_index, make_empty_intermediate_tensors_factory, make_layers,
     maybe_prefix)
 
-from vllm_ascend.ops.fused_moe import AscendSparseMoeBlock
-from vllm_ascend.platform import VllmConfig
+from vllm_ascend.ops.fused_moe import AscendFusedMoE
+
+
+class CustomSparseMoeBlock(Qwen3MoeSparseMoeBlock):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        nn.Module.__init__(self)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        if self.tp_size > config.num_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {config.num_experts}.")
+
+        self.gate = ReplicatedLinear(
+            config.hidden_size,
+            config.num_experts,
+            bias=False,
+            quant_config=None,
+            prefix=f"{prefix}.gate",
+        )
+
+        self.experts = AscendFusedMoE(
+            num_experts=config.num_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.moe_intermediate_size,
+            reduce_results=False,
+            renormalize=config.norm_topk_prob,
+            quant_config=quant_config,
+            prefix=f"{prefix}.experts",
+        )
+
+        self.top_k = config.num_experts_per_tok
+
+        self.dp_size = get_dp_group().world_size
+
+        self.tp_group = get_tp_group().device_group
+        self.tp_rank = get_tp_group().rank_in_group
+        self.ep_group = get_ep_group()
+
+        self.params_dtype = torch.get_default_dtype()
+
+    def forward(
+        self,
+        hidden_states,
+        attn_metadata=None,
+    ):
+        if attn_metadata is None:
+            attn_metadata = get_forward_context().attn_metadata
+        # when profile runs, force experts to load balanced tokens
+        # to avoid high memory consumption on a single rank.
+        enable_force_load_balance = get_forward_context().in_profile_run
+        is_prefill = get_forward_context().with_prefill
+
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+
+        hidden_states = self.experts(
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+            is_prefill=is_prefill,
+            top_k=self.top_k,
+            enable_force_load_balance=enable_force_load_balance,
+            shared_experts=None,
+        )
+
+        return hidden_states
 
 
 class CustomQwen3MoeDecoderLayer(Qwen3MoeDecoderLayer):
@@ -45,6 +122,7 @@ def __init__(
         config: PretrainedConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        vllm_config: Optional[VllmConfig] = None,
         prefix: str = "",
     ) -> None:
 
@@ -73,12 +151,22 @@ def __init__(
         layer_idx = extract_layer_index(prefix)
         mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
                            config.mlp_only_layers)
+        use_aclgraph = (vllm_config is not None
+                        and vllm_config.compilation_config.level
+                        == CompilationLevel.PIECEWISE
+                        and not vllm_config.model_config.enforce_eager)
         if (layer_idx not in mlp_only_layers) and (
                 config.num_experts > 0 and
             (layer_idx + 1) % config.decoder_sparse_step == 0):
-            self.mlp = AscendSparseMoeBlock(config=config,
-                                            quant_config=quant_config,
-                                            prefix=f"{prefix}.mlp")
+            if not use_aclgraph:
+                # FIXME: custom sparse moe block doesn't work with aclgraph.
+                self.mlp = CustomSparseMoeBlock(config=config,
+                                                quant_config=quant_config,
+                                                prefix=f"{prefix}.mlp")
+            else:
+                self.mlp = Qwen3MoeSparseMoeBlock(config=config,
+                                                  quant_config=quant_config,
+                                                  prefix=f"{prefix}.mlp")
         else:
             self.mlp = Qwen3MoeMLP(hidden_size=config.hidden_size,
                                    intermediate_size=config.intermediate_size,
@@ -115,6 +203,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 config=config,
                 cache_config=cache_config,
                 quant_config=quant_config,
+                vllm_config=vllm_config,
                 prefix=prefix),
             prefix=f"{prefix}.layers",
         )

vllm_ascend/ops/fused_moe.py

@@ -22,8 +22,6 @@
 import torch.distributed as dist
 import torch_npu
 from torch import nn
-from transformers import PretrainedConfig
-from vllm.attention import AttentionMetadata
 from vllm.config import get_current_vllm_config
 from vllm.distributed import (GroupCoordinator, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
@@ -37,7 +35,6 @@
     FusedMoEParallelConfig  # isort: skip
 from vllm.model_executor.layers.fused_moe.layer import (
     FusedMoE, UnquantizedFusedMoEMethod, determine_expert_map)
-from vllm.model_executor.layers.linear import ReplicatedLinear
 from vllm.model_executor.layers.quantization.base_config import \
     QuantizationConfig
 
@@ -1546,79 +1543,3 @@ def _forward_ms_fused_moe_comp(
         )
 
         return hidden_states
-
-
-class AscendSparseMoeBlock(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-        prefix: str = "",
-    ):
-        super().__init__()
-        self.tp_size = get_tensor_model_parallel_world_size()
-        if self.tp_size > config.num_experts:
-            raise ValueError(
-                f"Tensor parallel size {self.tp_size} is greater than "
-                f"the number of experts {config.num_experts}.")
-
-        ascend_config = get_ascend_config()
-        self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
-        self.enable_multistream_moe = (
-            ascend_config.torchair_graph_config.enable_multistream_moe)
-
-        self.gate = ReplicatedLinear(
-            config.hidden_size,
-            config.num_experts,
-            bias=False,
-            quant_config=None,
-            prefix=f"{prefix}.gate",
-        )
-
-        self.experts = AscendFusedMoE(
-            num_experts=config.num_experts,
-            top_k=config.num_experts_per_tok,
-            hidden_size=config.hidden_size,
-            intermediate_size=config.moe_intermediate_size,
-            reduce_results=False,
-            renormalize=config.norm_topk_prob,
-            quant_config=quant_config,
-            prefix=f"{prefix}.experts",
-        )
-
-        self.top_k = config.num_experts_per_tok
-
-        self.dp_size = get_dp_group().world_size
-
-        self.tp_group = get_tp_group().device_group
-        self.tp_rank = get_tp_group().rank_in_group
-        self.ep_group = get_ep_group()
-
-        self.params_dtype = torch.get_default_dtype()
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attn_metadata: Optional[AttentionMetadata] = None,
-    ) -> torch.Tensor:
-        if attn_metadata is None:
-            attn_metadata = get_forward_context().attn_metadata
-        # when profile runs, force experts to load balanced tokens
-        # to avoid high memory consumption on a single rank.
-        enable_force_load_balance = get_forward_context().in_profile_run
-        is_prefill = get_forward_context().with_prefill
-
-        # router_logits: (num_tokens, n_experts)
-        router_logits, _ = self.gate(hidden_states)
-
-        hidden_states = self.experts(
-            hidden_states=hidden_states,
-            router_logits=router_logits,
-            is_prefill=is_prefill,
-            top_k=self.top_k,
-            enable_force_load_balance=enable_force_load_balance,
-            shared_experts=None,
-        )
-
-        return hidden_states