cache_utils: fix QuantizedLayer to correctly propagate reorder_cache, crop, and batch ops to quantized buffers#45510
Open
GitGlimpse895 wants to merge 1 commit intohuggingface:mainfrom
Conversation
… crop, and batch ops to quantized buffers
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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?
QuantizedLayermaintains two separate storage regions: a full-precisionresidual buffer (
self.keys/self.values) and a quantized buffer(
self._quantized_keys/self._quantized_values). However, the fourmutation methods inherited from
DynamicLayer—reorder_cache,crop,batch_repeat_interleave, andbatch_select_indices— onlyoperated on the residual buffer, silently leaving the quantized buffer
untouched.
Concrete failure modes:
reorder_cache): the quantized buffer stays inoriginal beam order while the residual reorders, causing crossed
attention across beams with no error raised.
crop):cumulative_lengthdiverges from the actual stored state, corrupting subsequent
get_seq_lengthcalls.batch_select_indices,batch_repeat_interleave): batch dimension of the quantized bufferis never updated, producing mismatched batch sizes between the two
storage regions.
This PR overrides all four methods in
QuantizedLayer. Since_quantized_keys/_quantized_valuesare opaque backend objects forboth
QuantoQuantizedLayerandHQQQuantizedLayer, the fix uses adequantize → operate → re-quantize pattern, which is backend-agnostic
and does not compound quantization error meaningfully beyond what is
already introduced at the original quantization step.
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
@gante @SunMarc