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

.buildkite/test-pipeline.yaml

@@ -278,7 +278,7 @@ steps:
   command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py
   parallelism: 4
 
-- label: "PyTorch Fullgraph Smoke Test" # 9min
+- label: PyTorch Fullgraph Smoke Test # 9min
   fast_check: true
   source_file_dependencies:
   - vllm/
@@ -289,7 +289,7 @@ steps:
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
 
-- label: "PyTorch Fullgraph Test" # 18min
+- label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
   - vllm/
   - tests/compile

tests/basic_correctness/test_basic_correctness.py

@@ -9,7 +9,6 @@ import weakref
 import pytest
 
 from vllm import LLM
-from vllm.config import LoadFormat
 from vllm.platforms import current_platform
 
 from ..conftest import VllmRunner
@@ -34,7 +33,7 @@ def v1(run_with_both_engines):
 
 def test_vllm_gc_ed():
     """Verify vllm instance is GC'ed when it is deleted"""
-    llm = LLM("distilbert/distilgpt2", load_format=LoadFormat.RUNAI_STREAMER)
+    llm = LLM("distilbert/distilgpt2")
     weak_llm = weakref.ref(llm)
     del llm
     # If there's any circular reference to vllm, this fails
@@ -43,10 +42,10 @@ def test_vllm_gc_ed():
 
 
 @pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("backend", ["FLASH_ATTN", "XFORMERS", "FLASHINFER"])
+@pytest.mark.parametrize("backend", ["FLASH_ATTN"])
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [5])
-@pytest.mark.parametrize("enforce_eager", [False, True])
+@pytest.mark.parametrize("enforce_eager", [False])
 def test_models(
     hf_runner,
     model: str,
@@ -97,8 +96,8 @@ def test_models(
     "test_suite", [
         ("distilbert/distilgpt2", "ray", "", "L4"),
         ("distilbert/distilgpt2", "mp", "", "L4"),
-        ("meta-llama/Llama-2-7b-hf", "ray", "", "L4"),
-        ("meta-llama/Llama-2-7b-hf", "mp", "", "L4"),
+        ("meta-llama/Llama-3.2-1B-Instruct", "ray", "", "L4"),
+        ("meta-llama/Llama-3.2-1B-Instruct", "mp", "", "L4"),
         ("distilbert/distilgpt2", "ray", "", "A100"),
         ("distilbert/distilgpt2", "mp", "", "A100"),
         ("distilbert/distilgpt2", "mp", "FLASHINFER", "A100"),

tests/basic_correctness/test_cumem.py

@@ -4,11 +4,9 @@ import pytest
 import torch
 
 from vllm import LLM, SamplingParams
-from vllm.config import LoadFormat
 from vllm.device_allocator.cumem import CuMemAllocator
 from vllm.utils import GiB_bytes
 
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
 from ..utils import fork_new_process_for_each_test
 
 
@@ -121,7 +119,7 @@ def test_cumem_with_cudagraph():
     "model, use_v1",
     [
         # sleep mode with safetensors
-        (f"{MODEL_WEIGHTS_S3_BUCKET}/meta-llama/Llama-3.2-1B", True),
+        ("meta-llama/Llama-3.2-1B", True),
         # sleep mode with pytorch checkpoint
         ("facebook/opt-125m", False),
     ])
@@ -130,10 +128,7 @@ def test_end_to_end(model: str, use_v1: bool):
     os.environ["VLLM_USE_V1"] = "1" if use_v1 else "0"
     free, total = torch.cuda.mem_get_info()
     used_bytes_baseline = total - free  # in case other process is running
-    load_format = LoadFormat.AUTO
-    if "Llama" in model:
-        load_format = LoadFormat.RUNAI_STREAMER
-    llm = LLM(model, load_format=load_format, enable_sleep_mode=True)
+    llm = LLM(model, enable_sleep_mode=True)
     prompt = "How are you?"
     sampling_params = SamplingParams(temperature=0, max_tokens=10)
     output = llm.generate(prompt, sampling_params)

tests/conftest.py

@@ -24,7 +24,7 @@ from tests.models.utils import (TokensTextLogprobs,
 from vllm import LLM, SamplingParams
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
-from vllm.config import LoadFormat, TaskOption, TokenizerPoolConfig
+from vllm.config import TaskOption, TokenizerPoolConfig
 from vllm.connections import global_http_connection
 from vllm.distributed import (cleanup_dist_env_and_memory,
                               init_distributed_environment,
@@ -47,70 +47,6 @@ _SYS_MSG = os.path.join(_TEST_DIR, "system_messages", "sonnet3.5_nov2024.txt")
 
