From fda7955890978801727a388d05c14c301f4f7286 Mon Sep 17 00:00:00 2001 From: Jonah Bernard <96398205+jonahbernard@users.noreply.github.com> Date: Fri, 12 Jun 2026 16:57:46 -0700 Subject: [PATCH] [AMD][DFlash] Enable Fused KV Materialization (#27854) Co-authored-by: jonahbernard --- python/sglang/srt/speculative/dflash_worker_v2.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) 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()