{{ message }}
Repository navigation
Conversation
Member
Contributor
Author
|
Thanks, Tom — I pushed your concise three-loss regression into Fresh CPU validation (Python 3.12.14, Torch 2.14.1+cpu, Transformers 5.19.0, NumPy 2.5.4):
At the previous head |
Afloat16
marked this pull request as ready for review
October 14, 2026 13:35
Member
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Preserve encoder input masks across calls to the three uncached multi-vector losses by shallow-copying each input column dictionary before encoding. The returned scoring masks and tensor storage remain unchanged.
Identified through code analysis and reproduced with a composed multi-vector training objective: with classic query expansion (
attend=False), expansion positions are excluded from encoder attention but included in MaxSim. Writing that scoring mask back into the caller's features changes encoder attention for a later loss term. Related dense-mask work in #3944 does not isolate these loss inputs.The regression now uses the concise three-loss case suggested by the maintainer in the existing
test_misc.pymodule.Validation
Reproducer:
Fresh Linux CPU checks on Python 3.12.14, Torch 2.14.1+cpu, Transformers 5.19.0, NumPy 2.5.4:
4a3b5cd6ec718e421f57e824a41ed3fd99595df6.python -m pytest tests/multi_vector_encoder/losses/test_misc.py -q --tb=short: 48 passed, no skips.git diff --check: passed.The current production expressions are unchanged from the previously validated fix; this follow-up simplifies regression coverage and removes comments. The full repository suite, distributed/GPU execution, native pretrained VLM/Flash Attention and inference-provider/backend checks were not run in this follow-up. No performance improvement is claimed.
The previous-head Python 3.11/3.12 CI failures occur before test collection because pyarrow requires NumPy >=2 but those environments have 1.26.4. Updated-head hosted CI remains pending; no all-CI pass is claimed.