Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #5072 from Anai-Guo/fix/lumina-dimoo-attention-kwargs

Browse Source 
fix(lumina): pass to_compute_mask, not use_cache, to LLaDABlock.attention
This commit is contained in:
Vladimir Mandic
2026-08-31 09:46:30 +02:00
committed by GitHub
parent 94371a6215 3991941e5e
commit 4f6575a8bf
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pipelines/lumina_dimmo/lumina_dimoo.py
+3 -3
View File
@@ -874,10 +874,10 @@ class LLaDASequentialBlock(LLaDABlock):
if self._activation_checkpoint_fn is not None:
att, cache = self._activation_checkpoint_fn( # type: ignore
self.attention, q, k, v, attention_bias, layer_past=layer_past, use_cache=use_cache
self.attention, q, k, v, attention_bias, layer_past=layer_past
)
else:
att, cache = self.attention(q, k, v, attention_bias, layer_past=layer_past, use_cache=use_cache)
att, cache = self.attention(q, k, v, attention_bias, layer_past=layer_past)
x = x + self.dropout(att)
@@ -967,7 +967,7 @@ class LLaDALlamaBlock(LLaDABlock):
if self._activation_checkpoint_fn is not None:
att, cache = self._activation_checkpoint_fn( # type: ignore
self.attention, q, k, v, attention_bias, layer_past=layer_past, use_cache=use_cache
self.attention, q, k, v, attention_bias, layer_past=layer_past, to_compute_mask=to_compute_mask
)
else:
att, cache = self.attention(q, k, v, attention_bias, layer_past=layer_past, to_compute_mask=to_compute_mask)