👆 TouchNet [WIP]

A PyTorch native N-D parallel library for large-scale multimodal LLM (text/audio) training

• [2025/07/07] We support finetuning Qwen2-Audio-7B & Kimi-Audio-7B on ASR task! See WeneSpeech results for details.

👆 touchnet is highly motivated by torchtitan. Both of them are clean, minimal codebases for large-scale LLM training using native PyTorch. The main goal that differentiates 👆 touchnet from torchtitan is that 👆 touchnet focuses on multimodal LLM training where special data pipelines and model structures are needed. Please note that 👆 touchnet is currently in a pre-release state and under extensive development.

Our guiding principles when building 👆 touchnet are:

1. ⚡️ Blazing-fast checkpointable data loader with modular preprocessing and ​​fully random access​​ for large scale multimodal data
   • [New Storage Format] optimized for random access on sequentially saved tar files
   • Efficient [Sequence Packing] powered by [Flex Attention]
2. 🤗 Native integration with transformers models while get rid of structured trainer classes (e.g., [PyTorch-Lightning] or [HuggingFace Trainer])
   • Only reuse model definitions in transformers and leave other parts untouched
   • Entire training logic exposed in a single file [touchnet/bin/train.py], everything is under your control
3. 🛠️ Built-in profilers (CPU/GPU/memory) with flight recorder diagnostics.
   • [Nsys-like Profiler] to get optimization recommendations
   • [Memory Monitor] to debug OOM errors and improve memory usage
4. 🎯 N-D parallelism enabled through PyTorch native API and minimal lines of model code changes.
   • [FSDP2], why FSDP1 -> FSDP2?
   • [Tensor Parallel], [Context Parallel], [Pipeline Parallel] (PP WIP🚧), [Distributed Checkpoint]
5. ✨ Intuitive API design for rapid adoption & customization in minutes.
   • Supported tasks: [text/pretrain], [audio/pretrain], [audio/sft/asr], more tasks coming soon
   • Supported models: [Llama], [LlamaForASR] more models coming soon

Here is an end-to-end workflow for a traning job in 👆 TouchNet:

1. stage-1: Download dataset. We use load_dataset API in HuggingFace.datasets to download specific data.
2. stage-2: Convert dataset format to TouchDataset. see [touchnet/bin/make_data.py]
3. stage-3: (optional) Convert hf-format ckpt to torch distributed ckpt. see [touchnet/bin/convert_hf_to_dcp.py]
4. stage-4: Start training, either from scratch or from pretrained ckpt that has been converted in stage-3. see [touchnet/bin/train.py]
5. stage-5: Convert torch distributed ckpt to hf-format, enjoy HuggingFace ecosystem for inference and deployment. see [touchnet/bin/convert_dcp_to_hf.py]

For a more concrete example running those stages one by one, see [examples/audio/sft/asr/aishell/run.sh]

# NOTE(xcsong): Ensure that the linux system's glibc version is greater than or equal to 2.17 (see `ldd --version`)
#               (for example, Ubuntu 22.04 and later versions).
conda create -n touchnet python=3.10
conda activate touchnet
conda install -c conda-forge sox ffmpeg -y

# (Optional) install CUDA + cuDNN if they are not already available; change `prefix` to your install path.
# bash install_cuda_cudnn.sh

# Install TouchNet with GPU support (CUDA 12.6 - recommended)
pip install -e . --index-url https://download.pytorch.org/whl/cu126

# Or install with CUDA 11.8 support
# pip install -e . --index-url https://download.pytorch.org/whl/cu118

# For development with GPU support
# pip install -e '.[dev]' --index-url https://download.pytorch.org/whl/cu126

@misc{touchnet,
  title={TouchNet: A PyTorch native N-D parallel library for large-scale multimodal LLM (text/audio) training},
  author={Xingchen Song},
  year={2025},
  url={https://github.com/xingchensong/TouchNet},
}

1. This repo is highly motivated by torchtitan and we borrowed a lot of code from it.
2. This repo also benefits from Megatron-LM, WeNet, flame.

Thanks for their wonderful works.