Using MLX for distributed quantum simulation across Apple Silicon nodes

[deryakarl]

We’ve been developing Zilver, a distributed quantum simulation network that operates entirely on Apple Silicon. MLX is the main engine, capable of simulating state vectors up to 33 qubits on M-Ultra, along with density matrix and tensor network backends. It also supports batched circuit evaluation using vmap for variational algorithms.

We’d be grateful for your thoughts on two areas:

1. Memory management for long-running processes that handle sequential large tensor allocations. We’re currently using mx.set_memory_limit and mx.clear_cache() between jobs, but we’re wondering if there are more efficient ways to do this.

2. Distributed computation across multiple Apple Silicon devices. Do you have any experience or prior knowledge about MLX that we should be aware of?

If you are working on scientific computing or simulation with MLX, we would love to hear from you. You can find the code at https://github.com/Sirius-Quantum/zilver

Many thanks,
Derya