[Feat] QWen-1M context support[2/2]: Update block sparse attention backend (#5949)

2025-08-07 14:49:36 +08:00
parent a69b637014
commit b7cd743038
15 changed files with 2121 additions and 4 deletions
--- a/python/sglang/srt/disaggregation/decode_schedule_batch_mixin.py
+++ b/python/sglang/srt/disaggregation/decode_schedule_batch_mixin.py
@@ -76,6 +76,9 @@ class ScheduleBatchDisaggregationDecodeMixin:
            req_pool_indices, dtype=torch.int64, device=self.device
        )
        self.seq_lens = torch.tensor(seq_lens, dtype=torch.int64, device=self.device)
+        self.orig_seq_lens = torch.tensor(
+            seq_lens, dtype=torch.int32, device=self.device
+        )
        self.out_cache_loc = out_cache_loc
        self.seq_lens_sum = sum(seq_lens)