diff --git a/vllm/transformers_utils/configs/speculators/algos.py b/vllm/transformers_utils/configs/speculators/algos.py
index 405d5f5de1d1..31b60796d128 100644
--- a/vllm/transformers_utils/configs/speculators/algos.py
+++ b/vllm/transformers_utils/configs/speculators/algos.py
@@ -65,6 +65,7 @@ def update_dflash(config_dict: dict, pre_trained_config: dict) -> None:
     if config_dict.get("target_hidden_size") is not None:
         pre_trained_config["target_hidden_size"] = config_dict["target_hidden_size"]
 
+    # TODO: does this need to be shifted by 1 like in gpu_model_runner?
     aux_layer_ids = config_dict["aux_hidden_state_layer_ids"]
     pre_trained_config["eagle_aux_hidden_state_layer_ids"] = aux_layer_ids
 
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index b6bc942fc857..d7df9e4eeebe 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -4938,7 +4938,8 @@ def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
         if not layer_ids:
             dflash_config = getattr(hf_config, "dflash_config", None)
             if dflash_config and isinstance(dflash_config, dict):
-                layer_ids = dflash_config.get("target_layer_ids")
+                # Add 1 to convert DFlash's aux layer id semantics
+                layer_ids = [i + 1 for i in dflash_config.get("target_layer_ids", [])]
 
         if layer_ids and isinstance(layer_ids, (list, tuple)):
             return tuple(layer_ids)