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

[NPU] [BugFix] Fix discontinuous input for FIA operator in GLM4.7‑Flash (#36170)

Browse Source
This commit is contained in:
silencejade
2026-08-29 00:03:05 -07:00
committed by GitHub
parent 46334460a6
commit 78eef34356
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+1 -1
View File
@@ -1815,7 +1815,7 @@ class AscendAttnBackend(AttentionBackend):
torch.ops.npu.npu_fused_infer_attention_score( torch.ops.npu.npu_fused_infer_attention_score(
q[None, q_len_offset : q_len_offset + q_len], q[None, q_len_offset : q_len_offset + q_len],
k[None, q_len_offset : q_len_offset + q_len], k[None, q_len_offset : q_len_offset + q_len],
v[None, q_len_offset : q_len_offset + q_len], v[None, q_len_offset : q_len_offset + q_len].contiguous(),
num_heads=layer.tp_q_head_num, num_heads=layer.tp_q_head_num,
num_key_value_heads=layer.tp_k_head_num, num_key_value_heads=layer.tp_k_head_num,
input_layout="BSND", # todo, TND not supports q_heads!=k_heads input_layout="BSND", # todo, TND not supports q_heads!=k_heads