[Mypy] Better fixes for the `mypy` issues in `vllm/config` (#37902)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

benchmarks/benchmark_long_document_qa_throughput.py

@@ -42,7 +42,6 @@ details.
 
 import random
 import time
-from dataclasses import fields
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
@@ -124,7 +123,7 @@ def main(args):
 
     # Create the LLM engine
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM.from_engine_args(engine_args)
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
 
     print("------warm up------")

benchmarks/benchmark_prefix_caching.py

@@ -32,7 +32,6 @@ import dataclasses
 import json
 import random
 import time
-from dataclasses import fields
 
 from transformers import PreTrainedTokenizerBase
 
@@ -197,7 +196,7 @@ def main(args):
 
     engine_args = EngineArgs.from_cli_args(args)
 
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM.from_engine_args(engine_args)
 
     sampling_params = SamplingParams(
         temperature=0,

benchmarks/benchmark_prioritization.py

@@ -6,7 +6,6 @@ import argparse
 import json
 import random
 import time
-from dataclasses import fields
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
@@ -79,7 +78,7 @@ def run_vllm(
 ) -> float:
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM.from_engine_args(engine_args)
 
     assert all(
         llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])

examples/offline_inference/audio_language.py

@@ -9,7 +9,6 @@ on HuggingFace model repository.
 """
 
 import os
-from dataclasses import asdict
 from typing import Any, NamedTuple
 
 from huggingface_hub import snapshot_download
@@ -633,7 +632,7 @@ def main(args):
         req_data.engine_args.limit_mm_per_prompt or {}
     )
 
-    engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
+    engine_args = vars(req_data.engine_args) | {"seed": args.seed}
     if args.tensor_parallel_size is not None:
         engine_args["tensor_parallel_size"] = args.tensor_parallel_size
     llm = LLM(**engine_args)

examples/offline_inference/encoder_decoder_multimodal.py

@@ -8,7 +8,6 @@ the explicit/implicit prompt format on enc-dec LMMs for text generation.
 import os
 import time
 from collections.abc import Sequence
-from dataclasses import asdict
 from typing import NamedTuple
 
 from vllm import LLM, EngineArgs, PromptType, SamplingParams
@@ -91,13 +90,12 @@ def main(args):
     req_data = model_example_map[model]()
 
     # Disable other modalities to save memory
+    engine_args = req_data.engine_args
     default_limits = {"image": 0, "video": 0, "audio": 0}
-    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
-        req_data.engine_args.limit_mm_per_prompt or {}
-    )
-
-    engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
-    llm = LLM(**engine_args)
+    limit_mm_per_prompt = default_limits | (engine_args.limit_mm_per_prompt or {})
+    engine_args.limit_mm_per_prompt = limit_mm_per_prompt
+    engine_args.seed = args.seed
+    llm = LLM.from_engine_args(engine_args)
 
     prompts = req_data.prompts
 

examples/offline_inference/load_sharded_state.py

@@ -20,8 +20,6 @@ python load_sharded_state.py \
     --max-tokens 50
 """
 
-import dataclasses
-
 from vllm import LLM, EngineArgs, SamplingParams
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
@@ -64,7 +62,7 @@ def main():
     print(f"Tensor parallel size: {engine_args.tensor_parallel_size}")
 
     # Load the model using engine args
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM.from_engine_args(engine_args)
 
     # Prepare sampling parameters
     sampling_params = SamplingParams(

examples/offline_inference/save_sharded_state.py

@@ -21,7 +21,6 @@ llm = LLM(
 )
 """
 
-import dataclasses
 import os
 import shutil
 from pathlib import Path
@@ -60,7 +59,7 @@ def main(args):
     if not Path(model_path).is_dir():
         raise ValueError("model path must be a local directory")
     # Create LLM instance from arguments
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM.from_engine_args(engine_args)
     # Prepare output directory
     Path(args.output).mkdir(exist_ok=True)
     # Dump worker states to output directory

examples/offline_inference/vision_language.py

@@ -11,7 +11,6 @@ on HuggingFace model repository.
 import os
 import random
 from contextlib import contextmanager
-from dataclasses import asdict
 from typing import NamedTuple
 
 from huggingface_hub import snapshot_download
@@ -2434,13 +2433,13 @@ def main(args):
         req_data.engine_args.limit_mm_per_prompt or {}
     )
 
