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

tests/entrypoints/llm/test_prompt_validation.py

@@ -3,7 +3,6 @@
 import pytest
 
 from vllm import LLM
-from vllm.config import LoadFormat
 
 
 @pytest.fixture(autouse=True)
@@ -15,17 +14,13 @@ def v1(run_with_both_engines):
 
 
 def test_empty_prompt():
-    llm = LLM(model="s3://vllm-ci-model-weights/gpt2",
-              load_format=LoadFormat.RUNAI_STREAMER,
-              enforce_eager=True)
+    llm = LLM(model="openai-community/gpt2", enforce_eager=True)
     with pytest.raises(ValueError, match='Prompt cannot be empty'):
         llm.generate([""])
 
 
 @pytest.mark.skip_v1
 def test_out_of_vocab_token():
-    llm = LLM(model="s3://vllm-ci-model-weights/gpt2",
-              load_format=LoadFormat.RUNAI_STREAMER,
-              enforce_eager=True)
+    llm = LLM(model="openai-community/gpt2", enforce_eager=True)
     with pytest.raises(ValueError, match='out of vocabulary'):
         llm.generate({"prompt_token_ids": [999999]})