Repository navigation
Conversation
|
@Cyan4973 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D128229489. |
Cyan4973
force-pushed
the
export-D128229489
branch
from
November 2, 2026 19:38
d9eebaa to
013192c
Compare
meta-codesync
Bot
changed the title
Cyan4973
added a commit
to Cyan4973/openzl
that referenced
this pull request
Nov 2, 2026
Summary: Pull Request resolved: meta#1241 Pickles often hold NumPy arrays: `pickle.dump` of dicts of arrays, `pandas.to_pickle`, model weights. On Zenodo, `.pkl` files total 3.0 TB, and NumPy arrays make about 74% of their bytes. Today the `numpy` profile compresses pickles with the generic graph. This diff adds a pickle lexer (`custom_parsers/pickle_lexer.h`), which finds the bytes of NumPy arrays in pickles of protocols 2 to 5 by following their opcodes, without executing anything. The NumPy parser uses it: arrays go to the Numeric Transformer, the rest to zstd. A new `pickle` profile selects this graph. | 26 real pickles | zstd -19 | today | `-p pickle` | | --- | --- | --- | --- | | 16 with arrays (1.33 GB) | x1.565 | x1.488 | x1.779 | | 10 without arrays (1.00 GB) | x1.925 | x1.777 | x1.777 | Pickles without arrays compress as today. Arrays smaller than 16 KiB stay with the opcodes, where they compress better. Object arrays of `.npy` files come in the next diff. Differential Revision: D128229489
Summary: NumPy files hold the arrays of many scientific datasets and some model checkpoints, such as the JAX checkpoints of ViT and GenCast. A `.npy` file holds one array, and a `.npz` archive is a zip of `.npy` members. OpenZL had no profile for them, so they went through generic compression, which does no better than zstd on them. This diff adds a `numpy` CLI profile and `ZL_NumPy_registerGraph()`. - A small lexer reads the header of each `.npy` file, or of each stored `.npy` member of a `.npz` archive (through `zip_lexer`), to find the type and shape of its array. - Arrays of numbers (booleans, integers, floats, complex numbers, datetimes) go to the Numeric Transformer, which picks the transforms for each array. Other arrays, headers and zip records are compressed apart. - Chunking reuses the tensor chunker of the model parsers. | File | zstd -19 | `-p serial` | `-p numpy` | vs zstd -19 | |---|---|---|---|---| | ViT-B/16 JAX checkpoint, float32 `.npz` (413 MB) | x1.289 | x1.286 | **x1.427** | +10.7% | | GenCast JAX checkpoint, float32 `.npz` (230 MB) | x1.075 | x1.073 | **x1.190** | +10.7% | | rMD17 aspirin trajectories, float64 `.npz` (154 MB) | x1.168 | x1.048 | **x1.589** | +36.0% | | rMD17 benzene trajectories, float64 `.npz` (89 MB) | x1.607 | x1.348 | **x2.533** | +57.7% | | Neuropixels templates, float32 `.npy` (15 MB) | x11.25 | x10.30 | **x13.77** | +22.4% | | QuickDraw cat bitmaps, uint8 `.npy` (97 MB) | x2.699 | x2.460 | **x2.784** | +3.2% | Trade-offs: - The Numeric Transformer compresses at 25-50 MB/s on these files; decompression runs at 0.3-2.5 GB/s. On the float32 weights of ViT, the tensor graphs of the model parsers compress 14 to 18 times faster, for a ratio within 1%. Speeding up the Transformer is a separate effort, so the model parsers keep their tuned graphs. - Members deflated by `np.savez_compressed` (MedMNIST, Keras MNIST) are not inflated, so they gain nothing. - In a `.npz` archive, arrays smaller than 16 KiB share a stream with the headers around them, which the generic graph compresses: 0.1-0.2% of the bytes of ViT and GenCast. - Strings, objects, structured types and big-endian arrays are compressed as raw bytes. Differential Revision: D127896258
Summary: Pull Request resolved: meta#1241 Pickles often hold NumPy arrays: `pickle.dump` of dicts of arrays, `pandas.to_pickle`, model weights. On Zenodo, `.pkl` files total 3.0 TB, and NumPy arrays make about 74% of their bytes. Today the `numpy` profile compresses pickles with the generic graph. This diff adds a pickle lexer (`custom_parsers/pickle_lexer.h`), which finds the bytes of NumPy arrays in pickles of protocols 2 to 5 by following their opcodes, without executing anything. The NumPy parser uses it: arrays go to the Numeric Transformer, the rest to zstd. A new `pickle` profile selects this graph. | 26 real pickles | zstd -19 | today | `-p pickle` | | --- | --- | --- | --- | | 16 with arrays (1.33 GB) | x1.565 | x1.488 | x1.779 | | 10 without arrays (1.00 GB) | x1.925 | x1.777 | x1.777 | Pickles without arrays compress as today. Arrays smaller than 16 KiB stay with the opcodes, where they compress better. Object arrays of `.npy` files come in the next diff. Differential Revision: D128229489
Cyan4973
force-pushed
the
export-D128229489
branch
from
November 4, 2026 01:19
013192c to
82e08a6
Compare
This branch has not been deployed
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.
Summary:
Pickles often hold NumPy arrays:
pickle.dumpof dicts of arrays,pandas.to_pickle, model weights. On Zenodo,.pklfiles total 3.0 TB, and NumPy arrays make about 74% of their bytes. Today thenumpyprofile compresses pickles with the generic graph.This diff adds a pickle lexer (
custom_parsers/pickle_lexer.h), which finds the bytes of NumPy arrays in pickles of protocols 2 to 5 by following their opcodes, without executing anything. The NumPy parser uses it: arrays go to the Numeric Transformer, the rest to zstd. A newpickleprofile selects this graph.-p picklePickles without arrays compress as today. Arrays smaller than 16 KiB stay with the opcodes, where they compress better. Object arrays of
.npyfiles come in the next diff.Differential Revision: D128229489