diff --git a/python/sglang/srt/speculative/dflash_worker_v2.py b/python/sglang/srt/speculative/dflash_worker_v2.py
index 1184fb2d1..3e91ad5ef 100644
--- a/python/sglang/srt/speculative/dflash_worker_v2.py
+++ b/python/sglang/srt/speculative/dflash_worker_v2.py
@@ -238,7 +238,7 @@ class DFlashWorkerV2(BaseSpecWorker):
         self._draft_greedy_rank_index_buf: Optional[torch.Tensor] = None
         self._draft_greedy_selected_ids_buf: Optional[torch.Tensor] = None
         self._draft_greedy_index_cap: int = 0
-        self._use_fused_kv_materialize = is_cuda()
+        self._use_fused_kv_materialize = is_cuda() or is_hip()
         self._fused_kv_helper: Optional[object] = None
         if self._use_fused_kv_materialize:
             self._init_fused_kv_helper()