[UX][Startup] Account for CUDA graphs during memory profiling (#30515)

vllm/envs.py

@@ -244,6 +244,7 @@ if TYPE_CHECKING:
     VLLM_CUDA_COMPATIBILITY_PATH: str | None = None
     VLLM_ELASTIC_EP_SCALE_UP_LAUNCH: bool = False
     VLLM_ELASTIC_EP_DRAIN_REQUESTS: bool = False
+    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: bool = False
 
 
 def get_default_cache_root():
@@ -1628,6 +1629,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_ELASTIC_EP_DRAIN_REQUESTS": lambda: bool(
         int(os.getenv("VLLM_ELASTIC_EP_DRAIN_REQUESTS", "0"))
     ),
+    # If set to 1, enable CUDA graph memory estimation during memory profiling.
+    # This profiles CUDA graph memory usage to provide more accurate KV cache
+    # memory allocation. Disabled by default to preserve existing behavior.
+    "VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": lambda: bool(
+        int(os.getenv("VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS", "0"))
+    ),
 }