QDMI Adapter in Arvak, a Rust-native quantum compilation stack

[hiq-lab]

Servus and hello!

I have recently built a rust-native quantum compiler called Arvak and added a QDMI-Adapter to it to make it more useful and learn about MQSS. You can find the readme for the adapter here: https://github.com/hiq-lab/arvak/blob/main/adapters/arvak-adapter-qdmi/README.md

It mostly serves to test my idea of a HAL contract and Arvak is an example implementation. I would be extremely grateful if you could take a minute to look at it and tell me if the adapter makes sense to you and if you have any suggestestions.

Some links for context:
Project Page:
https://github.com/hiq-lab/arvak

General concept for HPC deployment:
https://github.com/hiq-lab/arvak/blob/main/docs/hpc-deployment.md

Website:
https://arvak.io

About the HAL Contract Idea in gerneral:
https://hal-contract.org/

Any help is welcome! Thanks a lot in advance!

best regards and Servus again :)

Daniel

[hiq-lab]

I made a mistake interpreting the MQSS structure, have updated QDMI integration now, kindly find a new version of the readme here: https://github.com/hiq-lab/arvak/blob/main/adapters/arvak-adapter-qdmi/README.md

[hiq-lab]

We’ve added an evaluator module to Arvak that provides structured observability over the quantum compilation pipeline. Key capabilities:
∙ QDMI Contract Checking – every gate classified as Safe/Conditional/Violating against device capabilities
∙ Emitter Compliance – native gate coverage analysis for IQM (PRX/CZ), IBM (SX/RZ/CX), and CUDA-Q targets with decomposition cost documentation
∙ HPC Scheduler Fitness – LRZ (qc_iqm) and LUMI (q_fiqci) walltime estimation, batch capacity, and fitness scoring
∙ Orchestration Analysis – hybrid quantum-classical DAG construction, critical path, and batchability

All output is structured JSON (schema v0.3.0) suitable for CI/CD integration and automated pre-screening before SLURM submission.

Design document with QDMI alignment details and integration notes: https://github.com/hiq-lab/arvak/blob/main/docs/arvak-eval-design.md

Here is a direct link to live the dashboard display of the evaluator module: https://arvak.io/app/#eval

Looking forward to your feedback!

[burgholzer]

Servus 👋🏼

thanks for creating this discussion!
While I did not have enough time to look through everything, I had a quick look at the library.
It is fairly interesting what one can build with AI these days.

A couple of bigger, higher-level observations:

• the library currently always assumes OpenQASM 3 as a circuit format. Not every QDMI device may support this.
• querying of device capabilities seems very limited so far. not even the coupling map is queried (correctly). neither is the native gate set or properties like T1, T2 times, gate durations, fidelities, etc.
• the library currently builds/links against the client interface of QDMI. It also blindly chooses the first available device. I am pretty sure that in most cases you would get access to a QDMI device implementation and not the driver implementation that is used, e.g., in an HPC center. So it would probably be better to directly consume the QDMI device interface. This makes things slightly more complicated, as the device headers are name shifted by a prefix, which makes code like https://github.com/munich-quantum-toolkit/core/blob/03e1f998fe1af08a44a747919ef8cb5ead7a390e/src/qdmi/driver/Driver.cpp#L88-L167 necessary. It would make for a cleaner integration though and would also give you access to device session parameters like BASE_URL.
• Tests for the "real" QDMI integration would probably help quite a bit. You could either use the example device library in this repository here, or one of the libraries from the MQT Core project (https://github.com/munich-quantum-toolkit/core/tree/main/src/qdmi/devices)

Hope that's enough feedback/input for now 😌

[hiq-lab]

Servus Lukas!

Thank you very much for your feedback! I wish I could blame it all on AI! Anyway, your pointers were very helpful and I am a little embarrassed by some of the obvious points I missed. I put in a lot of effort to follow them and build a new, accurate representation of QDMI that actually implements the standard, so I can focus on my real goal: the rust-native quantum compiler and the connectivity towards Qiskit et al. Please take a look, if you can find the time. Your feedback is invaluable to me.

So here is a report on how I have made use of your previous comment:
https://github.com/hiq-lab/arvak/blob/main/docs/internal/qdmi-device-interface-audit.md

Also I have integrated DDSIM compatibility as a standard test to my nightly CI pipeline and it just cleared it! (Caveat: In contrast to the audit doc, the current CI pipeline omits the DDSIM latest test. I will add that later on again to catch drifts in the versions. It slowed the process too much for now.)

If you (or anyone else reading this) would like to test any of the demos for VQE and more to see how fast the compiler is, you can request a key to a locked-down SSH demo here: https://arvak.io/demos/

A live shootout between compilers can be found here: https://arvak.io/demos/benchmark/

Best regards and thanks again!

Daniel

PS: the evaluator module: https://arvak.io/app/#eval is fun to play with!

[hiq-lab]

the latest mqt-core v3.4.1 is reflected in the SHIM now and the updated test passes too.