[v0.9.1][Feature] add moe alltoallv.

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>

Author: weijinqian_v1

vllm_ascend/attention/attention_v1.py

@@ -28,6 +28,7 @@
 from vllm.utils import direct_register_custom_op
 from vllm.v1.core.sched.output import SchedulerOutput
 
+from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
 from vllm_ascend.ops.attention import vanilla_chunked_prefill
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
                                nd_to_nz_2d, nd_to_nz_spec)
@@ -150,6 +151,18 @@ class AscendMetadata:
     # (num_tokens,)
     slot_mapping: torch.Tensor = None
 
+    def split_metadata_for_multistream(
+        self,
+        ms_split_config: MSAttentionMetadataSplitConfig,
+    ) -> list["AscendMetadata"]:
+        """Split metadata for multi-stream with AscendMetadata"""
+        from vllm_ascend.multistream.ms_split import model_input_split_v1_attn
+        return model_input_split_v1_attn(
+            ms_split_config=ms_split_config,
+            attn_metadata=self,
+            _metadata_cls=AscendMetadata,
+        )
+
 
 class AscendAttentionMetadataBuilder:
 

vllm_ascend/models/__init__.py

@@ -41,6 +41,10 @@ def register_model():
             "DeepseekV3ForCausalLM",
             "vllm_ascend.models.deepseek_dbo:CustomDeepseekDBOForCausalLM")
 
+        ModelRegistry.register_model(
+            "Qwen3MoeForCausalLM",
+            "vllm_ascend.models.qwen3_dbo:CustomQwen3MoeForCausalLMDBO")
+
     else:
         ModelRegistry.register_model(
             "DeepseekV2ForCausalLM",
@@ -50,9 +54,9 @@ def register_model():
             "DeepseekV3ForCausalLM",
             "vllm_ascend.models.deepseek_v3:CustomDeepseekV3ForCausalLM")
 
-    ModelRegistry.register_model(
-        "Qwen3MoeForCausalLM",
-        "vllm_ascend.models.qwen3_moe:CustomQwen3MoeForCausalLM")
+        ModelRegistry.register_model(
+            "Qwen3MoeForCausalLM",
+            "vllm_ascend.models.qwen3_moe:CustomQwen3MoeForCausalLM")
 
     ModelRegistry.register_model(
         "Qwen3ForCausalLM", "vllm_ascend.models.qwen3:CustomQwen3ForCausalLM")