 _M = TypeVar("_M")
 
-MODELS_ON_S3 = [
-    "distilbert/distilgpt2",
-    "meta-llama/Llama-2-7b-hf",
-    "meta-llama/Meta-Llama-3-8B",
-    "meta-llama/Llama-3.2-1B",
-    "meta-llama/Llama-3.2-1B-Instruct",
-    "openai-community/gpt2",
-    "ArthurZ/Ilama-3.2-1B",
-    "llava-hf/llava-1.5-7b-hf",
-    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
-    "ai21labs/Jamba-tiny-random",
-    "neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV",
-    "nm-testing/Phi-3-mini-128k-instruct-FP8",
-    "nm-testing/Qwen2-0.5B-Instruct-FP8-SkipQKV",
-    "neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV",
-    "nm-testing/Qwen2-1.5B-Instruct-FP8-K-V",
-    "ModelCloud/Qwen1.5-1.8B-Chat-GPTQ-4bits-dynamic-cfg-with-lm_head-symTrue",
-    "ModelCloud/Qwen1.5-1.8B-Chat-GPTQ-4bits-dynamic-cfg-with-lm_head-symFalse",
-    "AMead10/Llama-3.2-1B-Instruct-AWQ",
-    "shuyuej/Llama-3.2-1B-Instruct-GPTQ",
-    "ModelCloud/Qwen1.5-1.8B-Chat-GPTQ-4bits-dynamic-cfg-with-lm_head",
-    "ModelCloud/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit-10-25-2024",
-    "TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ",
-    "neuralmagic/Meta-Llama-3-8B-Instruct-FP8",
-    "amd/Llama-3.1-8B-Instruct-FP8-KV-Quark-test",
-    "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change",
-    "nm-testing/tinyllama-oneshot-w8-channel-a8-tensor",
-    "nm-testing/asym-w8w8-int8-static-per-tensor-tiny-llama",
-    "neuralmagic/Llama-3.2-1B-quantized.w8a8",
-    "nm-testing/Meta-Llama-3-8B-Instruct-W8A8-Dynamic-Asym",
-    "nm-testing/Meta-Llama-3-8B-Instruct-W8A8-Static-Per-Tensor-Sym",
-    "nm-testing/Meta-Llama-3-8B-Instruct-W8A8-Static-Per-Tensor-Asym",
-    "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change",
-    "nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2",
-    "nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2-asym",
-    "nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2",
-    "nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2-asym",
-    "nm-testing/tinyllama-oneshot-w4a16-channel-v2",
-    "nm-testing/tinyllama-oneshot-w4a16-group128-v2",
-    "nm-testing/tinyllama-oneshot-w8a16-per-channel",
-    "nm-testing/llama7b-one-shot-2_4-w4a16-marlin24-t",
-    "nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test",
-    "nm-testing/TinyLlama-1.1B-compressed-tensors-kv-cache-scheme",
-    "nm-testing/Meta-Llama-3-8B-Instruct-FP8-Dynamic-2of4-testing",
-    "nm-testing/Meta-Llama-3-8B-Instruct-FP8-Static-Per-Tensor-testing",
-    "nm-testing/Meta-Llama-3-8B-Instruct-FP8-Static-testing",
-    "nm-testing/Meta-Llama-3-8B-Instruct-FP8-Dynamic-IA-Per-Tensor-Weight-testing",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-chnl_wts_tensor_act_fp8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-tensor_wts_per_tok_dyn_act_fp8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-tensor_wts_tensor_act_fp8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_int8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-chnl_wts_tensor_act_int8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-tensor_wts_per_tok_dyn_act_int8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-gsm8k-pruned.2of4-tensor_wts_tensor_act_int8-BitM",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-INT8-Dynamic-IA-Per-Channel-Weight-testing",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-INT8-Static-testing",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-INT8-Dynamic-IA-Per-Tensor-Weight-testing",
-    "nm-testing/TinyLlama-1.1B-Chat-v1.0-2of4-Sparse-Dense-Compressor",
-    "nm-testing/llama2.c-stories42M-pruned2.4-compressed",
-]
-
-MODEL_WEIGHTS_S3_BUCKET = "s3://vllm-ci-model-weights"
-
 _PromptMultiModalInput = Union[List[_M], List[List[_M]]]
 
