-
Notifications
You must be signed in to change notification settings - Fork 173
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Pipeline Refactor][Text Generation][Continuous Batching] Integration #1409
Merged
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
dsikka
changed the title
Contin batch integ
[Pipeline Refactor][Text Generation][Continuous Batching] Integration
Nov 16, 2023
dsikka
requested review from
bfineran,
rahul-tuli and
dbogunowicz
and removed request for
bfineran
November 16, 2023 20:37
dbogunowicz
reviewed
Nov 17, 2023
bfineran
reviewed
Nov 20, 2023
…ng in text gen pipeline
bfineran
approved these changes
Nov 21, 2023
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM - as mentioned in the MLE last weekly, we'll need to update the UX around specifying continuous batching
dbogunowicz
approved these changes
Nov 21, 2023
Conditional approval -> please run the relevant integration/unit tests before landing |
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.
Summary
ContinuousBatchingScheduler
with theTextGeneration
PipelineNLEngineOperator
such that batches greater than size 1 can be handled. Specifically includes updating the input/output schema to include split/join methods required by theContinuousBatchingScheduler
and update therun
method such that inputs can be handled by any schedulerSPLIT
node, poll to identify the operator that is currently finished running and schedules the next operator in its split route, until all the operators within each sub graph (i.e. split route) have finished running. This is done through the help of a dataclassSplitRoute
Testing