Trace JAX from Python to the device: op-by-op dispatch vs jax.jit, through jaxpr, StableHLO, the XLA compiler, LLVM or Triton, and the PJRT runtime, with sources.
Follow a local LLM through llama.cpp: GGUF file, ggml compute graph, and the CPU, CUDA or Metal backend, with the language each kernel is written in and sources.
Trace a PyTorch model from Python to the hardware: eager mode vs torch.compile, CPU vs GPU, through ATen, TorchDynamo, TorchInductor, Triton, LLVM IR and PTX, with sources.
The number of stages and steps on each selectable path, counted from the explainers' data files. The counts show how many documented layers a path passes through; they say nothing about speed. Each row opens that explainer on that path.
| Explainer | Selection | Stages | Steps |
|---|---|---|---|
| JAX | Without jit · CPU | 4 | 3 |
| JAX | Without jit · NVIDIA GPU | 4 | 3 |
| JAX | Without jit · TPU | 4 | 3 |
| JAX | jax.jit · CPU | 9 | 8 |
| JAX | jax.jit · NVIDIA GPU | 12 | 13 |
| JAX | jax.jit · TPU | 9 | 8 |
| llama.cpp | CPU | 6 | 5 |
| llama.cpp | CUDA (NVIDIA) | 10 | 9 |
| llama.cpp | Metal (Apple) | 9 | 8 |
| PyTorch | Eager · CPU | 5 | 4 |
| PyTorch | Eager · NVIDIA GPU | 5 | 4 |
| PyTorch | torch.compile · CPU | 8 | 8 |
| PyTorch | torch.compile · NVIDIA GPU | 12 | 12 |
Stages that are also nodes in the Language Lineage atlas, so you can see what each tool is written in. CUDA appears in every explainer.
Each explainer is one versioned JSON file (how-jax-runs.json, how-llama-cpp-runs.json, how-pytorch-runs.json) under the same CC BY 4.0 license as the dataset. They are kept apart from the language lineage graph on purpose: "lowers to" and "dispatches to" describe how code executes, not what a tool is written in. Closed systems are not guessed at; where the public documentation stops, the diagram stops too. Found a mistake? Open an issue with the better source.
Explore the full graph →