 PromptImageInput = _PromptMultiModalInput[Image.Image]
@@ -742,14 +678,8 @@ class VllmRunner:
         enable_chunked_prefill: bool = False,
         swap_space: int = 4,
         enforce_eager: Optional[bool] = False,
-        load_format: Optional[LoadFormat] = None,
         **kwargs,
     ) -> None:
-        if model_name in MODELS_ON_S3 and not load_format:
-            model_name = (f"{MODEL_WEIGHTS_S3_BUCKET}/{model_name}")
-            load_format = LoadFormat.RUNAI_STREAMER
-        if not load_format:
-            load_format = LoadFormat.AUTO
         self.model = LLM(
             model=model_name,
             task=task,
@@ -764,7 +694,6 @@ class VllmRunner:
             max_model_len=max_model_len,
             block_size=block_size,
             enable_chunked_prefill=enable_chunked_prefill,
-            load_format=load_format,
             **kwargs,
         )
 

tests/engine/test_computed_prefix_blocks.py

@@ -2,16 +2,12 @@
 
 import pytest
 
-from vllm.config import LoadFormat
 from vllm.engine.arg_utils import EngineArgs
 from vllm.engine.llm_engine import LLMEngine
 from vllm.sampling_params import SamplingParams
 
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
 
-
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 @pytest.mark.parametrize("block_size", [16])
 def test_computed_prefix_blocks(model: str, block_size: int):
     # This test checks if we are able to run the engine to completion
@@ -28,7 +24,6 @@ def test_computed_prefix_blocks(model: str, block_size: int):
         "decoration.")
 
     engine_args = EngineArgs(model=model,
-                             load_format=LoadFormat.RUNAI_STREAMER,
                              block_size=block_size,
                              enable_prefix_caching=True)
 

tests/engine/test_detokenization.py

@@ -2,15 +2,11 @@
 
 import pytest
 
-from vllm.config import LoadFormat
 from vllm.entrypoints.llm import LLM
 from vllm.sampling_params import SamplingParams
 
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
 
-
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_computed_prefix_blocks(model: str):
     # This test checks if the engine generates completions both with and
     # without optional detokenization, that detokenization includes text
@@ -21,7 +17,7 @@ def test_computed_prefix_blocks(model: str):
         "paper clips? Is there an easy to follow video tutorial available "
         "online for free?")
 
-    llm = LLM(model=model, load_format=LoadFormat.RUNAI_STREAMER)
+    llm = LLM(model=model)
     sampling_params = SamplingParams(max_tokens=10,
                                      temperature=0.0,
                                      detokenize=False)

tests/engine/test_executor.py

@@ -6,17 +6,12 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import pytest
 
-from vllm.config import LoadFormat
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.engine.llm_engine import LLMEngine
 from vllm.executor.uniproc_executor import UniProcExecutor
 from vllm.sampling_params import SamplingParams
 
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
-
-RUNAI_STREAMER_LOAD_FORMAT = LoadFormat.RUNAI_STREAMER
-
 
 class Mock:
     ...
@@ -38,12 +33,10 @@ class CustomUniExecutor(UniProcExecutor):
 CustomUniExecutorAsync = CustomUniExecutor
 
 
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_custom_executor_type_checking(model):
     with pytest.raises(ValueError):
         engine_args = EngineArgs(model=model,
-                                 load_format=RUNAI_STREAMER_LOAD_FORMAT,
                                  distributed_executor_backend=Mock)
         LLMEngine.from_engine_args(engine_args)
     with pytest.raises(ValueError):
@@ -52,8 +45,7 @@ def test_custom_executor_type_checking(model):
         AsyncLLMEngine.from_engine_args(engine_args)
 
 
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_custom_executor(model, tmp_path):
     cwd = os.path.abspath(".")
     os.chdir(tmp_path)
@@ -62,7 +54,6 @@ def test_custom_executor(model, tmp_path):
 
         engine_args = EngineArgs(
             model=model,
-            load_format=RUNAI_STREAMER_LOAD_FORMAT,
             distributed_executor_backend=CustomUniExecutor,
             enforce_eager=True,  # reduce test time
         )
@@ -77,8 +68,7 @@ def test_custom_executor(model, tmp_path):
         os.chdir(cwd)
 
 
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_custom_executor_async(model, tmp_path):
     cwd = os.path.abspath(".")
     os.chdir(tmp_path)
@@ -87,7 +77,6 @@ def test_custom_executor_async(model, tmp_path):
 
         engine_args = AsyncEngineArgs(
             model=model,
-            load_format=RUNAI_STREAMER_LOAD_FORMAT,
             distributed_executor_backend=CustomUniExecutorAsync,
             enforce_eager=True,  # reduce test time
         )
@@ -106,8 +95,7 @@ def test_custom_executor_async(model, tmp_path):
         os.chdir(cwd)
 