-    engine_args = asdict(req_data.engine_args) | {
-        "seed": args.seed,
-        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
-    }
+    engine_args = req_data.engine_args
+    engine_args.seed = args.seed
+    mm_processor_cache_gb = 0 if args.disable_mm_processor_cache else 4
+    engine_args.mm_processor_cache_gb = mm_processor_cache_gb
     if args.tensor_parallel_size is not None:
-        engine_args["tensor_parallel_size"] = args.tensor_parallel_size
-    llm = LLM(**engine_args)
+        engine_args.tensor_parallel_size = args.tensor_parallel_size
+    llm = LLM.from_engine_args(engine_args)
 
     # Don't want to check the flag multiple times, so just hijack `prompts`.
     prompts = (

examples/offline_inference/vision_language_multi_image.py

@@ -8,7 +8,6 @@ using the chat template defined by the model.
 
 import os
 from argparse import Namespace
-from dataclasses import asdict
 from typing import NamedTuple
 
 from huggingface_hub import snapshot_download
@@ -1481,10 +1480,11 @@ def run_generate(
 ):
     req_data = model_example_map[model](question, image_urls)
 
-    engine_args = asdict(req_data.engine_args) | {"seed": seed}
+    engine_args = req_data.engine_args
+    engine_args.seed = seed
     if tensor_parallel_size is not None:
-        engine_args["tensor_parallel_size"] = tensor_parallel_size
-    llm = LLM(**engine_args)
+        engine_args.tensor_parallel_size = tensor_parallel_size
+    llm = LLM.from_engine_args(engine_args)
 
     sampling_params = SamplingParams(
         temperature=0.0, max_tokens=256, stop_token_ids=req_data.stop_token_ids
@@ -1521,10 +1521,11 @@ def run_chat(
         req_data.engine_args.limit_mm_per_prompt or {}
     )
 
-    engine_args = asdict(req_data.engine_args) | {"seed": seed}
+    engine_args = req_data.engine_args
+    engine_args.seed = seed
     if tensor_parallel_size is not None:
-        engine_args["tensor_parallel_size"] = tensor_parallel_size
-    llm = LLM(**engine_args)
+        engine_args.tensor_parallel_size = tensor_parallel_size
+    llm = LLM.from_engine_args(engine_args)
 
     sampling_params = (
         SamplingParams(

examples/pooling/embed/vision_embedding_offline.py

@@ -10,12 +10,11 @@ on HuggingFace model repository.
 """
 
 import argparse
-from dataclasses import asdict
 from pathlib import Path
 
 from PIL.Image import Image
 
-from vllm import LLM, EngineArgs
+from vllm import LLM
 from vllm.multimodal.utils import fetch_image
 from vllm.utils.print_utils import print_embeddings
 
@@ -28,14 +27,13 @@ multi_modal_data = {"image": fetch_image(image_url)}
 
 
 def run_clip(seed: int):
-    engine_args = EngineArgs(
+    llm = LLM(
         model="openai/clip-vit-base-patch32",
         runner="pooling",
         limit_mm_per_prompt={"image": 1},
+        seed=seed,
     )
 
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
-
     print("Text embedding output:")
     outputs = llm.embed(text, use_tqdm=False)
     print_embeddings(outputs[0].outputs.embedding)
@@ -53,15 +51,14 @@ def run_clip(seed: int):
 
 
 def run_e5_v(seed: int):
-    engine_args = EngineArgs(
+    llm = LLM(
         model="royokong/e5-v",
         runner="pooling",
         max_model_len=4096,
         limit_mm_per_prompt={"image": 1},
+        seed=seed,
     )
 
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
-
     llama3_template = "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n"  # noqa: E501
 
     print("Text embedding output:")
@@ -108,20 +105,20 @@ def run_qwen3_vl(seed: int):
 
         multi_modal_data["image"] = post_process_image(multi_modal_data["image"])
 
-    engine_args = EngineArgs(
-        model="Qwen/Qwen3-VL-Embedding-2B",
-        runner="pooling",
-        max_model_len=8192,
-        limit_mm_per_prompt={"image": 1},
-        mm_processor_kwargs={"do_resize": False} if smart_resize is not None else None,
-    )
     default_instruction = "Represent the user's input."
     image_placeholder = "<|vision_start|><|image_pad|><|vision_end|>"
     prompt_text = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant\n"
     prompt_image = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{image_placeholder}<|im_end|>\n<|im_start|>assistant\n"
     prompt_image_text = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{image_placeholder}{text}<|im_end|>\n<|im_start|>assistant\n"
 
