diff --git a/benchmarks/benchmark_long_document_qa_throughput.py b/benchmarks/benchmark_long_document_qa_throughput.py
index f64fd09ba..b50b310fd 100644
--- a/benchmarks/benchmark_long_document_qa_throughput.py
+++ b/benchmarks/benchmark_long_document_qa_throughput.py
@@ -40,9 +40,9 @@ LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
 details.
 """
 
-import dataclasses
 import random
 import time
+from dataclasses import fields
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
@@ -124,7 +124,7 @@ def main(args):
 
     # Create the LLM engine
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
 
     print("------warm up------")
diff --git a/benchmarks/benchmark_prefix_caching.py b/benchmarks/benchmark_prefix_caching.py
index e6391134f..e7759616e 100644
--- a/benchmarks/benchmark_prefix_caching.py
+++ b/benchmarks/benchmark_prefix_caching.py
@@ -32,6 +32,7 @@ import dataclasses
 import json
 import random
 import time
+from dataclasses import fields
 
 from transformers import PreTrainedTokenizerBase
 
@@ -196,7 +197,7 @@ def main(args):
 
     engine_args = EngineArgs.from_cli_args(args)
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     sampling_params = SamplingParams(
         temperature=0,
diff --git a/benchmarks/benchmark_prioritization.py b/benchmarks/benchmark_prioritization.py
index a35db0063..d83bb7e17 100644
--- a/benchmarks/benchmark_prioritization.py
+++ b/benchmarks/benchmark_prioritization.py
@@ -3,10 +3,10 @@
 """Benchmark offline prioritization."""
 
 import argparse
-import dataclasses
 import json
 import random
 import time
+from dataclasses import fields
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
@@ -79,7 +79,7 @@ def run_vllm(
 ) -> float:
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     assert all(
         llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])
diff --git a/vllm/benchmarks/latency.py b/vllm/benchmarks/latency.py
index a9d149666..758e5efed 100644
--- a/vllm/benchmarks/latency.py
+++ b/vllm/benchmarks/latency.py
@@ -3,10 +3,10 @@
 """Benchmark the latency of processing a single batch of requests."""
 
 import argparse
-import dataclasses
 import json
 import os
 import time
+from dataclasses import fields
 from typing import Any
 
 import numpy as np
@@ -85,7 +85,7 @@ def main(args: argparse.Namespace):
 
     # NOTE(woosuk): If the request cannot be processed in a single batch,
     # the engine will automatically process the request in multiple batches.
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
     assert llm.llm_engine.model_config.max_model_len >= (
         args.input_len + args.output_len
     ), (
diff --git a/vllm/benchmarks/mm_processor.py b/vllm/benchmarks/mm_processor.py
index 5900bbf99..4f31af0e0 100644
--- a/vllm/benchmarks/mm_processor.py
+++ b/vllm/benchmarks/mm_processor.py
@@ -14,10 +14,10 @@ Run:
 """
 
 import argparse
-import dataclasses
 import json
 import time
 from collections import defaultdict
+from dataclasses import fields
 from datetime import datetime
 from typing import TYPE_CHECKING, Any, Literal
 
@@ -225,7 +225,7 @@ def benchmark_multimodal_processor(
         args.seed = 0
 
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     tokenizer = llm.get_tokenizer()
     requests = get_requests(args, tokenizer)
diff --git a/vllm/benchmarks/startup.py b/vllm/benchmarks/startup.py
index 005625f61..405299938 100644
--- a/vllm/benchmarks/startup.py
+++ b/vllm/benchmarks/startup.py
@@ -9,7 +9,6 @@ and cache operations) for both cold and warm scenarios:
 """
 
 import argparse
-import dataclasses
 import json
 import multiprocessing
 import os
@@ -17,6 +16,7 @@ import shutil
 import tempfile
 import time
 from contextlib import contextmanager
+from dataclasses import fields
 from typing import Any
 
 import numpy as np
@@ -67,7 +67,7 @@ def run_startup_in_subprocess(engine_args, result_queue):
         # Measure total startup time
         start_time = time.perf_counter()
 
-        llm = LLM(**dataclasses.asdict(engine_args))
+        llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
         total_startup_time = time.perf_counter() - start_time
 
diff --git a/vllm/benchmarks/sweep/serve_workload.py b/vllm/benchmarks/sweep/serve_workload.py
index ca7ba09a5..a47668ff1 100644
--- a/vllm/benchmarks/sweep/serve_workload.py
+++ b/vllm/benchmarks/sweep/serve_workload.py
@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
 import math
-from dataclasses import asdict, dataclass
+from dataclasses import dataclass, fields
 from pathlib import Path
 from typing import ClassVar, Literal, get_args
 
@@ -267,7 +267,7 @@ class SweepServeWorkloadArgs(SweepServeArgs):
         base_args = SweepServeArgs.from_cli_args(args)
 
         return cls(
-            **asdict(base_args),
+            **{f.name: getattr(base_args, f.name) for f in fields(base_args)},
             workload_var=args.workload_var,
             workload_iters=args.workload_iters,
         )
diff --git a/vllm/benchmarks/throughput.py b/vllm/benchmarks/throughput.py
index 1af8cf900..4c6379d67 100644
--- a/vllm/benchmarks/throughput.py
+++ b/vllm/benchmarks/throughput.py
@@ -3,12 +3,12 @@
 """Benchmark offline inference throughput."""
 
 import argparse
-import dataclasses
 import json
 import os
 import random
 import time
 import warnings
+from dataclasses import fields
 from typing import Any
 
 import torch
@@ -53,7 +53,7 @@ def run_vllm(
 ) -> tuple[float, list[RequestOutput] | None]:
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
     assert all(
         llm.llm_engine.model_config.max_model_len
         >= (request.prompt_len + request.expected_output_len)
@@ -141,7 +141,7 @@ def run_vllm_chat(
     """
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     assert all(
         llm.llm_engine.model_config.max_model_len