Package 'safetensors'

Title: Safetensors File Format
Description: A file format for storing tensors that is secure (doesn't allow for code execution), fast and simple to implement. 'safetensors' also enables cross language and cross frameworks compatibility making it an ideal format for storing machine learning model weights.
Authors: Daniel Falbel [aut, cre], Posit [cph]
Maintainer: Daniel Falbel <[email protected]>
License: MIT + file LICENSE
Version: 0.1.2.9000
Built: 2024-11-16 05:50:59 UTC
Source: https://github.com/mlverse/safetensors

Help Index

Safe load a safetensors file

Description

Loads an safetensors file from disk.

Usage

safe_load_file(path, ..., framework = "torch", device = "cpu")

Arguments

path

Path to the file to load
...

Unused
framework

Framework to load the data into. Currently only torch is supported
device

Device to copy data once loaded

Value

A list with tensors in the file. The metadata attribute can be used to find metadata the metadata header in the file.

See Also

safetensors, safe_save_file()

Examples

if (rlang::is_installed("torch") && torch::torch_is_installed()) {
  tensors <- list(x = torch::torch_randn(10, 10))
  temp <- tempfile()
  safe_save_file(tensors, temp)
  safe_load_file(temp)
}

Writes a list of tensors to the safetensors format

Description

Writes a list of tensors to the safetensors format

Usage

safe_save_file(tensors, path, ..., metadata = NULL)

safe_serialize(tensors, ..., metadata = NULL)

Arguments

tensors

A named list of tensors. Currently only torch tensors are supported.
path

The path to save the tensors to. It can also be a binary connection, as eg, created with file().
...

Currently unused.
metadata

An optional string that is added to the file header. Possibly adding additional description to the weights.

Value

The path invisibly or a raw vector.

Functions

  • safe_serialize(): Serializes the tensors and returns a raw vector.

Examples

if (rlang::is_installed("torch") && torch::torch_is_installed()) {
  tensors <- list(x = torch::torch_randn(10, 10))
  temp <- tempfile()
  safe_save_file(tensors, temp)
  safe_load_file(temp)

  ser <- safe_serialize(tensors)
}

Low level control over safetensors files

Description

Low level control over safetensors files

Low level control over safetensors files

Details

Allows opening a connection to a safetensors file and query the tensor names, metadata, etc. Opening a connection only reads the file metadata into memory. This allows for more fined grained control over reading.

Public fields

con

the connection object with the file

metadata

an R list containing the metadata header in the file

framework

the framework used to return the tensors

device

the device to where tensors are copied

max_offset

the largest offset boundary that was visited. Mainly used in torch to find the end of the safetensors file.

Methods

Public methods

Method new()

Opens the connection with the file

Usage
safetensors$new(path, ..., framework = "torch", device = "cpu")
Arguments
path

Path to the file to load

...

Unused

framework

Framework to load the data into. Currently only torch is supported

device

Device to copy data once loaded

Method keys()

Get the keys (tensor names) in the file

Usage
safetensors$keys()

Method get_tensor()

Get a tensor from its name

Usage
safetensors$get_tensor(name)
Arguments
name

Name of the tensor to load

Method clone()

The objects of this class are cloneable with this method.

Usage
safetensors$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

if (rlang::is_installed("torch") && torch::torch_is_installed()) {
tensors <- list(x = torch::torch_randn(10, 10))
temp <- tempfile()
safe_save_file(tensors, temp)
f <- safetensors$new(temp)
f$get_tensor("x")
}