# FSDP2 (fully_shard): Per-Parameter Sharding + DTensor + 2024+ PyTorch Yeniliği

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-fsdp2-per-parameter-sharding-dtensor
> Updated: 2026-08-14T02:47:16.250Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part IV — Mid-Large Models (13B-70B+) + Distributed Internals
**TLDR:** FSDP2 (PyTorch 2.4+) — eski FSDP'in evrimi. Per-parameter sharding (FlatParameter pattern'ı atıldı), DTensor backbone, FQN (Fully Qualified Names) ile resumable checkpointing, mixed precision daha rahat. Llama 3.3 70B + FSDP2 + DCP (Distributed Checkpoint) reçetesi.