add qwen3-moe optimization

Signed-off-by: yangcheng (AJ) <yangcheng104@huawei.com>

Author: yangcheng (AJ)

tests/e2e/singlecard/test_offline_inference.py

@@ -35,6 +35,7 @@
 MODELS = [
     "Qwen/Qwen2.5-0.5B-Instruct",
     "Qwen/Qwen3-0.6B-Base",
+    "Qwen/Qwen3-30B-A3B",
 ]
 MULTIMODALITY_MODELS = ["Qwen/Qwen2.5-VL-3B-Instruct"]
 

vllm_ascend/models/qwen3_moe.py

@@ -16,7 +16,31 @@
 # Adapted from vllm/model_executor/models/qwen3_moe.py
 # This file is a part of the vllm-ascend project.
 
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.distributed as dist
+import torch_npu
+import vllm
+import vllm.envs as envs
+from torch import nn
+from transformers import PretrainedConfig
+from vllm.attention import AttentionMetadata
+from vllm.distributed import (get_tensor_model_parallel_world_size,
+                              get_tp_group)
+from vllm.distributed.parallel_state import get_dp_group
+from vllm.forward_context import get_forward_context
+from vllm.model_executor.layers.linear import ReplicatedLinear
+                                               
+from vllm.model_executor.layers.quantization import QuantizationConfig
+
+from vllm_ascend.ascend_config import get_ascend_config
+from vllm_ascend.distributed.parallel_state import get_ep_group
+from vllm_ascend.ops.fused_moe import AscendFusedMoE
+
 from vllm.model_executor.models.qwen3_moe import Qwen3MoeForCausalLM
+from transformers import PretrainedConfig
+from vllm.model_executor.layers.quantization import QuantizationConfig
 
 
 class CustomQwen3MoeForCausalLM(Qwen3MoeForCausalLM):
@@ -33,3 +57,89 @@ class CustomQwen3MoeForCausalLM(Qwen3MoeForCausalLM):
         "experts":
         ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
     }
+
+
+class AscendQwen3MoeSparseMoeBlock(nn.Module):
+    
+    top_k: int
+
+    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.
+        # TODO: need a better flag to indicate whether in profile run or not.
+        if attn_metadata is None:
+            # for profile run
+            is_prefill = True
+            enable_force_load_balance = True
+        else:
+            # is_prefill = attn_metadata.num_prefills > 0 is_prefill or
+            enable_force_load_balance = False
+            if hasattr(attn_metadata, 'with_prefill_across_dp'):
+                is_prefill = attn_metadata.with_prefill_across_dp
+
+        # 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
+
+
+vllm.model_executor.models.qwen3_moe.Qwen3MoeSparseMoeBlock = AscendQwen3MoeSparseMoeBlock