Update ensemble_qwen.py

ensemble_qwen.py

@@ -24,10 +24,9 @@ class EnsembleConfig(Qwen3Config):
 
 class EnsembleForCausalLM(PreTrainedModel, GenerationMixin):
     config_class = EnsembleConfig
-    _supports_flash_attn_2 = True          # <── NEW
-    _supports_sdpa        = True           # optional, lets users pick "sdpa"
+    _supports_flash_attn_2 = True
+    _supports_sdpa        = True
     main_input_name = "input_ids"
-    _tied_weights_keys = ["model_a.lm_head.weight", "model_b.lm_head.weight"]
     _tp_plan = {"model_a.lm_head": "colwise_rep", "model_b.lm_head": "colwise_rep"}
     
     def __init__(self, config: EnsembleConfig):