KO
|
EN
gitlite β search
Search
#typescript
#ai-agents
#ai
#dsh-plugin
#deepseek-harness
#open-source
#cli
#claude-code
#codex
#developer-tools
#react
#windows
ezpz
β 39
Open GitHub β
Write once, run anywhere; ezpz π
Download README (.md)
Explore Similar Repositories
fms-hf-tuning
:
π Collection of tuning recipes with HuggingFace SFTTrainer and PyTorch FSDP.
transformer_central
:
Various transformers for FSDP research
llm_parallelisms.c
:
LLM training parallelisms (DP, FSDP, TP, PP) in pure C
Mochi-Full-Finetuner
:
Code for full fintuing Mochi model with FSDP (and CP)
gradient-garden
:
Research platform for model training, evaluation, and experimentation across architectures, benchmarks, and recipes.
// repository documentation
Was this content helpful?
β 0
(0 ratings)
Select Rating:
β
β
β
β
β
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
ezpz
?
Download (.md)
# π ezpz > Write once, run anywhere. `ezpz` makes distributed PyTorch launches portable across any supported hardware {NVIDIA, AMD, Intel, MPS, CPU} with **zero code changes**. ## Features - **Multi-hardware** β automatic device detection and backend selection (CUDA/NCCL, XPU/CCL, MPS, CPU/Gloo) - **Zero code changes** β same script runs on a laptop, a single GPU, or a thousand-node supercomputer - **HPC integration** β native PBS and SLURM support with automatic hostfile discovery and rank assignment - **Metric tracking** β built-in `History` class for recording, plotting, and saving training metrics - **CLI tools** β `ezpz launch`, `ezpz test`, `ezpz doctor` for launching jobs, smoke-testing, and diagnostics ## Quick Install ```bash uv pip install git+https://github.com/saforem2/ezpz ``` ## Quick Start ```python import torch import ezpz rank = ezpz.setup_torch() # auto-detects device + backend device = ezpz.get_torch_device() model = torch.nn.Linear(128, 10).to(device) model = ezpz.wrap_model(model) # FSDP (default) # Multi-dim parallelism (TP/PP/CP) on XPU? Use ezpz.init_device_mesh_safe # instead of torch's init_device_mesh β works around xccl's missing # split_group on Aurora/Sunspot. See https://ezpz.cool/troubleshooting/. ``` ```bash # Same command everywhere -- Mac laptop, NVIDIA cluster, Intel Aurora: ezpz launch python3 train.py ``` For a side-by-side diff against the equivalent raw-torch boilerplate, see the [API Cheat Sheet](https://ezpz.cool/quickstart/#api-cheat-sheet). ## CLI ```bash ezpz launch python3 train.py # launch distributed training ezpz submit -N 2 -q debug -- python3 train.py # submit batch job to PBS/SLURM ezpz test # smoke-test your setup ezpz benchmark # run + compare example benchmarks ezpz doctor # diagnose environment issues ``` ## Why ezpz? Compared to the alternatives: - **vs raw `torchrun` / `mpirun` / `srun`**: one launcher that detects your scheduler, builds the right command, and works on a laptop too. - **vs `accelerate`**: lower surface area, no config files, designed for HPC schedulers from the ground up rather than retrofitted. - **vs `DeepSpeed`**: not an alternative β `ezpz` wraps your distributed init so you can still use DeepSpeed (or anything else) underneath. See the [full comparison](https://ezpz.cool/compare/) for details. ## Documentation Full documentation is available at [**ezpz.cool**](https://ezpz.cool). Useful entry points: - πββοΈ [Quickstart](https://ezpz.cool/quickstart/) β install β script β launch in 5 minutes - π [Distributed Training Tutorial](https://ezpz.cool/guides/distributed-training/) β progressive hello-world β FSDP+TP - π³ [Recipes](https://ezpz.cool/recipes/) β copy-pasteable patterns (checkpointing, gradient accumulation, MFU tracking) - π§ [Troubleshooting](https://ezpz.cool/troubleshooting/) β XPU FSDP2 hangs, NCCL/CCL errors, scheduler issues - π [Examples](https://ezpz.cool/examples/) β runnable end-to-end (FSDP, ViT, Diffusion, HF Trainer)
