Lightning-AI/lightning-thunder

Health v2 · maintenance only

Adoption not part of the score

1469 stars · 119 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Lightning Thunder is a source-to-source deep learning compiler for PyTorch that optimizes models for training and inference. It provides a Pythonic IR, composable optimization transforms (fusion, quantization, distributed parallelism), and an extensible dispatch mechanism to fusers and optimized kernel libraries.

Use cases

• speed up PyTorch model training and inference
• fuse PyTorch ops into optimized kernels
• quantize models to FP8/FP4 precision
• convert single-GPU models to distributed training
• profile and inspect PyTorch computation graphs
• write custom Triton kernels for PyTorch models
• avoid graph breaks when tracing PyTorch programs

When to choose

• you want faster PyTorch execution without leaving Python
• you need composable, inspectable optimization recipes across model families
• you want to target new hardware like NVIDIA Blackwell with custom kernels
• you need distributed TP/PP/DP transforms applied programmatically

When to avoid

• you need a battle-tested production compiler with broad ecosystem support (torch.compile)
• your models rely on ops Thunder does not yet cover
• you are not using PyTorch

Facets

library · maturity active

compiler machine-learning deep-learning llm-inference llm-training gpu-computing benchmarking deep-learning machine-learning compilers gpu-computing large-language-models performance python cross-platform pytorch-compiler kernel-fusion quantization distributed-training triton cuda-graphs source-to-source torch-compile-alternative gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Lightning-AI/lightning-thunder")