[SimpleFSDP] update simplefsdp readme (#1328)

Add notes to highlight users will need to pull latest pytorch to run
simplefsdp.

Author: ruisizhang123

torchtitan/experiments/simple_fsdp/README.md

@@ -3,6 +3,10 @@
 [![integration tests](https://github.com/pytorch/torchtitan/actions/workflows/integration_test_8gpu_simple_fsdp.yaml/badge.svg?branch=main)](https://github.com/pytorch/torchtitan/actions/workflows/integration_test_8gpu_simple_fsdp.yaml?query=branch%3Amain)
 [![arXiv](https://img.shields.io/badge/arXiv-2411.00284-b31b1b.svg)](https://arxiv.org/abs/2411.00284)
 
+💡 **Note**: SimpleFSDP's composability with Mixed Precision Training and Tensor Parallel requires updates from latest PyTorch, which can be installed (e.g., for CUDA 12.6) via
+```bash
+pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu126 --force-reinstall
+```
 
 This folder includes an experimental frontend implementation for [SimpleFSDP: Simpler Fully Sharded Data Parallel with torch.compile](https://arxiv.org/abs/2411.00284). SimpleFSDP is a compiler-based Fully Sharded Data Parallel (FSDP) framework, which has a simple implementation for maintenance and composability, allows full computation-communication graph tracing, and brings performance enhancement via compiler backend optimizations.