Skip to content

Latest commit

 

History

History

docs

README.md

MAX documentation

This directory contains the bulk of documentation about MAX, both for developers using MAX from the max package and for developers building with and contributing to the MAX framework through this repo.

Public website docs

Although the website frontend isn't included in this repo, you can find all the Markdown (MDX) files here:

Framework developer docs

The rest of the docs in this directory are for you if you've cloned this repo and are developing in the MAX framework. For example, if you're extending an existing model architectures, contributing new models, benchmarking, profiling, and debugging inside the MAX framework, then check out the following:

  • contributing/—MAX contributor handbook: Which areas of MAX accept contributions, how to get a change reviewed and merged, and what we expect from contributors.

  • design-docs/—Design docs: Engineering docs that describe how core Modular technologies work.

  • development.md—MAX framework developer guide: The getting started guide for development in the MAX framework codebase, introducing the Bazel build system.

  • contributing-models.md—Contributing new model architectures: How to add a new model architecture (directory layout, arch.py, model.py, config, weight adapters) and register it for max serve.

  • accuracy-debugging.md—Debug MAX model accuracy: Compare intermediate tensor outputs between a MAX pipeline and the PyTorch (Hugging Face) reference model using debug_model and compare_tensors to find numerical divergence.

  • max-benchmarking.md—Benchmarking a MAX endpoint: Using benchmark_serving.py to measure throughput, latency, and resource use for an OpenAI-compatible MAX model server.

  • kernel-benchmarking.md—Benchmarking Mojo kernels with kbench: Python toolkit for building and running Mojo kernel benchmarks, autotuning parameters, and analyzing performance across parameter grids.

  • kernel-profiling.md—Kernel profiling with Nsight Compute: Profile individual kernel performance on NVIDIA GPUs, install Nsight Compute, build with debug info, and run ncu to generate reports.

  • op-logging.md—Op logging in MAX: Enable op-level tracing to inspect operation launch and completion for debugging and performance analysis.

Other docs