YaFSDP is a high-performance Sharded Data Parallelism library for PyTorch, developed by Yandex. It provides two APIs: a high-level `YaFSDP` wrapper class and a lower-level `fully_shard` composable function, both designed to shard model parameters across data-parallel processes for training large transformer models. The library implements advanced techniques such as ragged sharding (via `RaggedShardDTensor`), efficient all-gather/reduce-scatter with optional YCCL acceleration, mixed precision, gradient accumulation, and hybrid parallel zones (HPZ). It is tightly integrated with PyTorch's distributed infrastructure and includes patches for Hugging Face libraries (transformers, accelerate, trl) to enable seamless usage in LLM training workflows. The codebase is well-structured, with clear separation of concerns across modules for parameter management, collectives, state, and tensor operations. It demonstrates high craftsmanship in its custom tensor subclass and communication optimizations, making it a valuable tool for large-scale distributed training.