Can TensorFlow keep up with the Generative AI and LLM trend in the coming years?

All answers to this question.

While it's true that many researchers use PyTorch for the initial LLM release, TensorFlow is where those models often go to be served at scale. KerasNLP and the newer Generative AI tools from Google are making it much easier to work with Transformers. Also, if you're using Google Cloud's Vertex AI, the integration with TensorFlow for fine-tuning models like BERT or PaLM-style architectures is extremely efficient. The framework is evolving to be more "module-heavy" to handle these large architectures better.

   Answered 2025-01-20 by Sandra Little

Does the ecosystem support PEFT (Parameter-Efficient Fine-Tuning) as well as the Hugging Face libraries do? That's the main concern for me.

   Answered 2025-01-23 by Jason Simmons

• Jason, Hugging Face actually has a TFTrainer and supports TensorFlow for many of their models. While the documentation for PEFT is more centered on PyTorch, the TF community is catching up rapidly. You can definitely implement LoRA and other techniques in TF; it just requires a bit more manual setup in the custom training loops for now.

     Commented 2025-01-25 by Ronald Bryant

TensorFlow is still very relevant here because of its ability to handle distributed training across TPU pods, which is essential for training LLMs from scratch.

   Answered 2025-01-28 by Thomas Jenkins

• True, Thomas. For anyone with access to TPU resources, TensorFlow is usually the most optimized path for training those massive parameter counts.

     Commented 2025-01-30 by Sandra Little