Avoid incorrect generations for KV caches containing more than sliding_window tokens #38156
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Gemma3 generates incoherent output when manually calling
forward
with an instance ofHybridCache
which contains more thansliding_window
tokens of content.This is because the call to
past_key_values.get_seq_len()
always returns the sequence length as measured by the cache of the very first layer. Because this is a local attention layer, itssequence_length
never extends beyondconfig.sliding_window
. This leads to an incorrect computation ofcache_position
and incoherent generations down the line.To fix it you can simply provide the correct
cache_position
manually.This behavior is impossible to fix without changing
get_seq_len
ofHybridCache
, so I propose to simply raise an informative error message for now.Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
@ArthurZucker