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

[NPU] enable mla prepare fused kernel only when being mla attn (#22024)

Browse Source
This commit is contained in:
khalilzhk
2026-04-08 00:49:16 +08:00
committed by GitHub
parent be42fbbbd7
commit 6131fb5882
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+2 -2
View File
@@ -860,7 +860,7 @@ class AscendAttnBackend(AttentionBackend):
sinks: Optional[torch.Tensor] = None, sinks: Optional[torch.Tensor] = None,
slopes: Optional[torch.Tensor] = None, slopes: Optional[torch.Tensor] = None,
): ):
if is_mla_preprocess_enabled(): if is_mla_preprocess_enabled() and self.use_mla:
# MLAPO and MLAPROLOG do save kv_cache # MLAPO and MLAPROLOG do save kv_cache
save_kv_cache = False save_kv_cache = False
if self.is_dllm_model: if self.is_dllm_model:
@@ -1773,7 +1773,7 @@ class AscendAttnBackend(AttentionBackend):
sinks: Optional[torch.Tensor] = None, sinks: Optional[torch.Tensor] = None,
slopes: Optional[torch.Tensor] = None, slopes: Optional[torch.Tensor] = None,
): ):
if is_mla_preprocess_enabled(): if is_mla_preprocess_enabled() and self.use_mla:
# MLAPO does saving kv_cache # MLAPO does saving kv_cache
save_kv_cache = False save_kv_cache = False
if topk_indices is not None: if topk_indices is not None: