[Core] Consolidate GB constant and enable float GB arguments (#7416)

vllm/engine/arg_utils.py

@@ -58,8 +58,8 @@ class EngineArgs:
     enable_prefix_caching: bool = False
     disable_sliding_window: bool = False
     use_v2_block_manager: bool = False
-    swap_space: int = 4  # GiB
-    cpu_offload_gb: int = 0  # GiB
+    swap_space: float = 4  # GiB
+    cpu_offload_gb: float = 0  # GiB
     gpu_memory_utilization: float = 0.90
     max_num_batched_tokens: Optional[int] = None
     max_num_seqs: int = 256
@@ -321,7 +321,7 @@ class EngineArgs:
                             default=EngineArgs.seed,
                             help='Random seed for operations.')
         parser.add_argument('--swap-space',
-                            type=int,
+                            type=float,
                             default=EngineArgs.swap_space,
                             help='CPU swap space size (GiB) per GPU.')
         parser.add_argument(