 
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_respect_ray(model):
     # even for TP=1 and PP=1,
     # if users specify ray, we should use ray.
@@ -116,7 +104,6 @@ def test_respect_ray(model):
     engine_args = EngineArgs(
         model=model,
         distributed_executor_backend="ray",
-        load_format=RUNAI_STREAMER_LOAD_FORMAT,
         enforce_eager=True,  # reduce test time
     )
     engine = LLMEngine.from_engine_args(engine_args)

tests/engine/test_skip_tokenizer_init.py

@@ -2,22 +2,19 @@
 
 import pytest
 
-from vllm.config import LoadFormat
 from vllm.entrypoints.llm import LLM
 from vllm.sampling_params import SamplingParams
 
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
 
-
-@pytest.mark.parametrize("model",
-                         [f"{MODEL_WEIGHTS_S3_BUCKET}/distilbert/distilgpt2"])
+@pytest.mark.parametrize("model", ["distilbert/distilgpt2"])
 def test_skip_tokenizer_initialization(model: str):
     # This test checks if the flag skip_tokenizer_init skips the initialization
     # of tokenizer and detokenizer. The generated output is expected to contain
     # token ids.
-    llm = LLM(model=model,
+    llm = LLM(
+        model=model,
         skip_tokenizer_init=True,
-              load_format=LoadFormat.RUNAI_STREAMER)
+    )
     sampling_params = SamplingParams(prompt_logprobs=True, detokenize=True)
 
     with pytest.raises(ValueError, match="cannot pass text prompts when"):

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,

tests/entrypoints/llm/test_collective_rpc.py

@@ -28,7 +28,7 @@ def test_collective_rpc(tp_size, backend):
         def echo_rank(self):
             return self.rank
 
-    llm = LLM(model="s3://vllm-ci-model-weights/Llama-3.2-1B-Instruct",
+    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct",
               enforce_eager=True,
               load_format="dummy",
               tensor_parallel_size=tp_size,

tests/entrypoints/llm/test_encode.py

@@ -6,10 +6,9 @@ from typing import List
 import pytest
 
 from vllm import LLM, PoolingParams, PoolingRequestOutput
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 
-MODEL_NAME = "s3://vllm-ci-model-weights/e5-mistral-7b-instruct"
+MODEL_NAME = "intfloat/e5-mistral-7b-instruct"
 
 PROMPTS = [
     "Hello, my name is",
@@ -33,7 +32,6 @@ def llm():
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
     llm = LLM(model=MODEL_NAME,
-              load_format=LoadFormat.RUNAI_STREAMER,
               max_num_batched_tokens=32768,
               tensor_parallel_size=1,
               gpu_memory_utilization=0.75,

tests/entrypoints/llm/test_generate.py

@@ -6,10 +6,9 @@ from typing import List
 import pytest
 
 from vllm import LLM, RequestOutput, SamplingParams
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 
-MODEL_NAME = "s3://vllm-ci-model-weights/distilgpt2"
+MODEL_NAME = "distilbert/distilgpt2"
 
 PROMPTS = [
     "Hello, my name is",
@@ -31,7 +30,6 @@ def llm():
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
     llm = LLM(model=MODEL_NAME,
-              load_format=LoadFormat.RUNAI_STREAMER,
               max_num_batched_tokens=4096,
               tensor_parallel_size=1,
               gpu_memory_utilization=0.10,

tests/entrypoints/llm/test_generate_multiple_loras.py

@@ -7,11 +7,10 @@ import pytest
 from huggingface_hub import snapshot_download
 
 from vllm import LLM
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.lora.request import LoRARequest
 
-MODEL_NAME = "s3://vllm-ci-model-weights/zephyr-7b-beta"
+MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
 
 PROMPTS = [
     "Hello, my name is",
@@ -28,7 +27,6 @@ def llm():
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
     llm = LLM(model=MODEL_NAME,
-              load_format=LoadFormat.RUNAI_STREAMER,
               tensor_parallel_size=1,
               max_model_len=8192,
               enable_lora=True,

tests/entrypoints/llm/test_guided_generate.py

@@ -7,13 +7,12 @@ import weakref
 import jsonschema
 import pytest
 
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.entrypoints.llm import LLM
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import GuidedDecodingParams, SamplingParams
 
