Support sliding window in CB#40688
Merged
ArthurZucker merged 40 commits intohuggingface:mainfrom Sep 9, 2025
Merged
Conversation
|
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. |
c3f6e76 to
dbf7b83
Compare
Collaborator
Author
|
Removing draft status, benchmark with / without PR gives:
with |
ArthurZucker
reviewed
Sep 8, 2025
src/transformers/generation/continuous_batching/continuous_api.py
Outdated
Show resolved
Hide resolved
ArthurZucker
approved these changes
Sep 8, 2025
Collaborator
ArthurZucker
left a comment
There was a problem hiding this comment.
Let's remove global var SLIDING_WINDOW
20cfcc5 to
0fe77e8
Compare
McPatate
approved these changes
Sep 9, 2025
| input_ids = torch.tensor([input_ids]).to("cuda") | ||
| attention_mask = torch.ones_like(input_ids) | ||
| outputs = model.generate(input_ids, attention_mask=attention_mask, generation_config=generation_config) | ||
| # attention_mask = torch.ones_like(input_ids) |
Member
There was a problem hiding this comment.
Suggested change
| # attention_mask = torch.ones_like(input_ids) |
?
Collaborator
Author
There was a problem hiding this comment.
Throws a warning otherwise
src/transformers/generation/continuous_batching/continuous_api.py
Outdated
Show resolved
Hide resolved
| self._generation_thread = threading.Thread(target=self._run_generation_loop) | ||
| self._generation_thread.start() | ||
| logger.info("Continuous batching manager started.") | ||
| logger.info(f"Continuous batching manager started at {time.time() - self.creation_time} seconds") |
Member
There was a problem hiding this comment.
Suggested change
| logger.info(f"Continuous batching manager started at {time.time() - self.creation_time} seconds") | |
| logger.info(f"Continuous batching manager started at {time.time() - self.creation_time}") |
This isn't really useful imo, usually you configure your logger to include a timestamp. I'd still keep the start_time in case we need it for metrics and whatnot
Collaborator
Author
There was a problem hiding this comment.
Kept it as debug with better message
src/transformers/generation/continuous_batching/continuous_api.py
Outdated
Show resolved
Hide resolved
src/transformers/generation/continuous_batching/continuous_api.py
Outdated
Show resolved
Hide resolved
Co-authored-by: Luc Georges <McPatate@users.noreply.github.com>
vasqu
reviewed
Sep 9, 2025
vijayabhaskar-ev
pushed a commit
to vijayabhaskar-ev/transformers
that referenced
this pull request
Oct 2, 2025
* CB example: better compare feature * Cache managers, still issue w/ effective length * WIP -- fix for effective length * Renames * Wroking, need better parity checks, we mind be missing 1 token * Small fixes * Fixed wrong attn mask and broke cache into pieces * Warmup is slowing down things, disabling it * Cache was too big, fixed * Simplified index objects * Added a profile option to the example * Avoid calls to memory reporing tools * Restore full attention read indices for better latency * Adressed some TODOS and style * Docstrings for cache managers * Docstrings for Schedulers * Refactor scheudlers * [Important] Cache fix for sliding window, check with small sw size * Updated doc for cache memory compute and cache as a whole * Moved a todo * Nits and style * Fix for when sliding window is smaller than max batch per token * Paged interface update * Support for FLash in new API * Fix example CB * Fix bug in CB for paged * Revert example * Style * Review compliance * Style * Styleeeee * Removed NO_SLIDING_WINDOW * Review huggingface#2 compliance * Better art * Turn cum_seqlens_k in a dict * Attn mask is now a dict * Update examples/pytorch/continuous_batching.py Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> * Adressed McPatate pro review * Style and fix --------- Co-authored-by: Luc Georges <McPatate@users.noreply.github.com>
snorkelopstesting1-a11y
pushed a commit
to snorkel-marlin-repos/huggingface_transformers_pr_40688_b9926daa-ae1d-440f-92f4-527975c3c44e
that referenced
this pull request
Oct 11, 2025
Original PR #40688 by remi-or Original: huggingface/transformers#40688
4 tasks
snorkelopstesting1-a11y
added a commit
to snorkel-marlin-repos/huggingface_transformers_pr_40688_b9926daa-ae1d-440f-92f4-527975c3c44e
that referenced
this pull request
Oct 11, 2025
Merged from original PR #40688 Original: huggingface/transformers#40688
snorkelopstesting1-a11y
pushed a commit
to snorkel-marlin-repos/huggingface_transformers_pr_40688_86a525f7-187b-44c1-9196-444938d2c9a2
that referenced
this pull request
Oct 11, 2025
Original PR #40688 by remi-or Original: huggingface/transformers#40688
4 tasks
snorkelopstesting1-a11y
added a commit
to snorkel-marlin-repos/huggingface_transformers_pr_40688_86a525f7-187b-44c1-9196-444938d2c9a2
that referenced
this pull request
Oct 11, 2025
Merged from original PR #40688 Original: huggingface/transformers#40688
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
5 participants
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.
This PR introduces support for sliding window attention in CB, using an allocator that can support both full attention and sliding window attention. Code works with almost similar performances (665 -> 655 tok/s in the example) but I am putting the PR in draft status until: