Pickling tokenizer fails due to builtins.CoreBPE

[jerheff]

I am using tiktoken in a dataset preprocessing step for a pytorch DataLoader. They support multiprocessing in creating batches which spawns workers. This fails with exception:

TypeError: cannot pickle 'builtins.CoreBPE' object

I am not familiar with Rust, but this thread seems to suggest that a few methods in the Rust implementation would enable pickling the tokenizer.

[sk-g]

+1

Fails with this error when used in multiprocessing context, eg

    train_dataset = train_dataset.map(
        tokenization_function,
        batched=True,
        batch_size=1000,
        num_proc=args.num_proc,
        load_from_cache_file=not args.overwrite_cache,
        desc=f"Running tokenizer on train dataset with {len(train_dataset)} items"
        )

[sk-g]

Related issue: https://github.com/openai/tiktoken/pull/181

[hauntsaninja]

I allow Encoding to be pickled in tiktoken 0.6. Please let me know if the implementation doesn't work well for you!

[jerheff]

@hauntsaninja Works for me. Thanks!