ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix for casual mask (#3011)

Browse source
This commit is contained in:
Lei Zhenyuan 2025-07-22 18:27:57 +08:00 • committed by GitHub
commit 37d0dedde5
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

@ -342,9 +342,9 @@ def LlamaAttention_fast_forward_inference(
A = torch_matmul(A, Vnn, out = Qn)
else:
if SDPA_HAS_GQA:
A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = False, enable_gqa = True)
A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = is_causal, enable_gqa = True)
else:
A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = False)
A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = is_causal)
pass
A = A.transpose(1, 2)
A = A.reshape(bsz, 1, attention_size)