Skip to content

mtmd: fix miscounting n_tokens #24656

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
ngxson merged 1 commit into from
Jun 15, 2026
Merged

mtmd: fix miscounting n_tokens #24656

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
17 changes: 8 additions & 9 deletions tools/mtmd/mtmd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,16 +96,15 @@ struct mtmd_image_tokens {
// [BOI] [row0 tokens + newline] ... [row(ny-1) tokens + newline] [EOI]
return (nx + 1) * ny + 2;
}
// [QWEN_VIDEO] this logic is quite ugly, it's mostly to make qwen-vl temporal merge work, can be improved in the future
if (batch_f32.entries.size() == 1 || n_temporal_merge == 1) {

@gabe-l-hart gabe-l-hart Jun 17, 2026

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.

@ngxson it looks like the removal of this short-circuit is what broke Granite 4 Vision. In the previous path, this triggers because n_temporal_merge == 1 which falls down into return nx * ny (724 * 1). Now without this, it falls down to return nx * ny * nz (724 * 1 * 7).

@gabe-l-hart gabe-l-hart Jun 17, 2026

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.

I think the issue is that for llava-next models, nz is the number of total tiles versus the bug in the upstream issue (#24634) where it's the number of images in a multi-image batch.

@ngxson ngxson Jun 17, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

@gabe-l-hart can you check if this PR resolves the problems? #24732

for ref, the problem was no about token counting, but it was because the llava-uhd try to construct a batch, that conflicts with #24384

@gabe-l-hart gabe-l-hart Jun 17, 2026

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.

Yep, that does it! Much better than my fix #24733

return nx * ny;
}
uint32_t nz = batch_f32.entries.size();
// TODO: simplify this by repeating the last frame until it fits the temporal merge
if (nz % n_temporal_merge != 0) {
nz = nz / n_temporal_merge + 1;
} else {
nz = nz / n_temporal_merge;
if (n_temporal_merge > 1) {
// [QWEN_VIDEO] this logic is quite ugly, it's mostly to make qwen-vl temporal merge work, can be improved in the future
// TODO: simplify this by repeating the last frame until it fits the temporal merge
if (nz % n_temporal_merge != 0) {
nz = nz / n_temporal_merge + 1;
} else {
nz = nz / n_temporal_merge;
}
}
return nx * ny * nz;
}
Expand Down