Eagle: GPT-OSS Eagle v2 support (#14920)

Signed-off-by: Izzy Putterman <iputterman@nvidia.com>

python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py

@@ -100,7 +100,10 @@ class EAGLEDraftExtendCudaGraphRunner:
                 (3, self.max_num_token), dtype=torch.int64
             )
 
-            if self.eagle_worker.speculative_algorithm.is_eagle3():
+            if (
+                self.eagle_worker.speculative_algorithm.is_eagle3()
+                and self.eagle_worker.eagle_use_aux_hidden_state
+            ):
                 self.hidden_states = torch.zeros(
                     (
                         self.max_num_token,

python/sglang/srt/speculative/eagle_worker.py

@@ -186,6 +186,15 @@ class EAGLEWorker(TpModelWorker):
         self.draft_tp_context = (
             draft_tp_context if server_args.enable_dp_attention else empty_context
         )
+        self.eagle_use_aux_hidden_state = False
+        if self.speculative_algorithm.is_eagle3():
+            self.eagle_use_aux_hidden_state = True
+            eagle_config = getattr(
+                self.draft_model_runner.model_config.hf_config, "eagle_config", {}
+            )
+            self.eagle_use_aux_hidden_state = eagle_config.get(
+                "use_aux_hidden_state", True
+            )
         with self.draft_tp_context(
             self.draft_model_runner.tp_group
         ), speculative_moe_backend_context(), speculative_moe_a2a_backend_context():
@@ -895,6 +904,7 @@ class EAGLEWorker(TpModelWorker):
             hidden_size = (
                 self.model_config.hidden_size * 3
                 if self.speculative_algorithm.is_eagle3()
+                and self.eagle_use_aux_hidden_state
                 else self.model_config.hidden_size
             )
             batch.spec_info = EagleDraftInput.create_idle_input(