[Core] Freeze gc during cuda graph capture to speed up init (#21146)

Signed-off-by: Codex <codex@openai.com>
Signed-off-by: mgoin <mgoin64@gmail.com>

vllm/envs.py

@@ -140,6 +140,7 @@ if TYPE_CHECKING:
     VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: Optional[int] = None
     VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 120
     VLLM_USE_CUDNN_PREFILL: bool = False
+    VLLM_ENABLE_CUDAGRAPH_GC: bool = False
     VLLM_LOOPBACK_IP: str = ""
 
 
@@ -968,6 +969,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_TRTLLM_DECODE_ATTENTION":
     lambda: os.getenv("VLLM_USE_TRTLLM_DECODE_ATTENTION", None),
 
+    # Controls garbage collection during CUDA graph capture.
+    # If set to 0 (default), enables GC freezing to speed up capture time.
+    # If set to 1, allows GC to run during capture.
+    "VLLM_ENABLE_CUDAGRAPH_GC":
+    lambda: bool(int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))),
+
     # Used to force set up loopback IP
     "VLLM_LOOPBACK_IP":
     lambda: os.getenv("VLLM_LOOPBACK_IP", ""),