[TorchComms + titan] Update ReadME for torchComms repo (#1992)

We already release TorchComms so we want to update the README. Also
there are some issues when it comes to PP degree larger than 2. We also
want to document it.

Author: fduwjj

torchtitan/experiments/torchcomms/README.md

@@ -4,14 +4,21 @@
 
 This folder provides a framework for composability testing with TorchComms and distributed training in TorchTitan. It enables flexible experimentation with distributed communication primitives and various parallelism strategies in PyTorch.
 
-> **TODO:** Additional documentation will be provided once TorchComms is publicly released.
+TorchComms repo: https://github.com/meta-pytorch/torchcomms
+
+Simple installation of TorchComms:
+```bash
+pip install --pre torch torchcomms --index-url https://download.pytorch.org/whl/nightly/cu128
+```
+
+If you want to compile torchcomms from source, please follow the instructions in the TorchComms repo.
 
 ### Quick Start
 
 The following command uses Llama 3 as an example:
 
 ```bash
-TEST_BACKEND=nccl TRAIN_FILE=torchtitan.experiments.torchcomms.train CONFIG_FILE="./torchtitan/models/llama3/train_configs/debug_model.toml" ./run_train.sh
+TEST_BACKEND=ncclx TRAIN_FILE=torchtitan.experiments.torchcomms.train CONFIG_FILE="./torchtitan/models/llama3/train_configs/llama3_8b.toml" ./run_train.sh
 ```
 
 ### Features
@@ -49,6 +56,7 @@ Locally tested with:
 ### Known Issues
 
 - **Memory Overhead** - TorchComms requires higher peak memory usage. As a workaround, we need to reduce `local_batch_size` to avoid out of memory error.
+- **Pipeline Parallelism** - Pipeline Parallelism is not supported yet when PP degree is larger than 2.
 
 ## Roadmap