diff --git a/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py b/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py index 93cc9dfbcf..cde274de13 100644 --- a/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py +++ b/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py @@ -88,6 +88,7 @@ from sglang.srt.utils import ( is_npu, is_pin_memory_available, ) +from sglang.srt.utils.common import async_h2d if TYPE_CHECKING: from sglang.srt.model_executor.model_runner import ModelRunner @@ -720,13 +721,16 @@ class DSparkWorkerV2(BaseSpecWorker): # A newly placed request's ring holds nothing below its cached prefix. placed = [i for i, reset in enumerate(batch.encoder_swa_reset) if reset] if placed: + # Pinned copies and a device gather: a list index or a pageable + # copy here would wait for the whole target forward. + placed_rows = async_h2d(placed, dtype=torch.int64, device=device) + valid_from = async_h2d( + [batch.prefix_lens[i] for i in placed], + dtype=torch.int64, + device=device, + ) draft_pool.set_swa_valid_from( - batch.req_pool_indices[placed], - torch.tensor( - [batch.prefix_lens[i] for i in placed], - dtype=torch.int64, - device=device, - ), + batch.req_pool_indices.index_select(0, placed_rows), valid_from ) # The draft KV goes to the slots the target prefill just wrote. cache_loc = self._kv_injector.ids_for(batch_output.kv_loc_plan)