Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

Dynamic atten don't repeat the last shape of atten masks

Browse Source
This commit is contained in:
Disty0
2024-10-25 13:45:20 +03:00
parent 42291a7e48
commit 7d56d5b72c
2 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/intel/ipex/attention.py
+2 -2
View File
@@ -136,11 +136,11 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
if do_split:
batch_size_attention, query_tokens, shape_three = query.shape[0], query.shape[1], query.shape[2]
hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
if attn_mask is not None and attn_mask.shape != query.shape:
if attn_mask is not None and attn_mask.shape[:-1] != query.shape[:-1]:
if len(query.shape) == 4:
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], shape_three // attn_mask.shape[2], 1))
else:
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], shape_three // attn_mask.shape[2]))
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], 1))
for i in range(batch_size_attention // split_slice_size):
start_idx = i * split_slice_size
end_idx = (i + 1) * split_slice_size
modules/sd_hijack_dynamic_atten.py
+2 -2
View File
@@ -57,11 +57,11 @@ def sliced_scaled_dot_product_attention(query, key, value, attn_mask=None, dropo
if do_split:
batch_size_attention, query_tokens, shape_three = query.shape[0], query.shape[1], query.shape[2]
hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
if attn_mask is not None and attn_mask.shape != query.shape:
if attn_mask is not None and attn_mask.shape[:-1] != query.shape[:-1]:
if len(query.shape) == 4:
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], shape_three // attn_mask.shape[2], 1))
else:
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], shape_three // attn_mask.shape[2]))
attn_mask = attn_mask.repeat((batch_size_attention // attn_mask.shape[0], query_tokens // attn_mask.shape[1], 1))
for i in range(batch_size_attention // split_slice_size):
start_idx = i * split_slice_size
end_idx = (i + 1) * split_slice_size