minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[FIX] NSA backend page_table overflow in speculative decoding target_verify (#19016)

Browse Source
This commit is contained in:
Xinyuan Tong
2026-03-05 16:04:58 -08:00
committed by GitHub
parent 13af7cbb02
commit 9ebffef1ef
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/nsa_backend.py
+3 -1
View File
@@ -752,9 +752,11 @@ class NativeSparseAttnBackend(
max_bs + 1, dtype=torch.int32, device=self.device max_bs + 1, dtype=torch.int32, device=self.device
), ),
# fake page_table for sparse_prefill # fake page_table for sparse_prefill
# Add extra columns for speculative draft tokens to avoid
# overflow during target_verify when max_seqlen_k = seq_len + num_draft_tokens
"page_table": torch.zeros( "page_table": torch.zeros(
max_num_tokens, max_num_tokens,
self.max_context_len, self.max_context_len + (self.speculative_num_draft_tokens or 0),
dtype=torch.int32, dtype=torch.int32,
device=self.device, device=self.device,
), ),