Revert "[V1] [Bugfix] eagle bugfix and enable correct lm_head for multimodal (#18034)" (#18600)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/transformers_utils/configs/eagle.py

@@ -70,6 +70,7 @@ class EAGLEConfig(PretrainedConfig):
 
         if self.model is not None:
             for k, v in self.model.to_dict().items():
+                if not hasattr(self, k):
                     setattr(self, k, v)
 
     @classmethod

vllm/v1/spec_decode/eagle.py

@@ -9,7 +9,6 @@ from vllm.distributed.parallel_state import get_pp_group
 from vllm.forward_context import set_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model
-from vllm.model_executor.models import supports_multimodal
 from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
 from vllm.triton_utils import tl, triton
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
@@ -311,9 +310,6 @@ class EagleProposer:
         if self.vllm_config.speculative_config.method != "eagle3" and \
                 hasattr(target_model, "lm_head"):
             logger.info("Loading EAGLE LM head weights from the target model.")
-            if supports_multimodal(target_model):
-                self.model.lm_head = target_model.get_language_model().lm_head
-            else:
             self.model.lm_head = target_model.lm_head
 
     @torch.inference_mode()