[V1][Spec Decode][Bugfix] Load quantize weights for EAGLE

vllm/transformers_utils/configs/eagle.py

@@ -52,13 +52,15 @@ def __init__(self,
                 assert self.model is not None, \
                     "model should not be None when method is eagle"
                 kwargs["architectures"] = [
-                    f"Eagle{arch}" for arch in self.model.architectures
+                    f"Eagle{arch}" if not arch.startswith("Eagle") \
+                        else arch for arch in self.model.architectures
                 ]
             elif method == "eagle3":
                 assert self.model is not None, \
                     "model should not be None when method is eagle3"
                 kwargs["architectures"] = [
-                    f"Eagle3{arch}" for arch in self.model.architectures
+                    f"Eagle3{arch}" if not arch.startswith("Eagle3") \
+                        else arch for arch in self.model.architectures
                 ]
             else:
                 raise ValueError(f"Invalid method {method}. \

vllm/v1/spec_decode/eagle.py

@@ -9,7 +9,8 @@
 from vllm.forward_context import set_forward_context
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model_loader
-from vllm.model_executor.model_loader.utils import set_default_torch_dtype
+from vllm.model_executor.model_loader.utils import (
+    process_weights_after_loading, set_default_torch_dtype)
 from vllm.model_executor.models import ModelRegistry
 from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
 from vllm.triton_utils import tl, triton
@@ -308,6 +309,9 @@ def load_model(self, target_model: nn.Module) -> None:
         loaded_weights = self.model.load_weights(
             loader.get_all_weights(draft_model_config, self.model))
 
+        process_weights_after_loading(self.model, draft_model_config,
+                                      target_device)
+
         # share embed_tokens with the target model if needed
         if get_pp_group().world_size == 1:
             assert "model.embed_tokens.weight" not in loaded_weights, \