[misc] add torch.compile compatibility check (#10618)

Signed-off-by: youkaichao <youkaichao@gmail.com>

tests/v1/engine/test_engine_core_client.py

@@ -81,7 +81,7 @@ def test_engine_core_client(monkeypatch, multiprocessing_mode: bool):
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
 
-        engine_args = EngineArgs(model=MODEL_NAME)
+        engine_args = EngineArgs(model=MODEL_NAME, compilation_config=3)
         vllm_config = engine_args.create_engine_config()
         executor_class = AsyncLLM._get_executor_cls(vllm_config)
         client = EngineCoreClient.make_client(

vllm/config.py

@@ -2394,6 +2394,20 @@ class VllmConfig:
             self.compilation_config.pass_config.enable_reshape = False
             self.compilation_config.level = CompilationLevel.PIECEWISE
 
+        if self.cache_config is not None and \
+            self.cache_config.cpu_offload_gb > 0 and \
+            self.compilation_config.level != CompilationLevel.NO_COMPILATION:
+            logger.warning(
+                "CPU offload is not supported with `torch.compile` yet."
+                " Disabling `torch.compile`.")
+            self.compilation_config.level = CompilationLevel.NO_COMPILATION
+
+        if self.lora_config is not None and self.compilation_config.level !=\
+             CompilationLevel.NO_COMPILATION:
+            logger.warning("LoRA is not supported with `torch.compile` yet. "
+                           "Disabling `torch.compile`.")
+            self.compilation_config.level = CompilationLevel.NO_COMPILATION
+
         current_platform.check_and_update_config(self)
 
     def __str__(self):

vllm/engine/arg_utils.py

@@ -197,6 +197,13 @@ class EngineArgs:
         if not self.tokenizer:
             self.tokenizer = self.model
 
+        # support `EngineArgs(compilation_config={...})`
+        # without having to manually construct a
+        # CompilationConfig object
+        if isinstance(self.compilation_config, (int, dict)):
+            self.compilation_config = CompilationConfig.from_cli(
+                json.dumps(self.compilation_config))
+
         # Setup plugins
         from vllm.plugins import load_general_plugins
         load_general_plugins()