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

[NPU]Ensure tensors allocated by empty_like are contiguous (#34935)

Browse Source
This commit is contained in:
Estrella-xx
2026-08-20 20:40:57 +08:00
committed by GitHub
parent a4ffb996db
commit c98f1ccedb
1 changed files with 6 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
+6 -2
View File
@@ -1587,7 +1587,9 @@ class AscendAttnBackend(AttentionBackend):
(q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
)
else:
attn_output = torch.empty_like(q)
attn_output = torch.empty_like(
q, memory_format=torch.contiguous_format
)
use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
@@ -2746,7 +2748,9 @@ class AscendAttnBackend(AttentionBackend):
(q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
)
else:
attn_output = torch.empty_like(q)
attn_output = torch.empty_like(
q, memory_format=torch.contiguous_format
)
use_gqa = layer.tp_q_head_num != layer.tp_k_head_num