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 2025-02-02 14:51:14 -08:00
commit 693ae8ddfc
1 changed files with 1 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

1
unsloth/models/llama.py
View file

@ -260,6 +260,7 @@ def LlamaAttention_fast_forward_inference(
if bsz == 1:
Qn *= self.scalar # See https://github.com/ggerganov/llama.cpp/issues/7805#issuecomment-2153349963
# It seems like doing (Q * scalar) @ K is better than (Q @ K) * scalar to stop overflows
print(Qn.shape, Knn.transpose(2, 3).shape, self.attention[:,:,:,:cached_len].shape, self.attention.shape, cached_len)
A = torch_matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,:cached_len])
# if attention_mask is not None: A += attention_mask # Must add attention_mask for batched
A[:] = torch_nn_functional_softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)