#pickle #pytorch #python #machine-learning #serialization

repugnant-pickle

Low effort scraping for Python pickle format files, including some support for loading PyTorch checkpoints

1 unstable release

0.0.1 Jan 11, 2024

#7 in #pickle


Used in web-rwkv-converter

Apache-2.0

54KB
900 lines

repugnant-pickle

Because it is, isn't it?

This is a Rust crate for dealing with the Python pickle format.

It also has support for opening PyTorch files (.pth, .pt) which are basically ZIP files with a pickle hidden inside.

What?

The Python pickle format. There are people who get paid way more than me who spent massive amounts of effort working on copy protection schemes that were less effective at preventing interoperation than the Python pickle format.

Why would anyone actually make such a thing? My theory is that the originator has a deep and abiding hate for their fellow humans. The only thing that gives them any pleasure is to cause pain to as many people as possible. To that end: the Python pickle format.

It's basically impossible to handle reliably unless you're Python. In an ideal world, you'd just leave the pickles alone but sadly that is not the world we find ourselves living in. Sometimes you need to get some data out of a pickle file and you don't want to embed a whole Python interpreter into your application or something crazy like that.

That's where repugnant-pickle comes in. It does what it can and gives you what it came up with.

The name is a riff on BeautifulSoup, a HTML scraper library. It didn't (doesn't?) try to parse HTML completely (which is insanely hard), it just gives you a relatively simple way to to to extract the data you need. Most of the time that's good enough.

Example

Here's the kind of output you could expect to get from parsing a PyTorch file:

[Build(
  Global(
    Raw(GLOBAL("collections", "OrderedDict"), [
      Seq(Tuple, []),
      Seq(Tuple, [
        Seq(Dict, [
          String("emb.weight"),
          Global(Raw(GLOBAL("torch._utils", "_rebuild_tensor_v2")), [
            Seq(Tuple, [
              PersId(Seq(Tuple, [
                String("storage"),
                Raw(GLOBAL("torch", "BFloat16Storage")),
                String("0"),
                String("cuda:0"),
                Int(430348288),
              ])),
              Int(327378944),
              Seq(Tuple, [Int(50277), Int(1024)]),
              Seq(Tuple, [Int(1024), Int(1)]),
              Bool(false),
              Global(Raw(GLOBAL("collections", "OrderedDict")), [
                Seq(Tuple, [])
              ]),
            ]),
          ]),
        ]),
        Seq(Dict, [
          String("blocks.0.ln1.weight"),
          Global(Raw(GLOBAL("torch._utils", "_rebuild_tensor_v2")), [
            // Etc
          ]),
        ]),
      ]),
    ]),
  ),
)]

What about serde-pickle?

Use it if you can. It absolutely will be a much nicer experience than this crate. However, there are things it can't handle such as persistant IDs. PyTorch files use persistant IDs, so you don't really have a choice in that case.

There are also some odd choices like forcing the Rust toolchain to version 1.41.

PyTorch

You can enable support for attempting to deal with PyTorch files with the torch feature. If your Torch file has weird stuff, you may have to deal with it manually. Look in src/torch.rs for an example of where to start.

Otherwise, you can use repugnant_pickle::torch::RepugnantTorchTensors::new_from_file to load the tensor metadata. For example, if it works, you'll get something like this:

RepugnantTorchTensors(
   [
       RepugnantTorchTensor {
           name: "emb.weight",
           device: "cuda:0",
           tensor_type: BFloat16,
           storage: "archive/data/0",
           storage_len: 430348288,
           storage_offset: 327378944,
           absolute_offset: 327445248,
           shape: [50277, 1024],
           stride: [1024, 1],
           requires_grad: false,
       },
       RepugnantTorchTensor {
           name: "blocks.0.ln1.weight",
           device: "cuda:0",
           tensor_type: BFloat16,
           storage: "archive/data/0",
           storage_len: 430348288,
           storage_offset: 13639680,
           absolute_offset: 13705984,
           shape: [1024],
           stride: [1],
           requires_grad: false,
       },
   ]

You'll have to calculate the length yourself from the type and shape. For example, the first tensor has shape [1024, 50277] and type BFloat16. So the length in bytes would be (1024 * 50277) * 2 and the data would start at offset 327445248 in the Torch file, or if you access the specific file archive/data/0 in the Torch ZIP, it would start at 327378944.

It worked on the .pth and .pt LLM models I have.

Source: Trust me bro.

Usage

Look at the examples in examples:

  • dump_raw.rs — Dumps the raw Pickle opcodes from a file.
  • dump.rs — Dumps the Values from a file.
  • dump_torch.rs — Dumps the tensor metadata for a PyTorch file.

Note that for the last one you'll need to have the torch feature enabled. You can also run the example like:

cargo run --features torch --example dump_raw -- /path/to/torchfile.pth

Warning: Don't try to run dump or dump_raw on a Torch file. It'll try to load the entire massive file as a pickle.

Caveats

This isn't very well tested. It probably won't delete all your cat pictures when it fails, but it could fail.

Also, be careful of recursive Pickle files. It won't crash, but it'll loop until it hits the recursion limit (about 250) and you'll get nested data up to that limit.

If you need to handle large Pickle files, ideally don't use this at all, but if you must then you'd probably be better off using the single Pickle op Nom parser in src/parsers.rs.

Dependencies

~1.3–2.5MB
~45K SLoC