ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update llama.py

Browse source
This commit is contained in:
Daniel Han-Chen 2024-01-23 02:13:01 +11:00
commit 5260ec2a0a
1 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
unsloth/models/llama.py
View file

@ -159,10 +159,11 @@ def LlamaAttention_fast_forward_inference(
# Vnn = Vn[:, :, None, :, :].expand(1, n_kv_heads, n_groups, cached_len, head_dim)
# Knn = Knn.reshape(1, n_heads, cached_len, head_dim)
# Vnn = Vnn.reshape(1, n_heads, cached_len, head_dim)
Knn = Kn[:, :, None, :, :].expand(n_kv_heads, n_groups, kv_seq_len, head_dim)
Vnn = Vn[:, :, None, :, :].expand(n_kv_heads, n_groups, kv_seq_len, head_dim)
Knn = Knn.reshape(n_heads, kv_seq_len, head_dim)
Vnn = Vnn.reshape(n_heads, kv_seq_len, head_dim)
new_seq_len = seq_len + 1
Knn = Kn[:, :, None, :].expand(n_kv_heads, n_groups, new_seq_len, head_dim)
Vnn = Vn[:, :, None, :].expand(n_kv_heads, n_groups, new_seq_len, head_dim)
Knn = Knn.reshape(n_heads, new_seq_len, head_dim)
Vnn = Vnn.reshape(n_heads, new_seq_len, head_dim)
else:
Knn, Vnn = Kn, Vn