Skip to content

Conversation

@ngxson
Copy link
Collaborator

@ngxson ngxson commented May 28, 2025

Fix #13865

Provided a model with 3 mrope sections, a rotated vector looks like this: [time, time+x, time+y]

This works with the assumption that mrope with x == y == 0 is identical to doing a neox rope with theta = +time

For example, when we want to shift the time to time-1, that mean we shift from [time, time+x, time+y] to [time-1, time-1+x, time-1+y] ; We can simply apply neox rope with theta = -1 to archive the same effect

@ngxson ngxson requested a review from ggerganov May 28, 2025 18:14
Copy link
Member

@ggerganov ggerganov left a comment

Choose a reason for hiding this comment

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

🪄

@ngxson ngxson merged commit 763d06e into ggml-org:master May 28, 2025
41 of 46 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Eval bug: KV cache shifting does not work for Qwen2.5VL

2 participants