ctranslate2-rs

Detect CUDA root directory automatically #55

Upgrade CTranslate2 to v4.3.0

• Fix: add Sync + Send to auto::Tokenizer

Auto tokenizer #49

Inspired by Hugging Face's Transformers' Auto Classes, this feature simplifies the tokenizer selection process by automatically choosing the correct tokenizer for a given model.

Stream API #52

Added a callback-based streaming API.

The following example demonstrates how to translate source and print results step-by-step:

let _ = translator.translate_batch(
    &vec![source],
    &TranslationOptions {
        // beam_size must be 1 to use the stream API.
        beam_size: 1,
        ..Default::default()
    },
    // Each time a new token is generated, the following callback closure is called.
    // In this example, it writes to the standard output sequentially.
    Some(&mut |r: GenerationStepResult| -> Result<()> {
        write!(out, "{}", r.text)?;
        out.flush()?;
        Ok(())
    }),
)?;

See the whole example code for more information.

Add config functions #47

Add Sent+Sync to Translator and Generator #45
Allow multiple backends #46

Translator and Generator support callback #44

Add support for Sentencepiece tokenizer #41

Add support for CUDA and cuDNN #37

Upgrade CTranslate2 to v4.2.1 #36

Windows support #35

Add support for Ruy #32

Add support for Intel MKL #30

Upgrade CTranslate2 to v4.2.0 #28

• fix build with Apple Silicon #26

Upgrade CTranslate2 to v3.24.0

Updates

• Upgrade CTranslate2 to v3.17.1

Updates

• Implement generator #12
• Add a generator example #13

• Translator::new takes device and config #9
• Translator::translate_batch takes options #10

• bump openssl from 0.10.54 to 0.10.55 #8