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 02:22:37 +11:00
commit d87ef86991
1 changed files with 3 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

5
unsloth/models/llama.py
View file

@ -168,11 +168,12 @@ def LlamaAttention_fast_forward_inference(
Knn, Vnn = Kn, Vn
# Attention
A = torch.matmul(Qn, Knn.transpose(2, 3))
# A = torch.matmul(Qn, Knn.transpose(2, 3))
A = torch.matmul(Qn, Knn.transpose(1, 2))
A *= 1.0 / math_sqrt(self.head_dim)
A[:] = torch.nn.functional.softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)
A = torch.matmul(A, Vnn, out = Qn)
A = A.transpose(1, 2)
# A = A.transpose(1, 2)
A = A.view(self.hidden_size)
# A = self.o_proj(A)