Posit AI Weblog: safetensors 0.1.0

safetensors is a brand new, easy, quick, and protected file format for storing tensors. The design of the file format and its authentic implementation are being led
by Hugging Face, and it’s getting largely adopted of their fashionable ‘transformers’ framework. The safetensors R bundle is a pure-R implementation, permitting to each learn and write safetensor recordsdata.

The preliminary model (0.1.0) of safetensors is now on CRAN.

Motivation

The principle motivation for safetensors within the Python group is safety. As famous
within the official documentation:

The principle rationale for this crate is to take away the necessity to use pickle on PyTorch which is utilized by default.

Pickle is taken into account an unsafe format, because the motion of loading a Pickle file can
set off the execution of arbitrary code. This has by no means been a priority for torch
for R customers, for the reason that Pickle parser that’s included in LibTorch solely helps a subset
of the Pickle format, which doesn’t embrace executing code.

Nonetheless, the file format has extra benefits over different generally used codecs, together with:

• Assist for lazy loading: You’ll be able to select to learn a subset of the tensors saved within the file.

• Zero copy: Studying the file doesn’t require extra reminiscence than the file itself.
  (Technically the present R implementation does makes a single copy, however that may
  be optimized out if we actually want it in some unspecified time in the future).

• Easy: Implementing the file format is easy, and doesn’t require complicated dependencies.
  Because of this it’s a superb format for exchanging tensors between ML frameworks and
  between totally different programming languages. As an illustration, you may write a safetensors file
  in R and cargo it in Python, and vice-versa.

There are extra benefits in comparison with different file codecs widespread on this area, and
you may see a comparability desk right here.

Format

The safetensors format is described within the determine under. It’s principally a header file
containing some metadata, adopted by uncooked tensor buffers.

Primary utilization

safetensors might be put in from CRAN utilizing:

set up.packages("safetensors")

We will then write any named checklist of torch tensors:

library(torch)
library(safetensors)

tensors <- checklist(
  x = torch_randn(10, 10),
  y = torch_ones(10, 10)
)

str(tensors)
#> Checklist of two
#>  $ x:Float [1:10, 1:10]
#>  $ y:Float [1:10, 1:10]

tmp <- tempfile()
safe_save_file(tensors, tmp)

It’s potential to go extra metadata to the saved file by offering a metadata
parameter containing a named checklist.

Studying safetensors recordsdata is dealt with by safe_load_file, and it returns the named
checklist of tensors together with the metadata attribute containing the parsed file header.

tensors <- safe_load_file(tmp)
str(tensors)
#> Checklist of two
#>  $ x:Float [1:10, 1:10]
#>  $ y:Float [1:10, 1:10]
#>  - attr(*, "metadata")=Checklist of two
#>   ..$ x:Checklist of three
#>   .. ..$ form       : int [1:2] 10 10
#>   .. ..$ dtype       : chr "F32"
#>   .. ..$ data_offsets: int [1:2] 0 400
#>   ..$ y:Checklist of three
#>   .. ..$ form       : int [1:2] 10 10
#>   .. ..$ dtype       : chr "F32"
#>   .. ..$ data_offsets: int [1:2] 400 800
#>  - attr(*, "max_offset")= int 929

At the moment, safetensors solely helps writing torch tensors, however we plan so as to add
help for writing plain R arrays and tensorflow tensors sooner or later.

Future instructions

The following model of torch will use safetensors as its serialization format,
which means that when calling torch_save() on a mannequin, checklist of tensors, or different
kinds of objects supported by torch_save, you’ll get a sound safetensors file.

That is an enchancment over the earlier implementation as a result of:

1. It’s a lot quicker. Greater than 10x for medium sized fashions. Could possibly be much more for big recordsdata.
   This additionally improves the efficiency of parallel dataloaders by ~30%.

2. It enhances cross-language and cross-framework compatibility. You’ll be able to prepare your mannequin
   in R and use it in Python (and vice-versa), or prepare your mannequin in tensorflow and run it
   with torch.

If you wish to strive it out, you may set up the event model of torch with:

remotes::install_github("mlverse/torch")

Photograph by Nick Fewings on Unsplash

Reuse

Textual content and figures are licensed below Inventive Commons Attribution CC BY 4.0. The figures which have been reused from different sources do not fall below this license and might be acknowledged by a notice of their caption: “Determine from …”.

Quotation

For attribution, please cite this work as

Falbel (2023, June 15). Posit AI Weblog: safetensors 0.1.0. Retrieved from https://blogs.rstudio.com/tensorflow/posts/2023-06-15-safetensors/

BibTeX quotation

@misc{safetensors,
  writer = {Falbel, Daniel},
  title = {Posit AI Weblog: safetensors 0.1.0},
  url = {https://blogs.rstudio.com/tensorflow/posts/2023-06-15-safetensors/},
  12 months = {2023}
}