[feat] Support EAGLE3 for Qwen (#7745)

Co-authored-by: 纬杭 <ximing.wxm@antgroup.com>
Co-authored-by: zyksir <zyksir@outlook.com>

python/sglang/srt/models/qwen3.py

@@ -2,7 +2,7 @@
 
 import logging
 from functools import partial
-from typing import Any, Dict, Iterable, Optional, Tuple
+from typing import Any, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import nn
@@ -325,6 +325,9 @@ class Qwen3ForCausalLM(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
+        # For EAGLE3 support
+        self.capture_aux_hidden_states = False
+
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.model.get_input_embeddings(input_ids)
 
@@ -346,10 +349,18 @@ class Qwen3ForCausalLM(nn.Module):
             pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
         if self.pp_group.is_last_rank:
             if not get_embedding:
                 return self.logits_processor(
-                    input_ids, hidden_states, self.lm_head, forward_batch
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
                 )
             else:
                 return self.pooler(hidden_states, forward_batch)
@@ -447,5 +458,20 @@ class Qwen3ForCausalLM(nn.Module):
     def load_kv_cache_scales(self, quantization_param_path: str) -> None:
         self.model.load_kv_cache_scales(quantization_param_path)
 
+    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
+        if not self.pp_group.is_last_rank:
+            return
+
+        self.capture_aux_hidden_states = True
+        if layer_ids is None:
+            num_layers = self.config.num_hidden_layers
+            self.model.layers_to_capture = [
+                2,
+                num_layers // 2,
+                num_layers - 3,
+            ]  # Specific layers for EAGLE3 support
+        else:
+            self.model.layers_to_capture = [val + 1 for val in layer_ids]
+
 
 EntryClass = Qwen3ForCausalLM