Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- --- a/sglang/srt/models/deepseek_v2.py
- +++ b/sglang/srt/models/deepseek_v2.py
- @@ -1201,6 +1201,10 @@
- self.next_skip_topk = False
- print('layer_id {} DSA skip_topk {} next_skip_topk {} is_nextn {}'.format(layer_id, self.skip_topk, self.next_skip_topk, is_nextn))
- + else:
- + self.skip_topk = False
- + self.next_skip_topk = False
- +
- self.kv_b_proj = ColumnParallelLinear(
- self.kv_lora_rank,
- self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
Advertisement
Add Comment
Please, Sign In to add comment