[Bugfix] Re-enable use_cudagraph in vLLM v1 (#19299)

Signed-off-by: Richard Zou <zou3519@gmail.com>

vllm/config.py

@@ -3918,12 +3918,14 @@ class CompilationConfig:
     constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`."""
 
     # CudaGraph compilation
-    use_cudagraph: bool = False
+    use_cudagraph: bool = envs.VLLM_USE_V1
     """Whether to use cudagraph inside compilation.
     - False: cudagraph inside compilation is not used.
     - True: cudagraph inside compilation is used. It requires
         that all input buffers have fixed addresses, and all
         splitting ops write their outputs to input buffers.
+    In the vLLM V1 Engine, this flag only applies for
+    CompilationLevel.PIECEWISE (aka -O3).
     Note that this is orthogonal to the cudagraph capture logic
     outside of compilation.
     TODO: move outside cudagraph logic into compilation.
@@ -4425,7 +4427,6 @@ class VllmConfig:
             # FIXME(rob): Add function to set all of these.
             if not self.compilation_config.custom_ops:
                 self.compilation_config.custom_ops = ["none"]
-            self.compilation_config.use_cudagraph = True
             self.compilation_config.cudagraph_num_of_warmups = 1
             self.compilation_config.pass_config.enable_fusion = False
             self.compilation_config.pass_config.enable_noop = False