chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[hybrid-model] clean up and consolidate redundant fields in RadixLinearAttention (#17660)

Browse Source
This commit is contained in:
Minglei Zhu
2026-01-27 10:37:58 -08:00
committed by GitHub
parent 8acd4d7d7e
commit d90c0837e5
4 changed files with 54 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/kimi_linear.py
+4 -3
View File
@@ -315,11 +315,12 @@ class KimiDeltaAttention(nn.Module):
self.linear_attn = RadixLinearAttention(
layer_id=self.layer_idx,
num_qk_heads=self.num_k_heads // self.attn_tp_size,
num_q_heads=self.num_k_heads // self.attn_tp_size,
num_k_heads=self.num_k_heads // self.attn_tp_size,
num_v_heads=self.num_v_heads // self.attn_tp_size,
head_qk_dim=self.head_k_dim,
head_q_dim=self.head_k_dim,
head_k_dim=self.head_k_dim,
head_v_dim=self.head_v_dim,
attention_tp_size=self.attn_tp_size,
conv_weights=conv_weights,
bias=bias,
A_log=self.A_log,