diff --git a/vllm/compilation/backends.py b/vllm/compilation/backends.py
index 1206424ae1e3f..f002a8ff905b1 100644
--- a/vllm/compilation/backends.py
+++ b/vllm/compilation/backends.py
@@ -265,7 +265,13 @@ def configure_post_pass(self):
 
     def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:
 
+        # when dynamo calls the backend, it means the bytecode
+        # transform and analysis are done
         compilation_counter.num_graphs_seen += 1
+        from .monitor import torch_compile_start_time
+        dynamo_time = time.time() - torch_compile_start_time
+        logger.info("Dynamo bytecode transform time: %.2f s", dynamo_time)
+        self.compilation_configs.compilation_time += dynamo_time
 
         # we control the compilation process, each instance can only be
         # called once
diff --git a/vllm/compilation/decorators.py b/vllm/compilation/decorators.py
index a32dced57e5b3..938430fe2a501 100644
--- a/vllm/compilation/decorators.py
+++ b/vllm/compilation/decorators.py
@@ -145,6 +145,7 @@ def _support_torch_compile(
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = '', **kwargs):
         old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
+        self.vllm_config = vllm_config
         # for CompilationLevel.DYNAMO_AS_IS , the upper level model runner
         # will handle the compilation, so we don't need to do anything here.
         self.do_not_compile = \
@@ -157,9 +158,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = '', **kwargs):
         TorchCompileWrapperWithCustomDispatcher.__init__(
             self, compilation_level=vllm_config.compilation_config.level)
 
-        if vllm_config.compilation_config.level == CompilationLevel.PIECEWISE:
-            start_monitoring_torch_compile(vllm_config.compilation_config)
-
     cls.__init__ = __init__
 
     def __call__(self, *args, **kwargs):
@@ -186,6 +184,8 @@ def __call__(self, *args, **kwargs):
                         raise ValueError(
                             "Unsupported dynamic dimensions"
                             f" {dims} for argument {k} with type {type(arg)}.")
+            # here, it is the starting point of the `torch.compile` process
+            start_monitoring_torch_compile(self.vllm_config.compilation_config)
 
         # if we don't use custom dispatcher, we can directly call the
         # compiled function and let torch.compile handle the dispatching,
diff --git a/vllm/compilation/monitor.py b/vllm/compilation/monitor.py
index f718e46423212..3348674b09af2 100644
--- a/vllm/compilation/monitor.py
+++ b/vllm/compilation/monitor.py
@@ -1,14 +1,19 @@
+import time
+
 from vllm.config import CompilationConfig, CompilationLevel
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)
 
+torch_compile_start_time: float = 0.0
+
 
 def start_monitoring_torch_compile(compilation_config: CompilationConfig):
-    pass
+    global torch_compile_start_time
+    torch_compile_start_time = time.time()
 
 
 def end_monitoring_torch_compile(compilation_config: CompilationConfig):
     if compilation_config.level == CompilationLevel.PIECEWISE:
-        logger.info("graph compilation takes %.2f s in total",
+        logger.info("torch.compile takes %.2f s in total",
                     compilation_config.compilation_time)