[qwen-vl] fix position ids#40490
Conversation
|
run-slow: qwen2_vl, qwen2_5_vl |
|
This comment contains run-slow, running the specified jobs: models: ['models/qwen2_5_vl', 'models/qwen2_vl'] |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
run-slow: qwen2_5_vl, qwen2_vl |
|
This comment contains run-slow, running the specified jobs: models: ['models/qwen2_5_vl', 'models/qwen2_vl'] |
|
prob cc @vasqu for FA2 and packed attention |
|
run-slow: qwen2_5_omni, qwen2_5_vl, qwen2_vl |
|
This comment contains run-slow, running the specified jobs: models: ['models/qwen2_5_omni', 'models/qwen2_5_vl', 'models/qwen2_vl'] |
vasqu
left a comment
There was a problem hiding this comment.
Just nits and question for my understanding
We could add a test with the specific image/prompt where the outputs are significantly different, but I don't think it is necessary
If it's not too much of a hassle, I'd appreciate it but no worries if not
|
[For maintainers] Suggested jobs to run (before merge) run-slow: qwen2_5_omni, qwen2_5_vl, qwen2_vl |
What does this PR do?
Fixes #40136 and fixes #40154
The accuracy on lm eval is restored. The issue was in position ids which weren't prepared correctly when generating. Using packed or unpacked positions doesn't affect anything, as I thought at first
Ig the values in position ids do not affect much generation results, given that slow tests weren't failing. We could add a test with the specific image/prompt where the outputs are significantly different, but I don't think it is necessary