New Progress Bar, Backoff, Batching#165
Open
soldni wants to merge 16 commits into
Open
Conversation
soldni
requested review from
kyleclo and
undfined
and removed request for
undfined
May 23, 2024 17:10
undfined
approved these changes
May 28, 2024
undfined
left a comment
undfined
left a comment
Contributor
There was a problem hiding this comment.
Looks nice! A few questions, mostly for my understanding and curious if we can run the BPP without a progress bar? It's small but has some perf impact in aggregate. Also, seems not useful to report per "worker" when distributing compute.
| PATH=/home/vscode/.cargo/bin:$PATH | ||
| cd dolma | ||
| source /home/vscode/miniforge3/bin/activate && pip install cmake "maturin[patchelf]>=1.1,<2.0" | ||
| source /home/vscode/miniforge3/bin/activate && pip install cmake "maturin>=1.5,<2.0" |
| "necessary>=0.4.3", | ||
| "charset-normalizer>=3.2.0", | ||
| "zstandard>=0.20.0", | ||
| "backoff>=2.0.0", |
Contributor
There was a problem hiding this comment.
Is this version required? There's 2 minor versions since this 2.0 release "2.2.1"
Comment on lines
+261
to
+263
| def __radd__(self: BPP, other: BPP) -> BPP: | ||
| """Combine two parallel processors into one.""" | ||
| return other.__add__(self) |
Contributor
There was a problem hiding this comment.
Can you describe when this is useful?
Comment on lines
+278
to
+280
| """Process multiple files. Naively calls process_single for each file, but can be overridden.""" | ||
| for src_path, dst_path, single_kwargs in zip(source_paths, destination_paths, kwargs): | ||
| cls.process_single(source_path=src_path, destination_path=dst_path, queue=queue, **single_kwargs) |
Contributor
There was a problem hiding this comment.
Maybe include an example of overloading this processing method?
| cls.get_logger().warning(message) | ||
|
|
||
| @classmethod | ||
| def _process_batch_and_save_status( |
Contributor
There was a problem hiding this comment.
Is this actually saving status or just the metadata outcome?
| len(all_process_kwargs), | ||
| ) | ||
| # no need to be wasteful with processes: we only need as many cores a the number of batches | ||
| num_processes = min(self.num_processes, len(batches)) |
Contributor
There was a problem hiding this comment.
Can you have more batches than available procs/cores?
| Args: | ||
| iterables (Iterable[T]): One or more iterables to group into batches. | ||
| batch_size (int): The size of each batch. Defaults to 1. | ||
| drop_last (bool): Whether to drop the last batch if it is smaller than `batch_size`. Defaults to False. |
Contributor
There was a problem hiding this comment.
What scenario would you want to drop the last batch?
Comment on lines
+190
to
+191
| if not hasattr(self, "PROGRESS_BAR_CLS"): | ||
| self.PROGRESS_BAR_CLS = BaseProgressBar.from_increment_function(self) |
Contributor
There was a problem hiding this comment.
Can this be run without a progress bar at all?
Comment on lines
+89
to
+127
| class TestBatching(TestCase): | ||
| def test_batching(self): | ||
| a = [1, 2, 3, 4, 5] | ||
| b = [6, 7, 8, 9, 0] | ||
|
|
||
| output = list(batch_iterator(a, b, batch_size=2)) | ||
| self.assertEqual(len(output), 3) | ||
| self.assertEqual(output[0], [(1, 2), (6, 7)]) | ||
| self.assertEqual(output[1], [(3, 4), (8, 9)]) | ||
| self.assertEqual(output[2], [(5,), (0,)]) | ||
|
|
||
| def test_single_batching(self): | ||
| a = [1, 2, 3, 4, 5] | ||
|
|
||
| output = list(batch_iterator(a, batch_size=2)) | ||
|
|
||
| self.assertEqual(len(output), 3) | ||
| self.assertEqual(output[0], [(1, 2)]) | ||
| self.assertEqual(output[1], [(3, 4)]) | ||
| self.assertEqual(output[2], [(5,)]) | ||
|
|
||
| def test_longer_batch_than_slice(self): | ||
| a = list(range(3)) | ||
| b = list(range(3, 6)) | ||
| c = list(range(6, 9)) | ||
|
|
||
| output = list(batch_iterator(a, b, c, batch_size=4)) | ||
|
|
||
| self.assertEqual(len(output), 1) | ||
| self.assertEqual(output[0], [(0, 1, 2), (3, 4, 5), (6, 7, 8)]) | ||
|
|
||
| def test_drop_last(self): | ||
| a = [1, 2, 3, 4, 5] | ||
| b = [6, 7, 8, 9, 0] | ||
|
|
||
| output = list(batch_iterator(a, b, batch_size=2, drop_last=True)) | ||
| self.assertEqual(len(output), 2) | ||
| self.assertEqual(output[0], [(1, 2), (6, 7)]) | ||
| self.assertEqual(output[1], [(3, 4), (8, 9)]) |
| dolma_tests_skip = os.environ.get(DOLMA_TESTS_SKIP_AWS_ENV_VAR) | ||
| LOGGER.info(f"{DOLMA_TESTS_SKIP_AWS_ENV_VAR}: {dolma_tests_skip}") | ||
| return (dolma_tests_skip or "false").lower() == "true" | ||
| dolma_tests_skip = yaml.safe_load(os.environ.get(DOLMA_TESTS_SKIP_AWS_ENV_VAR) or "false") |
Contributor
There was a problem hiding this comment.
safe_load is duplicative if casting to bool() regardless below right? More a nit than anything, I guess it wouldn't hurt...
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
2 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 adds three nice features to
BaseParallelProcessor:parallel.pyPoolWithDebugwrapper aroundmultiprocessing.Poolthat transparently disables multiprocessing when debuggingbackofflibrary to implement backoff and retries in case of failure