Repository navigation
[quantization] Add cached decode calibration for Gemma4 - #1078
Merged
Merged
Conversation
Slice text attention masks using the cache's active key span before updates so cropped sliding keys and shared-KV consumers receive aligned masks. Keep existing observer and fake-quantization boundaries unchanged. Add a synthetic prefill plus two cached decode regression with exact mask checks and parity against a full cache at the existing tolerance. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Resolve the opt-in global count and PTQ stage override consistently for prompt budgeting and execution, preserving the default prefill-only path. Run a fresh multimodal prefill and exactly D greedy single-token cached forwards through the prepared wrapper, with per-token PLE and sample-local cache/shared state. Validate processed inputs and actual wrapper capacity. Document the settings and supported padding. Cover control flow, settings, image routing, cache growth, PLE/shared-KV, sliding attention and conversion using deterministic CPU recording tests and real tiny Gemma4 wrappers. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
mhs4670go
changed the title
mhs4670go
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Gemma4 PTQ calibration currently runs only uncached prefill.
Add opt-in prefill followed by exactly D single-token cached forwards through the prepared wrapper so calibration can exercise the cache read path.