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

[CPU] Fix truncated KV prefix in intel_amx spec verify (#36430)

Browse Source
This commit is contained in:
Haotong Zou
2026-08-27 09:42:36 +08:00
committed by GitHub
parent 15688fea7d
commit 058f99c561
1 changed files with 0 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/intel_amx_backend.py
-2
View File
@@ -215,7 +215,6 @@ class IntelAMXAttnBackend(AttentionBackend):
seq_lens, extend_seq_lens, extend_start_loc, tree_mask = self.extend_metadata seq_lens, extend_seq_lens, extend_start_loc, tree_mask = self.extend_metadata
_, max_extend_len = self.forward_metadata _, max_extend_len = self.forward_metadata
seq_lens = forward_batch.seq_lens
if seq_lens.dtype != torch.int64: if seq_lens.dtype != torch.int64:
seq_lens = seq_lens.to(torch.int64) seq_lens = seq_lens.to(torch.int64)
@@ -262,7 +261,6 @@ class IntelAMXAttnBackend(AttentionBackend):
seq_lens = forward_batch.seq_lens seq_lens = forward_batch.seq_lens
q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim) q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
seq_lens = forward_batch.seq_lens
if seq_lens.dtype != torch.int64: if seq_lens.dtype != torch.int64:
seq_lens = seq_lens.to(torch.int64) seq_lens = seq_lens.to(torch.int64)