[ROCm] Using a more precise memory profiling (#12624)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>

vllm/platforms/rocm.py

@@ -169,4 +169,5 @@ class RocmPlatform(Platform):
                                  device: Optional[torch.types.Device] = None
                                  ) -> float:
         torch.cuda.reset_peak_memory_stats(device)
-        return torch.cuda.max_memory_allocated(device)
+        return torch.cuda.mem_get_info(device)[1] - torch.cuda.mem_get_info(
+            device)[0]