[ci] Use env var to control whether to use S3 bucket in CI (#13634)

tests/entrypoints/llm/test_chat.py

@@ -5,17 +5,12 @@ from typing import List
 import pytest
 
 from vllm import LLM
-from vllm.config import LoadFormat
 
-from ...conftest import MODEL_WEIGHTS_S3_BUCKET
 from ..openai.test_vision import TEST_IMAGE_URLS
 
-RUNAI_STREAMER_LOAD_FORMAT = LoadFormat.RUNAI_STREAMER
-
 
 def test_chat():
-    llm = LLM(model=f"{MODEL_WEIGHTS_S3_BUCKET}/Llama-3.2-1B-Instruct",
-              load_format=RUNAI_STREAMER_LOAD_FORMAT)
+    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct")
 
     prompt1 = "Explain the concept of entropy."
     messages = [
@@ -33,8 +28,7 @@ def test_chat():
 
 
 def test_multi_chat():
-    llm = LLM(model=f"{MODEL_WEIGHTS_S3_BUCKET}/Llama-3.2-1B-Instruct",
-              load_format=RUNAI_STREAMER_LOAD_FORMAT)
+    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct")
 
     prompt1 = "Explain the concept of entropy."
     prompt2 = "Explain what among us is."
@@ -71,8 +65,7 @@ def test_multi_chat():
                          [[TEST_IMAGE_URLS[0], TEST_IMAGE_URLS[1]]])
 def test_chat_multi_image(image_urls: List[str]):
     llm = LLM(
-        model=f"{MODEL_WEIGHTS_S3_BUCKET}/Phi-3.5-vision-instruct",
-        load_format=RUNAI_STREAMER_LOAD_FORMAT,
+        model="microsoft/Phi-3.5-vision-instruct",
         dtype="bfloat16",
         max_model_len=4096,
         max_num_seqs=5,