Skip to content

Fix Tensor Dimension Mismatch in Padding Operation for Batch Processing #254

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix Tensor Dimension Mismatch in Padding Operation for Batch Processing #254

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
951bed0
Fix tensor dimension mismatch in padding operation
sugary199 Apr 10, 2024
60ed6e2
Increase the difference in token numbers between the two examples
sugary199 Apr 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions examples/batch_chat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@
img_paths = ['examples/image1.webp',
'examples/image1.webp']
questions = ['Please describe this image in detail.',
'What is the text in this images?']
'What is the text in this images? Please describe it in detail.']

assert len(img_paths) == len(questions)

Expand Down Expand Up @@ -66,7 +66,8 @@
batch_inputs, batch_masks, batch_atten_masks = [], [], []
for inputs, im_mask in zip(inputs_list, masks_list):
if im_mask.shape[1] < max_len:
pad_embeds = torch.cat([pad_embed]*(max_len - im_mask.shape[1]))
pad_length = max_len - im_mask.shape[1]
pad_embeds = pad_embed.repeat(1, pad_length, 1)
pad_masks = torch.tensor([0]*(max_len - im_mask.shape[1])).unsqueeze(0).cuda()
inputs = torch.cat([pad_embeds, inputs['inputs_embeds']], dim=1)
atten_masks = torch.cat([pad_masks, torch.ones_like(im_mask)], dim=1)
Expand Down