Fix bug in dbrx model (#21445)
Co-authored-by: Jianzhao Xu <xujianchao@huawei.com>
This commit is contained in:
co-authored by
Jianzhao Xu
parent
646573e4e8
commit
3867c6431a
@@ -211,7 +211,7 @@ class DbrxAttention(nn.Module):
|
|||||||
self.head_dim = self.d_model // self.total_num_heads
|
self.head_dim = self.d_model // self.total_num_heads
|
||||||
self.total_num_kv_heads = config.attn_config.kv_n_heads
|
self.total_num_kv_heads = config.attn_config.kv_n_heads
|
||||||
self.clip_qkv = config.attn_config.clip_qkv
|
self.clip_qkv = config.attn_config.clip_qkv
|
||||||
self.rope_theta = config.attn_config.rope_parameters["rope_theta"]
|
self.rope_theta = config.attn_config.rope_theta
|
||||||
self.max_position = config.max_seq_len
|
self.max_position = config.max_seq_len
|
||||||
|
|
||||||
# pylint: disable=invalid-name
|
# pylint: disable=invalid-name
|
||||||
|
|||||||
Reference in New Issue
Block a user