Tricks from OpenAI gpt-oss YOU 🫵 can use with transformers

OpenAI's gpt-oss model series has prompted significant upgrades to the Hugging Face `transformers` library, making it more efficient to load, run, and fine-tune models. These enhancements include zero-build kernels, MXFP4 quantization, tensor and expert parallelism, and continuous batching with paged attention. A key feature is the zero-build kernel system, which downloads pre-built, specialized programs for tasks like RMSNorm and MoE layers directly from the Hub, avoiding complex local compilation. Users can enable these performance optimizations by passing a simple flag during model instantiation, and these improvements are designed to benefit many other models within the ecosystem.