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
+    llm = LLM(
+        model="Qwen/Qwen3-VL-Embedding-2B",
+        runner="pooling",
+        max_model_len=8192,
+        limit_mm_per_prompt={"image": 1},
+        mm_processor_kwargs={"do_resize": False} if smart_resize is not None else None,
+        seed=seed,
+    )
 
     print("Text embedding output:")
     outputs = llm.embed(prompt_text, use_tqdm=False)
@@ -149,14 +146,13 @@ def run_qwen3_vl(seed: int):
 
 
 def run_siglip(seed: int):
-    engine_args = EngineArgs(
+    llm = LLM(
         model="google/siglip-base-patch16-224",
         runner="pooling",
         limit_mm_per_prompt={"image": 1},
+        seed=seed,
     )
 
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
-
     print("Text embedding output:")
     outputs = llm.embed(text, use_tqdm=False)
     print_embeddings(outputs[0].outputs.embedding)
@@ -174,16 +170,15 @@ def run_siglip(seed: int):
 
 
 def run_vlm2vec_phi3v(seed: int):
-    engine_args = EngineArgs(
+    llm = LLM(
         model="TIGER-Lab/VLM2Vec-Full",
         runner="pooling",
         max_model_len=4096,
         trust_remote_code=True,
         mm_processor_kwargs={"num_crops": 4},
         limit_mm_per_prompt={"image": 1},
+        seed=seed,
     )
-
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
     image_token = "<|image_1|>"
 
     print("Text embedding output:")
@@ -259,7 +254,7 @@ def run_vlm2vec_qwen2vl(seed: int):
     processor.save_pretrained(merged_path)
     print("Done!")
 
-    engine_args = EngineArgs(
+    llm = LLM(
         model=merged_path,
         runner="pooling",
         max_model_len=4096,
@@ -268,9 +263,8 @@ def run_vlm2vec_qwen2vl(seed: int):
             "max_pixels": 12845056,
         },
         limit_mm_per_prompt={"image": 1},
+        seed=seed,
     )
-
-    llm = LLM(**asdict(engine_args) | {"seed": seed})
     image_token = "<|image_pad|>"
 
     print("Text embedding output:")

examples/pooling/score/vision_reranker_offline.py

@@ -10,7 +10,6 @@ multimodal documents (text + images/videos).
 
 from argparse import Namespace
 from collections.abc import Callable
-from dataclasses import asdict
 from pathlib import Path
 from typing import NamedTuple
 
@@ -125,7 +124,7 @@ def main(args: Namespace):
     model_request = model_example_map[args.model_name]()
     engine_args = model_request.engine_args
 
-    llm = LLM(**asdict(engine_args))
+    llm = LLM.from_engine_args(engine_args)
 
     print("Query: string & Document: string")
     outputs = llm.score(query, document)

tests/compile/test_config.py

@@ -414,9 +414,12 @@ def test_cudagraph_sizes_post_init(
         ctx,
         patch("vllm.config.parallel.cuda_device_count_stateless", return_value=tp_size),
     ):
+        kwargs = {}
+        if cudagraph_capture_sizes is not None:
+            kwargs["cudagraph_capture_sizes"] = cudagraph_capture_sizes
+        if max_cudagraph_capture_size is not None:
+            kwargs["max_cudagraph_capture_size"] = max_cudagraph_capture_size
         compilation_config = CompilationConfig(
-            cudagraph_capture_sizes=cudagraph_capture_sizes,
-            max_cudagraph_capture_size=max_cudagraph_capture_size,
             pass_config=PassConfig(
                 enable_sp=enable_sp,
                 fuse_norm_quant=True,
@@ -425,6 +428,7 @@ def test_cudagraph_sizes_post_init(
                 sp_min_token_num=512 if enable_sp else None,
             ),
             cudagraph_mode=cudagraph_mode,
+            **kwargs,
         )
         engine_args = EngineArgs(
             model="facebook/opt-125m",

tests/entrypoints/offline_mode/test_offline_mode.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Tests for HF_HUB_OFFLINE mode"""
 
