Write Once, Run Everywhere with Pytorch Transformers
summary
Hugging Face transformers is a pure PyTorch model-definition framework that now targets immediate downstream availability by going PyTorch-only, using a modular one-file model pattern and a single attention interface signature so alternative kernels (SDPA, flash attention, vLLM/SGLang) swap at runtime. Close integrations let vLLM and SGLang load a Transformers model in a few lines. A Transformers→MLX conversion skill is released that runs layer-by-layer logits comparisons, long-sequence generation and a non-agentic test harness; experiments show the approach works and speeds ports.
AI-written summary · about
chapters (9)
- 0:00Transformers as reference
- 1:27Readability and modularity
- 3:36Abstractions and attention
- 6:09Integrations with libraries
- 8:38Contract-based contributions
- 10:06Agents and open-source
- 12:13LLM release notes
- 14:42Transformers to MLX
- 17:57Future plans
related talks
01h 09mReinventing Machine Learning with Transformers and Hugging Face
25mTransformers: Standardizing Model Definitions Across the PyTorch Ecosystem
20mPyTorch x Transformers Journey
31mHugging Face ecosystem for Local AI/ ML
34mTransformers End to End Experience training...
40mPanel discussion "NLP"
26mRethinking the Transformer: Toward Native Multimodal Architectures
32mAn Introduction to Transfer Learning in NLP and HuggingFace