riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-30 17:52:35 +06:00
Code Issues Actions 30 Packages Projects Releases Wiki Activity

[InstructBlip] Fix int8/fp4 issues (#24888)

Browse Source 
* fix dtype issue

* revert `.float()`

* fix copies
This commit is contained in:
Younes Belkada 2023-07-18 19:24:36 +02:00 committed by GitHub
parent 3ec10e6c76
commit a9e067a45c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/transformers/models/instructblip/modeling_instructblip.py
View File

@ -558,7 +558,6 @@ class InstructBlipVisionModel(InstructBlipPreTrainedModel):
return self.embeddings
# Copied from transformers.models.blip_2.modeling_blip_2.Blip2QFormerMultiHeadAttention with Blip2->InstructBlip
class InstructBlipQFormerMultiHeadAttention(nn.Module):
def __init__(self, config, is_cross_attention=False):
super().__init__()
@ -659,13 +658,14 @@ class InstructBlipQFormerMultiHeadAttention(nn.Module):
attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
attention_scores_dtype = attention_scores.dtype
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in BertModel forward() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = nn.Softmax(dim=-1)(attention_scores).to(attention_scores_dtype)
if is_cross_attention and self.save_attention:
self.save_attention_map(attention_probs)
@ -1038,6 +1038,7 @@ class InstructBlipQFormerEmbeddings(nn.Module):
else:
embeddings = query_embeds
embeddings = embeddings.to(self.layernorm.weight.dtype)
embeddings = self.layernorm(embeddings)
embeddings = self.dropout(embeddings)
return embeddings