-import dataclasses
 import importlib
 import sys
 
@@ -12,7 +11,6 @@ import urllib3
 
 from vllm import LLM
 from vllm.distributed import cleanup_dist_env_and_memory
-from vllm.engine.arg_utils import EngineArgs
 
 MODEL_CONFIGS = [
     {
@@ -160,8 +158,7 @@ def test_model_from_huggingface_offline(monkeypatch: pytest.MonkeyPatch):
             # Need to re-import huggingface_hub
             # and friends to set up offline mode
             _re_import_modules()
-            engine_args = EngineArgs(model="facebook/opt-125m")
-            LLM(**dataclasses.asdict(engine_args))
+            LLM(model="facebook/opt-125m")
         finally:
             # Reset the environment after the test
             # NB: Assuming tests are run in online mode

tests/models/multimodal/generation/test_keye.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from dataclasses import asdict
 from typing import NamedTuple
 
 import pytest
@@ -29,14 +28,6 @@ def test_keye_vl(image_assets, question: str):
     images = [asset.pil_image for asset in image_assets]
     image_urls = [encode_image_url(image) for image in images]
 
-    engine_args = EngineArgs(
-        model=MODEL_NAME,
-        trust_remote_code=True,
-        max_model_len=8192,
-        max_num_seqs=5,
-        limit_mm_per_prompt={"image": len(image_urls)},
-    )
-
     placeholders = [{"type": "image", "image": url} for url in image_urls]
     messages = [
         {
@@ -54,8 +45,14 @@ def test_keye_vl(image_assets, question: str):
         messages, tokenize=False, add_generation_prompt=True
     )
 
-    engine_args = asdict(engine_args) | {"seed": 42}
-    llm = LLM(**engine_args)
+    llm = LLM(
+        model=MODEL_NAME,
+        trust_remote_code=True,
+        max_model_len=8192,
+        max_num_seqs=5,
+        limit_mm_per_prompt={"image": len(image_urls)},
+        seed=42,
+    )
 
     sampling_params = SamplingParams(
         temperature=0.0, max_tokens=256, stop_token_ids=None

tests/models/multimodal/generation/test_vit_backend_functionality.py

@@ -7,13 +7,12 @@ This test validates that each multimodal model can successfully generate outputs
 using different ViT attention backends. Tests are parametrized by model and backend.
 """
 
-from dataclasses import asdict
 from typing import Any
 
 import pytest
 from transformers import AutoProcessor
 
-from vllm import LLM, EngineArgs, SamplingParams
+from vllm import LLM, SamplingParams
 from vllm.multimodal.utils import encode_image_url
 from vllm.multimodal.video import sample_frames_from_video
 from vllm.platforms import current_platform
@@ -274,7 +273,7 @@ def run_llm_generate_test(config, mm_encoder_attn_backend, image_assets):
     limit_mm_per_prompt = config.get("limit_mm_per_prompt", {"image": len(images)})
 
     # Create engine
-    engine_args = EngineArgs(
+    llm = LLM(
         model=config["model_name"],
         trust_remote_code=True,
         max_model_len=config["max_model_len"],
@@ -283,11 +282,9 @@ def run_llm_generate_test(config, mm_encoder_attn_backend, image_assets):
         mm_encoder_attn_backend=mm_encoder_attn_backend,
         hf_overrides=dummy_hf_overrides,
         load_format="dummy",
+        seed=42,
     )
 
-    engine_dict = asdict(engine_args) | {"seed": 42}
-    llm = LLM(**engine_dict)
-
     # Generate
     sampling_params = SamplingParams(**config["sampling_params"])
     outputs = llm.generate(
@@ -318,7 +315,7 @@ def run_llm_chat_test(config, mm_encoder_attn_backend, image_assets):
     messages = build_dots_ocr_prompt([stop_sign_image], config)
 
     # Create engine
-    engine_args = EngineArgs(
+    llm = LLM(
         model=config["model_name"],
         trust_remote_code=True,
         max_model_len=config["max_model_len"],
@@ -327,11 +324,9 @@ def run_llm_chat_test(config, mm_encoder_attn_backend, image_assets):
         mm_encoder_attn_backend=mm_encoder_attn_backend,
         hf_overrides=dummy_hf_overrides,
         load_format="dummy",
+        seed=42,
     )
 
