v0.15.0

Summary

📢 KerasNLP is becoming KerasHub 📢, read more about it here.

This release contains a number of feature improvements:

• Added int8 quantization support.
  • Use the quantize() method to quantize any model.
  • Llama 2 and Llama 3 pre-quantized presets are available.
• PaliGemmaCausalLM will automatically resize input images during preprocessing.
• Added more converters for hugginface/transformers checkpoints.
  • Gemma 2, PaliGemma, GPT2, Bert, Albert, DistilBert, Bart.
• Class detection for huggingface/transformers checkpoints.
  • Call from_preset() on a base class, and we will find the correct subclass to create.
• Added Vicuna presets.
• Alias Classifier as TextClassifier, BertClassifier as BertTextClassifier.
• Added tokenizer.special_tokens and tokenizer.special_token_ids as convenient properties to view all special tokens on a pretrained tokenizer.

# Quantize an unquantized model.
lm = keras_nlp.models.CausalLM.from_preset(
    "gemma2_instruct_2b_en",
    dtype="bfloat16",
)
lm.quantize("int8")
# Load a pre-quantized model.
lm = keras_nlp.models.CausalLM.from_preset(
    "llama3_instruct_8b_en_int8",
    dtype="bfloat16",
)
# Convert a bert model in the huggingface/transformers format.
classifier = keras_nlp.models.TextClassifier.from_preset(
    "hf://google-bert/bert-base-uncased",
    num_classes=2,
)
# View all special tokens.
print(classifier.preprocessor.tokenizer.special_tokens)
print(classifier.preprocessor.tokenizer.special_token_ids)

Breaking changes

• On all backends, all strings and ragged output will be returned as python strings or python lists respectively.
  • This include preprocessing methods like tokenize() and detokenize().
  • This may break code that depended on tf.Tensor output on the tensorflow backend, but will lead to consistent output on all backends, which we believe will be an overall improvement.
  • Preprocessing layers can still always be included in a tf.data preprocessing pipeline, on any backend.

What's Changed

• Version bump to 0.14.0.dev0 by @grasskin in #1675
• Revert "Version bump to 0.14.0.dev0" by @grasskin in #1676
• Remove Keras pin, fix tests by @mattdangerw in #1681
• Add quantization support for Gemma, Gemma2 and PaliGemma by @james77777778 in #1670
• add vicuna preset by @sineeli in #1672
• Porting Gemma 2 transformers checkpoint by @ariG23498 in #1678
• Improve CI speed and resolve issues of run_quantization_check by @james77777778 in #1682
• Remove build_from_signature from MHA layers by @mattdangerw in #1687
• Refactoring: in CachedMultiHeadAttention call MHA methods instead of recoding the attention calculation by @apehex in #1684
• Porting PaliGemma transformers checkpoint by @ariG23498 in #1686
• Allow importing keras_nlp without tensorflow by @mattdangerw in #1660
• Add flag to gemma conversion script to specify local orbax by @mattdangerw in #1688
• Fix compatibility for earlier versions of Keras by @james77777778 in #1690
• Add a test against keras-nightly by @mattdangerw in #1693
• Fix dtype bugs in ReversibleEmbedding and LayerNorm by @james77777778 in #1692
• Partially revert #1687 by @mattdangerw in #1695
• Fix quantization test for XLNet by @james77777778 in #1699
• Add a HF BERT converter, improve safetensor loading by @mattdangerw in #1694
• Add a subtle fix for gemma 2 conversions by @mattdangerw in #1701
• One more small Gemma conversion fix by @mattdangerw in #1702
• Slightly more defensive handling of type for backbone by @mattdangerw in #1703
• Add support for converting Gemma 2 checkpoints by @mattdangerw in #1700
• Make it clearer what is running in the github action UI by @mattdangerw in #1707
• Try upgrading tensorflow pin by @mattdangerw in #1706
• Bump version to fix query norm in Gemma 2 9b by @mattdangerw in #1709
• Gemma: Add logit soft-capping to score function. by @RyanMullins in #1712
• Version bump HEAD to 0.15 by @mattdangerw in #1713
• Port gpt2 transformers checkpoint by @cosmo3769 in #1704
• Add soft capping to reversible embedding layer by @mattdangerw in #1718
• Add presets for gemma 2 2b by @mattdangerw in #1721
• Utilize to_numpy=True in quantize if available by @james77777778 in #1725
• Dynamic int8 quantization for Llama2 and Llama3 by @james77777778 in #1720
• Bump the python group with 2 updates by @dependabot in #1726
• Shield gemma shortnames by @mattdangerw in #1731
• Sliding window fixes by @mattdangerw in #1738
• Add int8 models to Llama2 and Llama3 by @james77777778 in #1734
• Port distilbert transformer checkpoint by @cosmo3769 in #1736
• Add support of kwargs to Backbone.from_preset and fix the dtype forwarding in Task.from_preset by @james77777778 in #1742
• Remove src init file contents by @mattdangerw in #1743
• Remove ROADMAP.md by @mattdangerw in #1773
• Fix nested list in args on keras.io by @mattdangerw in #1772
• Remove stale tf only examples by @mattdangerw in #1771
• Limit the default sequence length to 1024 for all models by @mattdangerw in #1770
• Consistent preprocessing output on all backends by @mattdangerw in #1777
• Port albert transformer checkpoint by @cosmo3769 in #1767
• Lower the default learning rate for albert by @mattdangerw in #1786
• Port bart transformer checkpoint by @cosmo3769 in #1783
• Add an option to disable default compilation by @mattdangerw in #1787
• Port mistral transformer checkpoint by @cosmo3769 in #1768
• [Bart]Fix missing weight port by @cosmo3769 in #1789
• Remove python 3.8 version in setup.py by @mattdangerw in #1792
• Class detection works for huggingface checkpoints by @mattdangerw in #1800
• Rename KerasNLP symbols for a multi-modal future by @mattdangerw in #1803
• Move preprocessing to base classes by @mattdangerw in #1807
• Add add_bos=False, add_eos=False to SentencePieceTokenizer.init() by @briango28 in #1811
• Only load a full task config when load_task_extras is passed by @mattdangerw in #1812
• Add image and audio converter classes by @mattdangerw in #1813
• Simplify registering "built-in" presets by @mattdangerw in #1818
• Support image and audio information in task summaries by @mattdangerw in #1819
• Take two of #1812, simpler classifier head loading by @mattdangerw in #1823
• Remove preprocessing layers we no longer use by @mattdangerw in #1824
• Version bump for dev release by @mattdangerw in #1825
• Version bump for dev release by @mattdangerw in #1830
• Version bump for 0.15.0 release by @mattdangerw in #1832

New Contributors

• @apehex made their first contribution in #1684
• @cosmo3769 made their first contribution in #1704

Full Changelog: v0.14.4...v0.15.0