[continuous batching] singleton pattern for scheduler #1391
Conversation
This file contains 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
rahul-tuli
reviewed
Nov 7, 2023
| """ | ||
| if _GLOBAL_SCHEDULER is not None: | ||
| return _GLOBAL_SCHEDULER | ||
| return cls(max_workers=1) |
There was a problem hiding this comment.
Do we need to assign the new instance to _GLOBAL_SCHEDULER?
bfineran
force-pushed
the
continuous-batching-singleton-scheduler
branch
from
85a5805
to
ad24f04
Compare
bfineran
force-pushed
the
continuous-batching-scheduler
branch
from
f0fb8d5
to
4b1472d
Compare
dsikka
reviewed
Nov 13, 2023
| ... | ||
| continuous_batching_scheduler = ContinuousBatchingScheduler.get_instance() | ||
| continuous_batching_scheduler.add_engine_operator(engine_operator) | ||
|
|
There was a problem hiding this comment.
If this is the lifecycle we're proposing, doesnt that mean we can only start with one worker at any given point?
There was a problem hiding this comment.
yes, I think in general we will only want one worker so that for a given batch the engine will have full access to all resources. otherwise we'd need to run the engines in multistream mode which I am not sure is supported for KV cache right now
bfineran
force-pushed
the
continuous-batching-scheduler
branch
from
4b1472d
to
e2a0761
Compare
bfineran
force-pushed
the
continuous-batching-singleton-scheduler
branch
from
ad24f04
to
e48005f
Compare
bfineran
added a commit
that referenced
this pull request
Dec 6, 2023
* Pipelines Refactor - Initial Impl (#1287) * [Pipeline Refactor] Additional functionality, engine operator, linear router and image classification pipeline/operators/example (#1325) * initial functionality and working example with image classification * remove testing image * update args * initial functionality and working example with image classification * remove testing image * pr comments * defines schemas for operators and test * add image classification test, PR comments * fix input/output handling in pipeline and operator base classes to be more generic; remove context * add additional operator input message * typo fix * [v2] EngineOperator updates to make continuous batching easier (#1371) * [v2] EngineOperator updates to make continuous batching easier * test fixes * [Pipeline Refactor] Update routes, text generation initial functionality (#1348) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * [Pipeline Refactor] Additional Operators, Route update and completed generation functionality (#1356) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * initial functionality and working example with image classification * updates func * prompt inference, initial functionality * finish generation operators and update routes * further breakdown operators * add operators * fix can_operate condition * update can_operate to not rely on the inference_state * rebase + update * fix condition * fix capacity settting again * typo fixes * [Pipeline Refactor] Split/Join Functionality for multiple prompts (#1384) * add split/join functionality * update router to include split/join in parent class, refactor pipeline code to remove repeat code, update map function * process multiple generations * move map to base class * [Pipeline Refactor] Unit Testing for Text Generation Operators (#1392) * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * fix name * [Continuous Batching] Queue Implementation to support batching grouping and prioritization (#1373) * [Continuous Batching] Queue Implementation to support batching grouping and prioritization * has_key method * thread safety * add blocking option for pop_batch * update docstring * allow mutex to be shared across continuous batching objects * revert last commit * [Continuous Batching] Executor thread for running continuous batching (#1374) * [Continuous Batching] Executor thread for running continuous batching * quality * ensure that executor stops when main thread does - clean up test hack * [ContinuousBatching] ContinuousBatchingScheduler Implementation (#1375) * [ContinuousBatching] ContinuousBatchingScheduler Implementation * cleanup unnecessary stop condition * [continuous batching] singleton pattern for scheduler (#1391) * [continuous batching] singleton pattern for scheduler * catch from review * [Pipeline Refactor][Text-Generation] Create a helper function for creating engine_inputs (#1364) * rebasing off my initial commit * cleanups * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * Delete tests/deepsparse/v2/unit/text_generation/test_msic.py --------- Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com> * [Pipeline Refactor][Text-Generation] Refactor `transformers` helpers functions (#1394) * add split/join functionality * update router to include split/join in parent class, refactor pipeline code to remove repeat code, update map function * process multiple generations * initial commit * fix error * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * Delete tests/deepsparse/v2/unit/text_generation/test_msic.py * pipeline runs, but incorrectly * Revert "pipeline runs, but incorrectly" This reverts commit 51c4ee6. * PR review comments --------- Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com> * [Text Generation][V2] End-to-end tests (#1402) * initial commit * initial commit * its working now * beautification * thank you Dipika <3 * ready to review * [Pipeline Refactor][Text Generation][Continuous Batching] Integration (#1409) * update split/join * use map * update * run end-to-end * clean-up * fix bug with batch size, introduce SplitRoute dataclass * update tests to use new inputs/outputs * use the normal scheduler for internal kv_cache * add pipeline inpuits * clean-up * change engine type, update docstrings, update override function to be more generic * move subgraph functionality to its own function; clean-up cont batching in text gen pipeline * update linear pathway to also use subgraph execution * rebase fix * fix tests * [Pipeline Refactor] Operator Registry (#1420) * initial registry functionality * use sparsezoo mixin * [Pipeline Refactor] Fix Operator scheduling to fix issue with slow execution (#1453) * fix scheduling to fix issue with engine running very slowly; introduce new completed attribute for Subgraph instead of checking instance type * fix warning message * [Pipeline Refactor] Add `Pipeline.create` method to initialize pipelines (#1457) * add pipeline create method for pipeline creation using the operator registry * add instance check * [Pipeline Refactor] async (#1380) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * initial functionality and working example with image classification * updates func * prompt inference, initial functionality * finish generation operators and update routes * further breakdown operators * add operators * fix can_operate condition * update can_operate to not rely on the inference_state * rebase + update * fix condition * async initial functionality * fix capacity settting again * add blocking * more testing * update to use split/join * fix * rebase fix * remove index * change event loop * rebase fix * update async run to use new operator scheduling properly * rebase fixes (#1458) * more fixes (#1459) --------- Co-authored-by: Benjamin Fineran <bfineran@users.noreply.github.com> Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>
dbogunowicz
added a commit
that referenced
this pull request
Jan 2, 2024
* Pipelines Refactor - Initial Impl (#1287) * [Pipeline Refactor] Additional functionality, engine operator, linear router and image classification pipeline/operators/example (#1325) * initial functionality and working example with image classification * remove testing image * update args * initial functionality and working example with image classification * remove testing image * pr comments * defines schemas for operators and test * add image classification test, PR comments * fix input/output handling in pipeline and operator base classes to be more generic; remove context * add additional operator input message * typo fix * [v2] EngineOperator updates to make continuous batching easier (#1371) * [v2] EngineOperator updates to make continuous batching easier * test fixes * [Pipeline Refactor] Update routes, text generation initial functionality (#1348) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * [Pipeline Refactor] Additional Operators, Route update and completed generation functionality (#1356) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * initial functionality and working example with image classification * updates func * prompt inference, initial functionality * finish generation operators and update routes * further breakdown operators * add operators * fix can_operate condition * update can_operate to not rely on the inference_state * rebase + update * fix condition * fix capacity settting again * typo fixes * add split/join functionality * update router to include split/join in parent class, refactor pipeline code to remove repeat code, update map function * process multiple generations * initial commit * fix error * [Pipeline Refactor] Split/Join Functionality for multiple prompts (#1384) * add split/join functionality * update router to include split/join in parent class, refactor pipeline code to remove repeat code, update map function * process multiple generations * move map to base class * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * [Pipeline Refactor] Unit Testing for Text Generation Operators (#1392) * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * fix name * Delete tests/deepsparse/v2/unit/text_generation/test_msic.py * [Continuous Batching] Queue Implementation to support batching grouping and prioritization (#1373) * [Continuous Batching] Queue Implementation to support batching grouping and prioritization * has_key method * thread safety * add blocking option for pop_batch * update docstring * allow mutex to be shared across continuous batching objects * revert last commit * [Continuous Batching] Executor thread for running continuous batching (#1374) * [Continuous Batching] Executor thread for running continuous batching * quality * ensure that executor stops when main thread does - clean up test hack * [ContinuousBatching] ContinuousBatchingScheduler Implementation (#1375) * [ContinuousBatching] ContinuousBatchingScheduler Implementation * cleanup unnecessary stop condition * [continuous batching] singleton pattern for scheduler (#1391) * [continuous batching] singleton pattern for scheduler * catch from review * [Pipeline Refactor][Text-Generation] Create a helper function for creating engine_inputs (#1364) * rebasing off my initial commit * cleanups * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * Delete tests/deepsparse/v2/unit/text_generation/test_msic.py --------- Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com> * pipeline runs, but incorrectly * it works for a single sequence * cleanup. now lets figure out how to run multiple sequences * [Pipeline Refactor][Text-Generation] Refactor `transformers` helpers functions (#1394) * add split/join functionality * update router to include split/join in parent class, refactor pipeline code to remove repeat code, update map function * process multiple generations * initial commit * fix error * unit testing for text generation operators * additional changes * unit testing completion * remove debug * fix * add todo * more clean-up * fix test * add docstrings/comments * break out tests to individual unit test files; add conftest and make scope of fixtures module to help with speed * Delete tests/deepsparse/v2/unit/text_generation/test_msic.py * pipeline runs, but incorrectly * Revert "pipeline runs, but incorrectly" This reverts commit 51c4ee6. * PR review comments --------- Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com> * [Text Generation][V2] End-to-end tests (#1402) * initial commit * initial commit * its working now * beautification * thank you Dipika <3 * ready to review * integration tests pass * [Pipeline Refactor][Text Generation][Continuous Batching] Integration (#1409) * update split/join * use map * update * run end-to-end * clean-up * fix bug with batch size, introduce SplitRoute dataclass * update tests to use new inputs/outputs * use the normal scheduler for internal kv_cache * add pipeline inpuits * clean-up * change engine type, update docstrings, update override function to be more generic * move subgraph functionality to its own function; clean-up cont batching in text gen pipeline * update linear pathway to also use subgraph execution * rebase fix * fix tests * [Pipeline Refactor] Operator Registry (#1420) * initial registry functionality * use sparsezoo mixin * fix tricky rebase * one more cleanup * got tests to work after rebase. implementing SPLIT and JOIN in linearouter now * pipeline working, with GraphRouter. Needs some more testing * ready for review * cleanup * simplify after PR review round * [Pipeline Refactor] Fix Operator scheduling to fix issue with slow execution (#1453) * fix scheduling to fix issue with engine running very slowly; introduce new completed attribute for Subgraph instead of checking instance type * fix warning message * [Pipeline Refactor] Add `Pipeline.create` method to initialize pipelines (#1457) * add pipeline create method for pipeline creation using the operator registry * add instance check * [Pipeline Refactor] async (#1380) * initial functionality and working example with image classification * remove testing image * rebase fixes * initial functionality and working example with image classification * text gen * updates func * prompt inference, initial functionality * remove image; update state docstring * Fix typo * add todo for split/join * remove context, clean-up args, remove prefill_preprocess_operaator * fix docstrings * initial functionality and working example with image classification * updates func * prompt inference, initial functionality * finish generation operators and update routes * further breakdown operators * add operators * fix can_operate condition * update can_operate to not rely on the inference_state * rebase + update * fix condition * async initial functionality * fix capacity settting again * add blocking * more testing * update to use split/join * fix * rebase fix * remove index * change event loop * rebase fix * update async run to use new operator scheduling properly * rebase fixes (#1458) * more fixes (#1459) * bring back functionalities that were lost in v2 during rebasing * Update src/deepsparse/transformers/helpers.py * ready for review * bring tests back" * quality * original readme * addressing Dipikas comments * Update src/deepsparse/transformers/pipelines/text_generation/pipeline_no_kv_cache.py * addressing PR review --------- Co-authored-by: Benjamin Fineran <bfineran@users.noreply.github.com> Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>
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.
singleton pattern for continuous batching scheduler is favorable so that engine requests coming from different pipelines are scheduled together and do not interfere with each other.
this diff provides a simple implementation and extends the docstring with a small example