Skip to content

[bugfix] Fix tensor device in Idefics2, Idefics3, and SmolVLM #39975

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

Merged
ArthurZucker merged 4 commits into from
Aug 13, 2025
+21 −12
Merged

[bugfix] Fix tensor device in Idefics2, Idefics3, and SmolVLM #39975

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ab27606
[bugfix] ensure correct tensor device in Idefics2, Idefics3, and Smol…
qgallouedec Aug 7, 2025
36371b0
to cuda
qgallouedec Aug 12, 2025
6ac95c6
Merge branch 'main' into fix-smolvlm-idefics
qgallouedec Aug 12, 2025
b5aa571
Merge branch 'main' into fix-smolvlm-idefics
qgallouedec Aug 12, 2025
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
8 changes: 4 additions & 4 deletions src/transformers/models/idefics2/modeling_idefics2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,11 +144,11 @@ def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.B
position_ids = torch.full(size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0)

for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
nb_patches_h = p_attn_mask[:, 0].sum()
nb_patches_w = p_attn_mask[0].sum()
nb_patches_h = p_attn_mask[:, 0].sum().item()
nb_patches_w = p_attn_mask[0].sum().item()

@zucchini-nlp zucchini-nlp Aug 7, 2025

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

let's move all computation to 'cuda' then and avoid ''item()"

@qgallouedec qgallouedec Aug 12, 2025

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

h_indices = torch.arange(nb_patches_h, device=pixel_values.device, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, device=pixel_values.device, dtype=pixel_values.dtype)
h_indices = torch.arange(nb_patches_h, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, dtype=pixel_values.dtype)
Comment thread
zucchini-nlp marked this conversation as resolved.
Outdated
Comment thread
zucchini-nlp marked this conversation as resolved.
Outdated

fractional_coords_h = h_indices / nb_patches_h * (1 - 1e-6)
fractional_coords_w = w_indices / nb_patches_w * (1 - 1e-6)
Expand Down
8 changes: 4 additions & 4 deletions src/transformers/models/idefics3/modeling_idefics3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,11 +144,11 @@ def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.B
position_ids = torch.full(size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0)

for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
nb_patches_h = p_attn_mask[:, 0].sum()
nb_patches_w = p_attn_mask[0].sum()
nb_patches_h = p_attn_mask[:, 0].sum().item()
nb_patches_w = p_attn_mask[0].sum().item()

h_indices = torch.arange(nb_patches_h, device=pixel_values.device, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, device=pixel_values.device, dtype=pixel_values.dtype)
h_indices = torch.arange(nb_patches_h, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, dtype=pixel_values.dtype)
Comment thread
qgallouedec marked this conversation as resolved.
Outdated
Comment thread
qgallouedec marked this conversation as resolved.
Outdated

fractional_coords_h = h_indices / nb_patches_h * (1 - 1e-6)
fractional_coords_w = w_indices / nb_patches_w * (1 - 1e-6)
Expand Down
8 changes: 4 additions & 4 deletions src/transformers/models/smolvlm/modeling_smolvlm.py
View file
Open in desktop

@ArthurZucker ArthurZucker Aug 12, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

test_export_smolvlm_vision_encoder does not seem to have been triggered

Original file line number Diff line number Diff line change
Expand Up @@ -139,11 +139,11 @@ def forward(self, pixel_values: torch.FloatTensor, patch_attention_mask: torch.B
position_ids = torch.full(size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0)

for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
nb_patches_h = p_attn_mask[:, 0].sum()
nb_patches_w = p_attn_mask[0].sum()
nb_patches_h = p_attn_mask[:, 0].sum().item()
nb_patches_w = p_attn_mask[0].sum().item()

h_indices = torch.arange(nb_patches_h, device=pixel_values.device, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, device=pixel_values.device, dtype=pixel_values.dtype)
h_indices = torch.arange(nb_patches_h, dtype=pixel_values.dtype)
w_indices = torch.arange(nb_patches_w, dtype=pixel_values.dtype)
Comment thread
qgallouedec marked this conversation as resolved.
Outdated
Comment thread
qgallouedec marked this conversation as resolved.
Outdated

fractional_coords_h = h_indices / nb_patches_h * (1 - 1e-6)
fractional_coords_w = w_indices / nb_patches_w * (1 - 1e-6)
Expand Down
Loading