docker_mlx_cpp — Give any Docker container Metal GPU access (107 ops)

[ilessiorobotflowlabs]

docker_mlx_cpp — The NVIDIA Container Toolkit for Mac

We built an open-source tool that gives any Docker container access to the full MLX compute stack on Apple Silicon.

The problem

Docker on Mac runs Linux VMs. Metal GPU can't be passed through. Every ML workload falls back to CPU inside containers.

The solution

A host-side MLX daemon with direct Metal GPU access, proxied to containers via HTTP:

Container → mlx-gateway:8080 → MLX Daemon:12435 → Metal GPU

What containers get

• 107 GPU operations (matmul, softmax, conv2d, attention, FFT, normalization, etc.)
• LLM inference (mlx-lm, 50+ architectures)
• VLM (mlx-vlm, image + text)
• Audio (mlx-audio: Whisper STT + Kokoro TTS)
• Image gen (mflux: FLUX)
• Embeddings (mlx-embeddings)
• Training (LoRA/QLoRA via mlx-lm tuner)
• OpenAI-compatible API — drop-in replacement

One-line install

curl -fsSL https://raw.githubusercontent.com/RobotFlow-Labs/docker_mlx_cpp/main/install.sh | bash

Benchmarks (M5, 24GB)

Operation	Performance
Matmul 1024x1024	~95 TFLOPS
Flash Attention	1.6ms
Conv2d	0.4ms
FFT2	0.5ms
Test suite	107/108 passing

Repo: https://github.com/RobotFlow-Labs/docker_mlx_cpp

MIT licensed. Contributions welcome. Would love feedback from the MLX community!