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

[Fix] DO NOT skip save_kv_cache for dllm (#19020)

Browse Source
This commit is contained in:
DarkSharpness
2026-02-20 23:20:29 +08:00
committed by GitHub
parent f23a23cc05
commit 52852404c8
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/flashinfer_backend.py
+1 -1
View File
@@ -812,7 +812,7 @@ class FlashInferAttnBackend(AttentionBackend):
or layer.attn_type == AttentionType.ENCODER_ONLY or layer.attn_type == AttentionType.ENCODER_ONLY
): ):
causal = False causal = False
if save_kv_cache and layer.attn_type == AttentionType.ENCODER_ONLY: if not self.is_dllm_model and layer.attn_type == AttentionType.ENCODER_ONLY:
save_kv_cache = False save_kv_cache = False
if self.forward_metadata.extend_no_prefix: if self.forward_metadata.extend_no_prefix: