[core][model] yet another cpu offload implementation (#6496)

Co-authored-by: Michael Goin <michael@neuralmagic.com>

vllm/config.py

@@ -433,6 +433,7 @@ class CacheConfig:
         num_gpu_blocks_override: Optional[int] = None,
         sliding_window: Optional[int] = None,
         enable_prefix_caching: bool = False,
+        cpu_offload_gb: float = 0,
     ) -> None:
         self.block_size = block_size
         self.gpu_memory_utilization = gpu_memory_utilization
@@ -441,6 +442,7 @@ class CacheConfig:
         self.cache_dtype = cache_dtype
         self.sliding_window = sliding_window
         self.enable_prefix_caching = enable_prefix_caching
+        self.cpu_offload_gb = cpu_offload_gb
         self._verify_args()
         self._verify_cache_dtype()
         self._verify_prefix_caching()