-MODEL_NAME = "s3://vllm-ci-model-weights/Qwen2.5-1.5B-Instruct"
+MODEL_NAME = "Qwen/Qwen2.5-1.5B-Instruct"
 GUIDED_DECODING_BACKENDS = ["outlines", "lm-format-enforcer", "xgrammar"]
 
 
@@ -21,9 +20,7 @@ GUIDED_DECODING_BACKENDS = ["outlines", "lm-format-enforcer", "xgrammar"]
 def llm():
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
-    llm = LLM(model=MODEL_NAME,
-              load_format=LoadFormat.RUNAI_STREAMER,
-              max_model_len=1024)
+    llm = LLM(model=MODEL_NAME, max_model_len=1024)
 
     with llm.deprecate_legacy_api():
         yield weakref.proxy(llm)

tests/entrypoints/llm/test_lazy_outlines.py

@@ -6,7 +6,6 @@ from contextlib import nullcontext
 from vllm_test_utils import BlameResult, blame
 
 from vllm import LLM, SamplingParams
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 
 
@@ -44,8 +43,7 @@ def run_normal():
     sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
     # Create an LLM without guided decoding as a baseline.
-    llm = LLM(model="s3://vllm-ci-model-weights/distilgpt2",
-              load_format=LoadFormat.RUNAI_STREAMER,
+    llm = LLM(model="distilbert/distilgpt2",
               enforce_eager=True,
               gpu_memory_utilization=0.3)
     outputs = llm.generate(prompts, sampling_params)
@@ -61,8 +59,7 @@ def run_normal():
 
 def run_lmfe(sample_regex):
     # Create an LLM with guided decoding enabled.
-    llm = LLM(model="s3://vllm-ci-model-weights/distilgpt2",
-              load_format=LoadFormat.RUNAI_STREAMER,
+    llm = LLM(model="distilbert/distilgpt2",
               enforce_eager=True,
               guided_decoding_backend="lm-format-enforcer",
               gpu_memory_utilization=0.3)

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]})

tests/metrics/test_metrics.py

@@ -8,21 +8,17 @@ import ray
 from prometheus_client import REGISTRY
 
 from vllm import EngineArgs, LLMEngine
-from vllm.config import LoadFormat
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.engine.metrics import RayPrometheusStatLogger
 from vllm.sampling_params import SamplingParams
-
-from ..conftest import MODEL_WEIGHTS_S3_BUCKET
+from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET
 
 MODELS = [
     "distilbert/distilgpt2",
 ]
 
-RUNAI_STREAMER_LOAD_FORMAT = LoadFormat.RUNAI_STREAMER
-
 
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["float"])
@@ -146,9 +142,8 @@ def test_metric_set_tag_model_name(vllm_runner, model: str, dtype: str,
         metrics_tag_content = stat_logger.labels["model_name"]
 
     if served_model_name is None or served_model_name == []:
-        actual_model_name = f"{MODEL_WEIGHTS_S3_BUCKET}/{model}"
-        assert metrics_tag_content == actual_model_name, (
-            f"Metrics tag model_name is wrong! expect: {actual_model_name!r}\n"
+        assert metrics_tag_content == f"{MODEL_WEIGHTS_S3_BUCKET}/{model}", (
+            f"Metrics tag model_name is wrong! expect: {model!r}\n"
             f"actual: {metrics_tag_content!r}")
     else:
         assert metrics_tag_content == served_model_name[0], (
@@ -174,10 +169,11 @@ async def test_async_engine_log_metrics_regression(
     when disable_log_stats=False
     (see: https://github.com/vllm-project/vllm/pull/4150#pullrequestreview-2008176678)
     """
-    engine_args = AsyncEngineArgs(model=model,
+    engine_args = AsyncEngineArgs(
+        model=model,
         dtype=dtype,
         disable_log_stats=disable_log_stats,
-                                  load_format=RUNAI_STREAMER_LOAD_FORMAT)
+    )
     async_engine = AsyncLLMEngine.from_engine_args(engine_args)
     for i, prompt in enumerate(example_prompts):
         results = async_engine.generate(
@@ -189,7 +185,7 @@ async def test_async_engine_log_metrics_regression(
         async for _ in results:
             pass
 
-    assert_metrics(async_engine.engine, disable_log_stats,
+    assert_metrics(model, async_engine.engine, disable_log_stats,
                    len(example_prompts))
 
