[Qwen3] Enable dual-chunk-attention support for Qwen3 models. (#21924)

Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>

vllm/model_executor/models/qwen3.py

@@ -23,7 +23,7 @@
 # limitations under the License.
 """Inference-only Qwen3 model compatible with HuggingFace weights."""
 from collections.abc import Iterable
-from typing import Optional, Union
+from typing import Any, Optional, Union
 
 import torch
 from torch import nn
@@ -47,14 +47,16 @@ from vllm.sequence import IntermediateTensors
 from .interfaces import SupportsLoRA, SupportsPP
 from .qwen2 import Qwen2MLP as Qwen3MLP
 from .qwen2 import Qwen2Model
-from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
+from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
+                    maybe_prefix)
 
 logger = init_logger(__name__)
 
 
 class Qwen3Attention(nn.Module):
 
-    def __init__(self,
+    def __init__(
+        self,
         hidden_size: int,
         num_heads: int,
         num_kv_heads: int,
@@ -67,7 +69,9 @@ class Qwen3Attention(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
         rope_scaling: Optional[tuple] = None,
         prefix: str = "",
-                 attn_type: str = AttentionType.DECODER) -> None:
+        attn_type: str = AttentionType.DECODER,
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
+    ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
         tp_size = get_tensor_model_parallel_world_size()
@@ -89,6 +93,7 @@ class Qwen3Attention(nn.Module):
         self.kv_size = self.num_kv_heads * self.head_dim
         self.scaling = self.head_dim**-0.5
         self.rope_theta = rope_theta
+        self.dual_chunk_attention_config = dual_chunk_attention_config
 
         self.qkv_proj = QKVParallelLinear(
             hidden_size,
@@ -113,15 +118,22 @@ class Qwen3Attention(nn.Module):
             max_position=max_position,
             base=self.rope_theta,
             rope_scaling=rope_scaling,
+            dual_chunk_attention_config=dual_chunk_attention_config,
         )
-        self.attn = Attention(self.num_heads,
+        self.attn = Attention(
+            self.num_heads,
             self.head_dim,
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             cache_config=cache_config,
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
-                              attn_type=attn_type)
+            attn_type=attn_type,
+            **{
+                "layer_idx": extract_layer_index(prefix),
+                "dual_chunk_attention_config": dual_chunk_attention_config,
+            } if dual_chunk_attention_config else {},
+        )
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
 
@@ -161,6 +173,9 @@ class Qwen3DecoderLayer(nn.Module):
         # Requires transformers > 4.32.0
         rope_theta = getattr(config, "rope_theta", 1000000)
         rope_scaling = getattr(config, "rope_scaling", None)
+        dual_chunk_attention_config = getattr(config,
+                                              "dual_chunk_attention_config",
+                                              None)
 
         # By default, Qwen3 uses causal attention as it is a decoder-only model.
         # You can override the HF config with `is_causal=False` to enable
@@ -185,6 +200,7 @@ class Qwen3DecoderLayer(nn.Module):
             rope_scaling=rope_scaling,
             prefix=f"{prefix}.self_attn",
             attn_type=attn_type,
+            dual_chunk_attention_config=dual_chunk_attention_config,
         )
         self.mlp = Qwen3MLP(
             hidden_size=self.hidden_size,

vllm/model_executor/models/qwen3_moe.py

@@ -185,6 +185,7 @@ class Qwen3MoeAttention(nn.Module):
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -208,6 +209,7 @@ class Qwen3MoeAttention(nn.Module):
         self.scaling = self.head_dim**-0.5
         self.rope_theta = rope_theta
         self.max_position_embeddings = max_position_embeddings
+        self.dual_chunk_attention_config = dual_chunk_attention_config
 
         self.qkv_proj = QKVParallelLinear(hidden_size,
                                           self.head_dim,
@@ -229,14 +231,21 @@ class Qwen3MoeAttention(nn.Module):
             max_position=max_position_embeddings,
             base=rope_theta,
             rope_scaling=rope_scaling,
+            dual_chunk_attention_config=dual_chunk_attention_config,
         )
-        self.attn = Attention(self.num_heads,
+        self.attn = Attention(
+            self.num_heads,
             self.head_dim,
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             cache_config=cache_config,
             quant_config=quant_config,
-                              prefix=f"{prefix}.attn")
+            prefix=f"{prefix}.attn",
+            **{
+                "layer_idx": extract_layer_index(prefix),
+                "dual_chunk_attention_config": dual_chunk_attention_config,
+            } if dual_chunk_attention_config else {},
+        )
 
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
@@ -280,6 +289,9 @@ class Qwen3MoeDecoderLayer(nn.Module):
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings",
                                           8192)
+        dual_chunk_attention_config = getattr(config,
+                                              "dual_chunk_attention_config",
+                                              None)
         self.self_attn = Qwen3MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -293,6 +305,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
             cache_config=cache_config,
             quant_config=quant_config,
             prefix=f"{prefix}.self_attn",
+            dual_chunk_attention_config=dual_chunk_attention_config,
         )
 
         # `mlp_only_layers` in the config.