make trtllm attn backend's init_forward_metadat non blocking (#13802)

2025-11-23 22:35:40 +01:00
parent 2892265d4c
commit 18403f6bfe
1 changed files with 3 additions and 1 deletions
--- a/python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -564,7 +564,9 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
            seq_lens = forward_batch.seq_lens - forward_batch.extend_prefix_lens
            cum_seq_lens_q = torch.cat(
                (
-                    torch.tensor([0], device=forward_batch.seq_lens.device),
+                    torch.zeros(
+                        1, dtype=torch.int32, device=forward_batch.seq_lens.device
+                    ),
                    torch.cumsum(seq_lens, dim=0),
                )
            ).int()