# emlx_axon v0.5.0 - Table of Contents

> Axon model rewrites to swap supported nodes for EMLX.Fast Metal shaders

## Pages

- [EMLXAxon](readme.md)

## Modules

- [EMLXAxon](EMLXAxon.md): Axon model rewrites that swap supported nodes to `EMLX.Fast` Metal shaders.
- [EMLXAxon.Application](EMLXAxon.Application.md): OTP Application for EMLXAxon.
- [EMLXAxon.Llama.DenseLoader](EMLXAxon.Llama.DenseLoader.md): Builds native Llama `%EMLXAxon.Llama.Model.State{}` structs from standard
Bumblebee Llama dense parameters or Hugging Face safetensors files.
- [EMLXAxon.Llama.Generate](EMLXAxon.Llama.Generate.md): Autoregressive token generation loop.
- [EMLXAxon.Llama.Model](EMLXAxon.Llama.Model.md): Llama model state struct and forward pass for native dense weights.
- [EMLXAxon.Llama.Model.State](EMLXAxon.Llama.Model.State.md): Loaded model weights and config.
- [EMLXAxon.Llama.Native](EMLXAxon.Llama.Native.md): Compiler compatible dense Llama operations backed by the EMLX native plugin.
- [EMLXAxon.Llama.Sampler](EMLXAxon.Llama.Sampler.md): Three sampling strategies for autoregressive token generation.
- [EMLXAxon.MLX4BitParams](EMLXAxon.MLX4BitParams.md): Loads Qwen3 weights from an MLX-4bit safetensors checkpoint into Bumblebee
Axon params format (BF16, Bumblebee `{in, out}` key convention).
- [EMLXAxon.QuantizeParams](EMLXAxon.QuantizeParams.md): Post-load param quantization for Bumblebee models.
- [EMLXAxon.Qwen3.Attention](EMLXAxon.Qwen3.Attention.md): Grouped-query attention (GQA) for Qwen3, with a preallocated KV cache.
- [EMLXAxon.Qwen3.DenseLoader](EMLXAxon.Qwen3.DenseLoader.md): Builds native Qwen3 `%EMLXAxon.Qwen3.Model.State{}` structs from standard
Bumblebee Qwen3 dense parameters or Hugging Face safetensors files.
- [EMLXAxon.Qwen3.Generate](EMLXAxon.Qwen3.Generate.md): Autoregressive token generation loop.
- [EMLXAxon.Qwen3.Layers](EMLXAxon.Qwen3.Layers.md): Stateless layer primitives: RMSNorm, SwiGLU.
- [EMLXAxon.Qwen3.Loader](EMLXAxon.Qwen3.Loader.md): Loads a `lmstudio-community/Qwen3-*-MLX-4bit` checkpoint from disk into
an `%EMLXAxon.Qwen3.Model.State{}` struct.
- [EMLXAxon.Qwen3.Model](EMLXAxon.Qwen3.Model.md): Qwen3 model state struct and forward pass for quantized and dense weights.
- [EMLXAxon.Qwen3.Model.State](EMLXAxon.Qwen3.Model.State.md): Loaded model weights and config.
- [EMLXAxon.Qwen3.Native](EMLXAxon.Qwen3.Native.md): Compiler compatible Qwen3 operations backed by the EMLX native plugin.
- [EMLXAxon.Qwen3.Sampler](EMLXAxon.Qwen3.Sampler.md): Three sampling strategies for autoregressive token generation.
- [EMLXAxon.TextGeneration](EMLXAxon.TextGeneration.md): A `Nx.Serving`-compatible wrapper around native EMLXAxon text generation.