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

2026-02-20 23:20:29 +08:00
parent f23a23cc05
commit 52852404c8
1 changed files with 1 additions and 1 deletions
--- a/python/sglang/srt/layers/attention/flashinfer_backend.py
+++ b/python/sglang/srt/layers/attention/flashinfer_backend.py
@@ -812,7 +812,7 @@ class FlashInferAttnBackend(AttentionBackend):
                or layer.attn_type == AttentionType.ENCODER_ONLY
            ):
                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

            if self.forward_metadata.extend_no_prefix: