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 2024-07-25 00:19:32 -07:00
commit ef42e61d68
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

@ -96,7 +96,7 @@ def fix_prepare_inputs_for_generation(module):
pass
pass
torch_matmul = torch.matmul
def LlamaAttention_fast_forward_inference(
self,
hidden_states: torch.Tensor,
@ -238,10 +238,10 @@ 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
A = torch.matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,: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)
A = torch.matmul(A, Vnn, out = Qn)
A = torch_matmul(A, Vnn, out = Qn)
else:
A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = False)
pass