feat: add overlap function for multi-blocks compression#1850
Open
ZaneMark wants to merge 3 commits into
Open
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
Adds overlap support for multi-block compression/tuning so adjacent block windows can share a configurable number of blocks, aiming to improve continuity during calibration.
Changes:
- Introduce
--nblocks_overlap/nblocks_overlapconfiguration and validate it (0 <= nblocks_overlap < nblocks). - Implement overlap-aware block-window iteration and progress accounting.
- Add logic to compute “stride” outputs for the next window and snapshot/restore overlapped tail blocks.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.
| File | Description |
|---|---|
| auto_round/compressors/data_driven.py | Implements overlap window iteration, stride-based next-input selection, and tail snapshot/restore helpers. |
| auto_round/compressors/config.py | Plumbs nblocks_overlap through extra config/dataclass. |
| auto_round/compressors/base.py | Parses nblocks_overlap and enforces basic validity constraints. |
| auto_round/main.py | Adds the --nblocks_overlap CLI flag and forwards it into tuning config. |
Comment on lines
+445
to
+447
| for module in tail_modules: | ||
| state = {key: value.detach().cpu().clone() for key, value in module.state_dict().items()} | ||
| attrs = {} |
Comment on lines
41
to
47
| lr: float = None, | ||
| lr_scheduler: Callable = None, | ||
| minmax_lr: float = None, | ||
| nblocks: int = 1, | ||
| nblocks_overlap: int = 0, | ||
| to_quant_block_names: Union[str, list, None] = None, | ||
| scale_dtype: str = "fp16", |
Comment on lines
+711
to
+722
| def _get_block_window_starts(self, block_names: list, nblocks: int) -> list[int]: | ||
| overlap = self.nblocks_overlap if nblocks > 1 else 0 | ||
| stride = nblocks - overlap | ||
| block_starts = [] | ||
| block_idx = 0 | ||
| while block_idx < len(block_names): | ||
| remaining = len(block_names) - block_idx | ||
| if block_idx > 0 and overlap > 0 and remaining <= overlap: | ||
| break | ||
| block_starts.append(block_idx) | ||
| block_idx += stride | ||
| return block_starts |
Contributor
|
as discussed offline, the implementation probably has bugs from algorithm perspective |
Contributor
Author
|
10 tasks acc_norm W2A16
W4A16
|
Contributor
|
/azp run Unit-Test-CUDA-AutoRound |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Contributor
|
as overlapping will cause 2X cost, comparing it with baseline directly is not very fair. So please comparing yours with iters x N, e.g., comparing n2o1 with iters 400 and n4o3 with iters 800. |
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.
Description
Introducing overlap support for multi-block compression, enabling overlapping block windows during tuning.
Motivation:
Key changes:
--nblocks_overlapto control overlap between adjacent block windowsType of Change
New feature
Checklist Before Submitting
/azp run Unit-Test-CUDA-AutoRound.To Do
Outcome
Qwen3-8B
Llama-3.1-8B-Instruct