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 02:03:25 +11:00
commit 00f50876f7
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/models/llama.py
View file

@ -112,9 +112,8 @@ def LlamaAttention_fast_forward_inference(
Xn = hidden_states.view(self.hidden_size)
K1, V1 = past_key_value
seq_len = K1.shape[1]
position_id = seq_len + 1
K1 = K1.view(n_kv_heads, position_id, head_dim)
V1 = V1.view(n_kv_heads, position_id, head_dim)
K1 = K1.view(n_kv_heads, seq_len, head_dim)
V1 = V1.view(n_kv_heads, seq_len, head_dim)
# LoRA or general matrix multiplication
dtype = Xn.dtype
@ -135,6 +134,7 @@ def LlamaAttention_fast_forward_inference(
# kv_seq_len = K1.shape[-2] + 1
# cos, sin = self.rotary_emb(Vn, seq_len = kv_seq_len)
# Qn, Kn = inplace_rope_embedding(Qn, Kn, cos, sin, position_ids)
position_id = seq_len + 1
cos = self.rotary_emb.cos_cached[position_id]
sin = self.rotary_emb.sin_cached[position_id]
h = head_dim // 2