[Core] Run garbage collector after CUDA graph capture to fix throughput regression (vllm-project#24128)

Signed-off-by: Gregory Shtrasberg <[email protected]>
Co-authored-by: Gregory Shtrasberg <[email protected]>

Author: 2 people

vllm/v1/worker/gpu_model_runner.py

@@ -2885,6 +2885,7 @@ def freeze_gc():
             finally:
                 if should_freeze:
                     gc.unfreeze()
+                    gc.collect()
 
         # Trigger CUDA graph capture for specific shapes.
         # Capture the large shapes first so that the smaller shapes