[torch.compile] rework compile control with piecewise cudagraph (#9715)

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

vllm/platforms/tpu.py

@@ -11,7 +11,7 @@ from .interface import Platform, PlatformEnum
 if "VLLM_TORCH_COMPILE_LEVEL" not in os.environ:
     os.environ["VLLM_TORCH_COMPILE_LEVEL"] = str(CompilationLevel.DYNAMO_ONCE)
 
-assert envs.VLLM_TORCH_COMPILE_LEVEL < CompilationLevel.INDUCTOR,\
+assert envs.VLLM_TORCH_COMPILE_LEVEL < CompilationLevel.PIECEWISE,\
      "TPU does not support Inductor."
 
 set_torch_compile_backend("openxla")