EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Bug fix] Fix ascend mla in aclgraph (#9925)

Browse Source
This commit is contained in:
alanhe151220037
2025-09-09 00:49:43 +08:00
committed by GitHub
parent 0096798ed6
commit 8085aca791
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/attention/ascend_backend.py
View File

@@ -368,7 +368,7 @@ class AscendAttnBackend(AttentionBackend):
-1, layer.tp_v_head_num, self.page_size, self.kv_lora_rank -1, layer.tp_v_head_num, self.page_size, self.kv_lora_rank
) )
q_nope = q.view(-1, layer.tp_q_head_num, 1, self.kv_lora_rank) q_nope = q.view(-1, layer.tp_q_head_num, 1, self.kv_lora_rank).contiguous()
q_rope = q_rope.view(-1, layer.tp_q_head_num, 1, self.qk_rope_head_dim) q_rope = q_rope.view(-1, layer.tp_q_head_num, 1, self.qk_rope_head_dim)
if self.forward_metadata.seq_lens_cpu_int is None: if self.forward_metadata.seq_lens_cpu_int is None:
actual_seq_len_kv = self.forward_metadata.seq_lens_cpu_list actual_seq_len_kv = self.forward_metadata.seq_lens_cpu_list