[OpenVINO] Introduce extended quantization dataset options like 'wikitext2:seq_len=128' #1564
Conversation
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
- Add dataset_kwargs attribute and parsing logic to OVQuantizationConfigBase - Parse "dataset:seq_len=value" syntax with validation - Rename seqlen parameter to seq_len in _prepare_causal_lm_calibration_data - Thread dataset_kwargs through all calibration helpers - Add comprehensive unit and integration tests Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Apply black and ruff formatting to changes Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Revert unrelated formatting change to minimize diff Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Address PR feedback: merge tests, update docstrings, use explicit seq_len parameter - Merge TestDatasetIntegration into TestDatasetParsing and move to test_quantization.py - Remove standalone test files (test_dataset_parsing.py, test_dataset_integration.py) - Update dataset docstrings in all OVQuantizationConfigBase child classes - Update CLI --dataset help text in openvino.py - Update dataset documentation in export.mdx - Change from **dataset_kwargs unpacking to explicit seq_len= parameter - Remove unnecessary comment about integrating seq_len Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Add test cases for new dataset format with seq_len option - Add test case to test_exporters_cli.py using wikitext2:seq_len=64 - Add test case to test_quantization.py using c4:seq_len=64 Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Address final PR feedback: remove empty line, adjust test samples, remove redundant tests - Remove empty line between seq_len assignment and tokenizer initialization - Change test to use num_samples=1 instead of 100 for faster execution - Add num_samples=1 to test configuration in test_quantization.py - Remove redundant integration test methods (causal_lm, gsm8k, text_to_text, text_encoder) - Remove redundant list_dataset_backward_compatibility test Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com> Fix serialization and seq_len parameter passing issues - Handle dataset_kwargs deserialization for backward compatibility by extracting it from kwargs - Only pass seq_len to helper functions when it exists in dataset_kwargs to avoid overriding defaults with None - Use conditional kwargs construction for all calibration helpers Co-authored-by: nikita-savelyevv <23343961+nikita-savelyevv@users.noreply.github.com>
nikita-savelyevv
changed the title
|
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. |
nikita-savelyevv
changed the title
added 4 commits
December 17, 2025 17:07
nikita-savelyevv
requested review from
IlyasMoutawwakil,
echarlaix,
popovaan and
rkazants
echarlaix
reviewed
Dec 22, 2025
Collaborator
echarlaix
left a comment
echarlaix
left a comment
There was a problem hiding this comment.
Thanks for the addition!
ljaljushkin
reviewed
Dec 23, 2025
ljaljushkin
approved these changes
Jan 2, 2026
echarlaix
approved these changes
Jan 5, 2026
Collaborator
echarlaix
left a comment
echarlaix
left a comment
There was a problem hiding this comment.
Thanks a lot for the great PR @nikita-savelyevv
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?
Sequence length during quantization calibration data collection can now be provided as a string like
"wikitext2:seq_len=128". The motivation is that sometimes it makes sense to adjust default sequence length value, and this way it will be possible to configure it in those particular cases. For example for configs inside_DEFAULT_4BIT_WQ_CONFIGS.Based on PR by Copilot: nikita-savelyevv#4
Ticket: CVS-176623
Before submitting