[v1] EngineArgs for better config handling for v1 (#10382)

Signed-off-by: rickyx <rickyx@anyscale.com>

tests/v1/engine/test_async_llm.py

@@ -32,6 +32,9 @@ async def generate(engine: AsyncLLM, request_id: str,
 
 @pytest.mark.asyncio
 async def test_load(monkeypatch):
+    # TODO(rickyx): Remove monkeypatch once we have a better way to test V1
+    # so that in the future when we switch, we don't have to change all the
+    # tests.
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
 

tests/v1/engine/test_engine_args.py

@@ -0,0 +1,42 @@
+import pytest
+
+from vllm import envs
+from vllm.config import VllmConfig
+from vllm.engine.arg_utils import EngineArgs
+from vllm.usage.usage_lib import UsageContext
+
+if not envs.VLLM_USE_V1:
+    pytest.skip(
+        "Skipping V1 tests. Rerun with `VLLM_USE_V1=1` to test.",
+        allow_module_level=True,
+    )
+
+
+def test_defaults():
+    engine_args = EngineArgs(model="facebook/opt-125m")
+
+    # Assert V1 defaults
+    assert (engine_args.enable_prefix_caching
+            ), "V1 turns on prefix caching by default"
+
+
+def test_defaults_with_usage_context():
+    engine_args = EngineArgs(model="facebook/opt-125m")
+    vllm_config: VllmConfig = engine_args.create_engine_config(
+        UsageContext.LLM_CLASS)
+
+    assert vllm_config.scheduler_config.max_num_seqs == 1024
+    assert vllm_config.scheduler_config.max_num_batched_tokens == 8192
+
+    engine_args = EngineArgs(model="facebook/opt-125m")
+    vllm_config = engine_args.create_engine_config(
+        UsageContext.OPENAI_API_SERVER)
+    assert vllm_config.scheduler_config.max_num_seqs == 1024
+    assert vllm_config.scheduler_config.max_num_batched_tokens == 2048
+
+
+def test_prefix_cache_disabled_with_multimodel():
+    engine_args = EngineArgs(model="llava-hf/llava-1.5-7b-hf")
+
+    vllm_config = engine_args.create_engine_config(UsageContext.LLM_CLASS)
+    assert not vllm_config.cache_config.enable_prefix_caching

tests/v1/engine/test_engine_core.py

@@ -43,7 +43,8 @@ def test_engine_core(monkeypatch):
         m.setenv("VLLM_USE_V1", "1")
         """Setup the EngineCore."""
         engine_args = EngineArgs(model=MODEL_NAME)
-        vllm_config = engine_args.create_engine_config()
+        vllm_config = engine_args.create_engine_config(
+            usage_context=UsageContext.UNKNOWN_CONTEXT)
         executor_class = AsyncLLM._get_executor_cls(vllm_config)
 
         engine_core = EngineCore(vllm_config=vllm_config,

tests/v1/engine/test_engine_core_client.py

@@ -82,7 +82,8 @@ def test_engine_core_client(monkeypatch, multiprocessing_mode: bool):
         m.setenv("VLLM_USE_V1", "1")
 
         engine_args = EngineArgs(model=MODEL_NAME, compilation_config=3)
-        vllm_config = engine_args.create_engine_config()
+        vllm_config = engine_args.create_engine_config(
+            UsageContext.UNKNOWN_CONTEXT)
         executor_class = AsyncLLM._get_executor_cls(vllm_config)
         client = EngineCoreClient.make_client(
             vllm_config,
@@ -153,7 +154,8 @@ async def test_engine_core_client_asyncio(monkeypatch):
         m.setenv("VLLM_USE_V1", "1")
 
         engine_args = EngineArgs(model=MODEL_NAME)
-        vllm_config = engine_args.create_engine_config()
+        vllm_config = engine_args.create_engine_config(
+            usage_context=UsageContext.UNKNOWN_CONTEXT)
         executor_class = AsyncLLM._get_executor_cls(vllm_config)
         client = EngineCoreClient.make_client(
             vllm_config,