🔴 🔴 🔴 supersede paligemma forward to shift pos id indexing #36859
Conversation
molbap
changed the title
|
run slow: gemma3 |
|
run_slow: gemma3 |
1 similar comment
|
run_slow: gemma3 |
|
This comment contains run-slow, running the specified jobs: This comment contains run-slow, running the specified jobs: models: ['models/gemma3'] |
|
@molbap BTW, this is not the only place where this shift occurs, it also happens in |
|
Nice catch, thanks a lot @oceanusxiv 🤗 |
|
run_slow: gemma3 |
|
This comment contains run-slow, running the specified jobs: This comment contains run-slow, running the specified jobs: models: ['models/gemma3'] |
|
run_slow: gemma3 |
|
This comment contains run-slow, running the specified jobs: This comment contains run-slow, running the specified jobs: models: ['models/gemma3'] |
|
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. |
|
don't merge yet- the 4b crops test with pan_and_scan is broken with this change. Unsure yet if related to #36864 , checking |
|
Ok, I think unrelated as well, https://huggingface.co/google/gemma-3-4b-pt/commit/2166bd1fba050eab3ed4fdd0b591ad1e11c96802 added another eos_token which causes the generation to stop earlier. Confirming, fixing in #36882 then we can merge |
|
Not sure what is causing the difference, but it's not big AFAIK and other integration tests pass. I'll merge now. |
What does this PR do?
Should fix #36856
This is because, in the modular file,
Gemma3ForConditionalGenerationinherits fromPaliGemmaForConditionalGenerationwhere this specific fix happens. But it is specific to PaliGemma input ordering and should not have been propagated to Gemma3.