🗺️ KerasHub Roadmap 🗺️

[mattdangerw]

Here's an overview of the features we intend to work on in the near future. Feel free to comment on this issue below to suggest new directions for us to improve the library!

Models

We are always adding new models to the library. Here's what is currently on our radar.

• NLP models
  • Continue Gemma* releases.
    • As more models are released in the Gemma family of models, we will bring these to KerasHub on an ongoing basis.
  • DeepSeek R1 add DeepSeek R1 to KerasHub #2077
  • Llama 3.1, 3.2, 3.3 Add Llama 3.1, 3.2, 3.3 to KerasHub #2076
  • Qwen 2.5 Add Qwen 2.5 to KerasHub #2078
  • ModernBERT Add ModernBERT to KerasHub #2079
  • Mixtral.
  • all-MiniLM.
• RecSys models
  • BERT4Rec Add BERT4Rec to KerasHub #2080
  • SASRec Add SASRec to KerasHub #2081
• Vision models
  • LayoutLmV3.
  • DINOv2 Add DINOv2 to KerasHub #2082
  • ControlNet
• Audio models
  • Add a high-level Whisper speech to text task with generate() support.
  • Moonshine Add Moonshine to KerasHub #2083

Feature improvements

• Feature extractor task support.
• Weight file sharding for large (e.g. 10GB+) models Weight file sharding for large (e.g. 10GB+) models #2084
• DoRA Add DoRA support in KerasHub #2072
• Improved generation.
  • Move generation functionality to base classes #1861.
  • Directly use the backbone functional graph for CausalLM.generate() #1862.
  • Add support for JetStream generative inference for all KerasHub LLMs #1863.
• Improved quantization support (here and core Keras).
  • int4 support, QAT, more quantization options.
• Improved multi-host training support on Jax.
  • Add auto variable sharding for all backbones/tasks #1689.
  • Guide for multi-host distributed training with KerasHub #1850.
• An option for pythonic preprocessing decoupled from tf.data.
• Support RLHF and other instruction fine-tuning options beyond supervised fine-tuning. Support RLHF and other instruction fine-tuning options beyond supervised fine-tuning. #2073
• Add high level API for Whisper Streaming- realtime transcription Add a high-level speech to text task with generate() support. #2074
• Reducing inference latency using KVPress Reducing inference latency using KVPress #2075
• Speculative decoding

Integrations

• Continue to add conversion support for Huggingface Transformers and Timm checkpoints.
  • See this blog post.
• Support JetStream, Add support for JetStream generative inference for all KerasHub LLMs #1863.
  • Allow native, high-throughput Jax LLM inference on TPUs.

[rlcauvin]

May I suggest adding the Universal Sentence Encoder to KerasHub?