-    engine_dict = asdict(engine_args) | {"seed": 42}
-    llm = LLM(**engine_dict)
-
     # Generate using chat
     sampling_params = SamplingParams(**config["sampling_params"])
     outputs = llm.chat(messages=messages, sampling_params=sampling_params)

tests/models/multimodal/generation/test_voxtral_realtime.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import contextlib
-from dataclasses import asdict
 
 import pytest
 import pytest_asyncio
@@ -75,7 +74,7 @@ def tokenizer() -> MistralTokenizer:
 @pytest.fixture
 def engine():
     engine_args = EngineArgs(**ENGINE_CONFIG)
-    llm = LLM(**asdict(engine_args))
+    llm = LLM.from_engine_args(engine_args)
     try:
         yield llm
     finally:

tests/v1/kv_connector/unit/test_example_connector.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from dataclasses import asdict
 from typing import NamedTuple
 
 import pytest
 from PIL import Image
 
-from vllm import LLM, EngineArgs, SamplingParams
+from vllm import LLM, SamplingParams
 from vllm.assets.image import ImageAsset
 from vllm.config import AttentionConfig, KVTransferConfig
 from vllm.multimodal.utils import encode_image_url
@@ -129,24 +128,6 @@ def test_shared_storage_connector_hashes(tmp_path, attn_backend):
     # Using tmp_path as the storage path to store KV
     print(f"KV storage path at: {str(tmp_path)}")
 
-    # Configure the ExampleConnector
-    kv_transfer_config = KVTransferConfig(
-        kv_connector="ExampleConnector",
-        kv_role="kv_both",
-        kv_connector_extra_config={"shared_storage_path": str(tmp_path)},
-    )
-
-    engine_args = EngineArgs(
-        model=MODEL_NAME,
-        max_model_len=8192,
-        max_num_seqs=1,
-        gpu_memory_utilization=0.4,
-        attention_config=AttentionConfig(backend=attn_backend),
-        enforce_eager=True,
-        kv_transfer_config=kv_transfer_config,
-        limit_mm_per_prompt={"image": 2},
-    )
-
     # don't put this import at the top level
     # it will call torch.accelerator.device_count()
     from transformers import AutoProcessor
@@ -163,8 +144,20 @@ def test_shared_storage_connector_hashes(tmp_path, attn_backend):
     assert image_1 != image_2, "The images should not be identical"
 
     # Create the LLM instance
-    engine_args = asdict(engine_args)
-    llm = LLM(**engine_args)
+    llm = LLM(
+        model=MODEL_NAME,
+        max_model_len=8192,
+        max_num_seqs=1,
+        gpu_memory_utilization=0.4,
+        attention_config=AttentionConfig(backend=attn_backend),
+        enforce_eager=True,
+        kv_transfer_config=KVTransferConfig(
+            kv_connector="ExampleConnector",
+            kv_role="kv_both",
+            kv_connector_extra_config={"shared_storage_path": str(tmp_path)},
+        ),
+        limit_mm_per_prompt={"image": 2},
+    )
 
     # Prepare the input cases
     input_cases = [

vllm/benchmarks/latency.py

@@ -6,7 +6,6 @@ import argparse
 import json
 import os
 import time
-from dataclasses import fields
 from typing import Any
 
 import numpy as np
@@ -85,7 +84,7 @@ 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(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM.from_engine_args(engine_args)
     assert llm.llm_engine.model_config.max_model_len >= (
         args.input_len + args.output_len
     ), (

vllm/benchmarks/mm_processor.py

@@ -17,7 +17,6 @@ import argparse
 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 +224,7 @@ def benchmark_multimodal_processor(
         args.seed = 0
 
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM.from_engine_args(engine_args)
 
     tokenizer = llm.get_tokenizer()
     requests = get_requests(args, tokenizer)

vllm/benchmarks/startup.py

@@ -16,7 +16,6 @@ import shutil
 import tempfile
 import time
 from contextlib import contextmanager
-from dataclasses import fields
 from typing import Any
 
 import numpy as np
@@ -67,7 +66,7 @@ def run_startup_in_subprocess(engine_args, result_queue):
         # Measure total startup time
         start_time = time.perf_counter()
 
-        llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+        llm = LLM.from_engine_args(engine_args)
 
         total_startup_time = time.perf_counter() - start_time