 
@@ -204,10 +200,11 @@ def test_engine_log_metrics_regression(
     max_tokens: int,
     disable_log_stats: bool,
 ) -> None:
-    engine_args = EngineArgs(model=model,
+    engine_args = EngineArgs(
+        model=model,
         dtype=dtype,
         disable_log_stats=disable_log_stats,
-                             load_format=RUNAI_STREAMER_LOAD_FORMAT)
+    )
     engine = LLMEngine.from_engine_args(engine_args)
     for i, prompt in enumerate(example_prompts):
         engine.add_request(
@@ -218,7 +215,8 @@ def test_engine_log_metrics_regression(
     while engine.has_unfinished_requests():
         engine.step()
 
-    assert_metrics(engine, disable_log_stats, len(example_prompts))
+    assert_metrics(f"{MODEL_WEIGHTS_S3_BUCKET}/{model}", engine,
+                   disable_log_stats, len(example_prompts))
 
 
 @pytest.mark.parametrize("model", MODELS)
@@ -285,14 +283,15 @@ def test_metric_spec_decode_interval(
 ) -> None:
     k = 5
 
-    engine_args = EngineArgs(model=model,
+    engine_args = EngineArgs(
+        model=model,
         dtype=dtype,
         disable_log_stats=False,
         gpu_memory_utilization=0.4,
         speculative_model=model,
         num_speculative_tokens=k,
         enforce_eager=True,
-                             load_format=RUNAI_STREAMER_LOAD_FORMAT)
+    )
 
     engine = LLMEngine.from_engine_args(engine_args)
 
@@ -359,7 +358,7 @@ def test_metric_spec_decode_interval(
         cleanup_dist_env_and_memory()
 
 
-def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
+def assert_metrics(model: str, engine: LLMEngine, disable_log_stats: bool,
                    num_requests: int) -> None:
     if disable_log_stats:
         with pytest.raises(AttributeError):
@@ -370,7 +369,7 @@ def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
         # Ensure the count bucket of request-level histogram metrics matches
         # the number of requests as a simple sanity check to ensure metrics are
         # generated
-        labels = {'model_name': engine.model_config.model}
+        labels = {'model_name': model}
         request_histogram_metrics = [
             "vllm:e2e_request_latency_seconds",
             "vllm:request_prompt_tokens",

tests/models/test_initialization.py

@@ -7,7 +7,6 @@ from transformers import PretrainedConfig
 
 from vllm import LLM
 
-from ..conftest import MODELS_ON_S3
 from .registry import HF_EXAMPLE_MODELS
 
 
@@ -43,11 +42,8 @@ def test_can_initialize(model_arch):
 
     with patch.object(LLM.get_engine_class(), "_initialize_kv_caches",
                       _initialize_kv_caches):
-        model_name = model_info.default
-        if model_name in MODELS_ON_S3:
-            model_name = f"s3://vllm-ci-model-weights/{model_name.split('/')[-1]}"
         LLM(
-            model_name,
+            model_info.default,
             tokenizer=model_info.tokenizer,
             tokenizer_mode=model_info.tokenizer_mode,
             speculative_model=model_info.speculative_model,

tests/mq_llm_engine/test_abort.py

@@ -10,8 +10,8 @@ import pytest
 from tests.mq_llm_engine.utils import RemoteMQLLMEngine, generate
 from vllm.engine.arg_utils import AsyncEngineArgs
 
-MODEL = "s3://vllm-ci-model-weights/gemma-1.1-2b-it"
-ENGINE_ARGS = AsyncEngineArgs(model=MODEL, load_format="runai_streamer")
+MODEL = "google/gemma-1.1-2b-it"
+ENGINE_ARGS = AsyncEngineArgs(model=MODEL)
 RAISED_ERROR = KeyError
 RAISED_VALUE = "foo"
 EXPECTED_TOKENS = 250

tests/mq_llm_engine/test_error_handling.py

@@ -21,10 +21,8 @@ from vllm.lora.request import LoRARequest
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils import FlexibleArgumentParser
 
-MODEL = "s3://vllm-ci-model-weights/gemma-1.1-2b-it"
-ENGINE_ARGS = AsyncEngineArgs(model=MODEL,
-                              load_format="runai_streamer",
-                              enforce_eager=True)
+MODEL = "google/gemma-1.1-2b-it"
+ENGINE_ARGS = AsyncEngineArgs(model=MODEL, enforce_eager=True)
 RAISED_ERROR = KeyError
 RAISED_VALUE = "foo"