From 7e226b1b71030c053381b2a9f6ff8a82845e7eac Mon Sep 17 00:00:00 2001 From: xiaofei-zheng Date: Wed, 14 Oct 2026 11:06:07 +0800 Subject: [PATCH] [AMD] Use the preallocated EAGLE topk=1 draft chain on HIP (#43640) Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> --- python/sglang/srt/speculative/eagle_worker_v2.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/speculative/eagle_worker_v2.py b/python/sglang/srt/speculative/eagle_worker_v2.py index 2ddca4bc97..a91fca4691 100644 --- a/python/sglang/srt/speculative/eagle_worker_v2.py +++ b/python/sglang/srt/speculative/eagle_worker_v2.py @@ -735,13 +735,13 @@ class EagleDraftWorker(EagleDraftWorkerBase): self.topk == 1 and topk_index.shape[0] <= self._topk1_parents_prealloc.shape[0] ) - # Materialize the chain directly only when the CUDA kernel can write + # Materialize the chain directly only when the Triton kernel can write # every subsequent column. Other topk=1 paths retain the token list and # assemble it with one final cat instead of launching a copy per step. draft_tokens_topk1 = None if ( topk1_chain_fits - and _is_cuda + and (_is_cuda or _is_hip) and self.hot_token_id is None and not needs_draft_probs ):