Skip to content

GitLab

"...git@developer.sourcefind.cn:chenpangpang/transformers.git" did not exist on "accccdd0087263a1e494e9c9ec30a43043ff3905"
Unverified Commit 4ad5adaf authored by amyeroberts's avatar amyeroberts Committed by GitHub
Browse files

Fix copies for DBRX - neuron fix (#30610)

parent f9530258
Hide whitespace changes
Inline Side-by-side
Showing with 5 additions and 2 deletions
src/transformers/models/dbrx/modeling_dbrx.py
View file @ 4ad5adaf
...@@ -1256,8 +1256,11 @@ class DbrxModel(DbrxPreTrainedModel): ...@@ -1256,8 +1256,11 @@ class DbrxModel(DbrxPreTrainedModel):
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
if attention_mask.dim() == 2: if attention_mask.dim() == 2:
mask_length = attention_mask.shape[-1] mask_length = attention_mask.shape[-1]
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0) padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype) padding_mask = padding_mask == 0
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
padding_mask, min_dtype
)
elif attention_mask.dim() == 4: elif attention_mask.dim() == 4:
# backwards compatibility: we allow passing a 4D attention mask shorter than the input length with # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
# cache. In that case, the 4D attention mask attends to the newest tokens only. # cache. In that case, the 4D attention mask attends to the newest tokens only.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment