[core] overhaul memory profiling and fix backward compatibility (#10511)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/engine/arg_utils.py

@@ -487,11 +487,12 @@ class EngineArgs:
             help='The fraction of GPU memory to be used for the model '
             'executor, which can range from 0 to 1. For example, a value of '
             '0.5 would imply 50%% GPU memory utilization. If unspecified, '
-            'will use the default value of 0.9. This is a global gpu memory '
-            'utilization limit, for example if 50%% of the gpu memory is '
-            'already used before vLLM starts and --gpu-memory-utilization is '
-            'set to 0.9, then only 40%% of the gpu memory will be allocated '
-            'to the model executor.')
+            'will use the default value of 0.9. This is a per-instance '
+            'limit, and only applies to the current vLLM instance.'
+            'It does not matter if you have another vLLM instance running '
+            'on the same GPU. For example, if you have two vLLM instances '
+            'running on the same GPU, you can set the GPU memory utilization '
+            'to 0.5 for each instance.')
         parser.add_argument(
             '--num-gpu-blocks-override',
             type=int,