Date: 18.08.2025

TensorFlow-friendly causal LMs in 🤗 Transformers 4.x

This table lists decoder-only (causal LM) model families that have TensorFlow classes in Transformers 4.x and publish TensorFlow weights (tf_model.h5) on the Hugging Face Hub.
If tf_model.h5 is present for the checkpoint, you can load with TFAutoModelForCausalLM.from_pretrained(...) without converting from PyTorch.

Model family	Hugging Face repo	TF weights present? (Conversion needed?)
OpenAI GPT (GPT‑1)	openai-community/openai-gpt	Yes — tf_model.h5 available → No conversion needed
GPT‑2 (family)	openai-community/gpt2	Yes — tf_model.h5 available → No conversion needed
DistilGPT‑2	distilbert/distilgpt2	Yes — tf_model.h5 available → No conversion needed
DialoGPT (GPT‑2 based)	microsoft/DialoGPT-medium	Yes — tf_model.h5 available → No conversion needed
CTRL	Salesforce/ctrl	Yes — tf_model.h5 available → No conversion needed
Transformer‑XL	transfo-xl/transfo-xl-wt103	Yes — tf_model.h5 available → No conversion needed
XLNet	xlnet/xlnet-base-cased	Yes — tf_model.h5 available → No conversion needed
XLM (CLM variant)	FacebookAI/xlm-clm-ende-1024	Yes — tf_model.h5 available → No conversion needed
OPT (small)	facebook/opt-125m	Yes — tf_model.h5 available → No conversion needed
OPT (2.7B)	facebook/opt-2.7b	Yes — tf_model.h5 available → No conversion needed
GPT‑J‑6B	EleutherAI/gpt-j-6b	Yes — tf_model.h5 available → No conversion needed

Notes

• Transformers v4.x only. TensorFlow classes (e.g., TFAutoModelForCausalLM, TFGPT2LMHeadModel) are part of the 4.x line. In Transformers v5, TF support has been removed. Pin transformers<5 for TF usage.
• Per‑checkpoint variation. Some orgs publish TF weights for certain sizes but not others. Always check the target checkpoint’s Files & versions tab.
  • GPT‑2 medium/large/xl also provide TF weights (tf_model.h5).
• If there is no tf_model.h5 but the family has a TF class: you can often load from PyTorch with from_pt=True and then save_pretrained(...) to create a local tf_model.h5.
• Modern LLMs (e.g., LLaMA, Mistral, Mixtral, Gemma) generally do not ship TF classes in Transformers; they are PyTorch‑only in the library.
• Keras 3 compatibility: with TF ≥ 2.16 use tf-keras~=2.16 and set TF_USE_LEGACY_KERAS=1 before importing TensorFlow so transformers 4.x works smoothly.