[AMD] DSv4: bring HIP compress-state pool into the memory_saver KV_CACHE region (#31747)

Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>

python/sglang/srt/mem_cache/deepseek_v4_compress_state.py

@@ -153,11 +153,6 @@ class CompressStatePool:
         allocation boilerplate. Requires ``self._size`` and ``self.last_dim``
         to be set already.
         """
-        if _is_hip:
-            self.kv_score_buffer = KVAndScore(
-                torch.empty((self._size, self.last_dim), dtype=dtype, device=device)
-            )
-        else:
         self.memory_saver_adapter = TorchMemorySaverAdapter.create(
             enable=enable_memory_saver
         )

python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py

@@ -646,9 +646,6 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool):
 
         self._init_compressed_layer_mapping()
 
-        if _is_hip:
-            self._init_paged_compress_states(False)
-        else:
         self._init_paged_compress_states(enable_memory_saver)
 
     def get_unified_kv(self, layer_id: int) -> torch.Tensor: