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

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

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)