diff --git a/vllm/model_executor/input_metadata.py b/vllm/model_executor/input_metadata.py
index ebba0ba0a261..01bba70ac10a 100644
--- a/vllm/model_executor/input_metadata.py
+++ b/vllm/model_executor/input_metadata.py
@@ -1,5 +1,5 @@
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, fields
+from typing import Optional, Any, Dict
 
 import torch
 
@@ -31,3 +31,12 @@ class InputMetadata:
     def __post_init__(self):
         # will not appear in the __repr__ and __init__
         self.attn_bias = None
+
+    def asdict_zerocopy(self) -> Dict[str, Any]:
+        """Similar to dataclasses.asdict, but avoids deepcopying."""
+        # Note that if we add dataclasses as fields, they will need
+        # similar handling.
+        return {
+            field.name: getattr(self, field.name)
+            for field in fields(self)
+        }
diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py
index 1ef783da6d08..27213887ed26 100644
--- a/vllm/worker/model_runner.py
+++ b/vllm/worker/model_runner.py
@@ -1,5 +1,4 @@
 import contextlib
-import dataclasses
 import time
 from typing import Dict, List, Optional, Tuple, Set, Union
 
@@ -527,7 +526,7 @@ class ModelRunner:
                 "lora_requests": lora_requests,
                 "lora_mapping": lora_mapping,
             }
-            metadata_dict.update(dataclasses.asdict(input_metadata))
+            metadata_dict.update(input_metadata.asdict_zerocopy())
             broadcast_tensor_dict(metadata_dict, src=0)
         else:
             metadata_dict = broadcast_tensor_dict(src=0)