[Misc] unify variable for LLM instance (#20996)

Signed-off-by: Andy Xie <andy.xning@gmail.com>

tests/models/multimodal/generation/test_pixtral.py

@@ -180,8 +180,7 @@ def test_chat(
     ) as vllm_model:
         outputs = []
         for msg in MSGS:
-            output = vllm_model.model.chat(msg,
-                                           sampling_params=SAMPLING_PARAMS)
+            output = vllm_model.llm.chat(msg, sampling_params=SAMPLING_PARAMS)
 
             outputs.extend(output)
 
@@ -217,7 +216,7 @@ def test_multi_modal_placeholders(vllm_runner, prompt,
             max_model_len=8192,
             limit_mm_per_prompt=LIMIT_MM_PER_PROMPT,
     ) as vllm_model:
-        outputs = vllm_model.model.generate(prompt)
+        outputs = vllm_model.llm.generate(prompt)
 
         assert len(outputs) == 1, f"{len(outputs)=}"
         output: RequestOutput = outputs[0]

tests/models/multimodal/generation/test_whisper.py

@@ -106,7 +106,7 @@ def run_test(
             tensor_parallel_size=tensor_parallel_size,
             distributed_executor_backend=distributed_executor_backend,
     ) as vllm_model:
-        llm = vllm_model.model
+        llm = vllm_model.llm
 
         sampling_params = SamplingParams(
             temperature=0,

tests/models/multimodal/generation/vlm_utils/core.py

@@ -85,7 +85,7 @@ def run_test(
                      enforce_eager=enforce_eager,
                      task=task,
                      **vllm_runner_kwargs_) as vllm_model:
-        tokenizer = vllm_model.model.get_tokenizer()
+        tokenizer = vllm_model.llm.get_tokenizer()
 
         vllm_kwargs: dict[str, Any] = {}
         if get_stop_token_ids is not None:

tests/models/multimodal/pooling/test_dse_qwen2_vl.py

@@ -96,7 +96,7 @@ def _run_test(
                      dtype=dtype,
                      enforce_eager=True,
                      max_model_len=8192) as vllm_model:
-        tokenizer = vllm_model.model.get_tokenizer()
+        tokenizer = vllm_model.llm.get_tokenizer()
         texts = [
             # this is necessary because vllm_model.embed will not apply any
             # templating to the prompt, and therefore lacks an image_pad

tests/models/multimodal/pooling/test_jinavl_reranker.py

@@ -56,7 +56,7 @@ def vllm_reranker(
             mm_processor_kwargs=mm_processor_kwargs,
             limit_mm_per_prompt=limit_mm_per_prompt,
     ) as vllm_model:
-        outputs = vllm_model.model.score(query, documents)
+        outputs = vllm_model.llm.score(query, documents)
 
     return [output.outputs.score for output in outputs]