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 2024-02-04 03:42:41 +11:00
commit 63816fc119
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
unsloth/models/llama.py
View file

@ -321,8 +321,8 @@ def LlamaAttention_fast_forward(
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(Q.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(Q.dtype)
attn_weights = torch.nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
A = torch.matmul(attn_weights, V)
# Must be contiguous or else results are False!