Pass missing arguments to `SeamlessM4Tv2ConformerEncoderLayer.forward…

…()` when gradient checkpointing is enabled (huggingface#31945) * pass missing arguments when gradient checkpointing is enabled for SeamlessM4Tv2 * fix same bug in SeamlessM4Tv1 * pass args, not kwargs
zucchini-nlp · Jul 24, 2024 · 0995c77 · 0995c77
1 parent fc42335
commit 0995c77
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 0 deletions.
diff --git a/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py b/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py
@@ -842,6 +842,8 @@ def forward(
                         hidden_states,
                         attention_mask,
                         relative_position_embeddings,
+                        output_attentions,
+                        conv_attention_mask,
                     )
                 else:
                     layer_outputs = layer(

diff --git a/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py b/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py
@@ -788,6 +788,8 @@ def forward(
                         layer.__call__,
                         hidden_states,
                         attention_mask,
+                        output_attentions,
+                        conv_attention_mask,
                     )
                 else:
                     layer_outputs = layer(