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 2025-02-20 07:36:23 -08:00
commit bb0d3f900f
1 changed files with 2 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

7
unsloth/models/llama.py
View file

@ -247,7 +247,7 @@ def LlamaAttention_fast_forward_inference(
# Grouped query attention
_, _, cached_len, _ = Knn.shape
if bsz == 1 or not SDPA_HAS_GQA and n_groups != 1:
if True:
Knn = Knn[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, cached_len, head_dim)
Vnn = Vnn[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, cached_len, head_dim)
Knn = Knn.reshape(bsz, n_heads, cached_len, head_dim)
@ -266,10 +266,7 @@ def LlamaAttention_fast_forward_inference(
A[:] = torch_nn_functional_softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)
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)
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 = False)
pass
A = A.transpose(1, 2)
A = A.reshape(bsz, 1, attention_size)