Migrating from Weights & Biases
Coming from Weights & Biases.
EUVYRA covers the experiment tracking teams use W&B for and adds what happens after training: jobs on your own GPUs, serving behind an approval gate, and governance evidence for the AI Act. euvyra import wandb brings your runs along.
Weights & Biases today
- CoreWeave announced on 5 May 2025 that it had completed its acquisition of Weights & Biases.
- W&B's documentation lists three ways to run it: a fully managed service in W&B's cloud, a dedicated cloud on isolated infrastructure, and W&B Self-Managed on your own infrastructure.
- W&B Weave is W&B's product for tracing and evaluating LLM applications and agents.
If W&B works for you, these facts may not change anything. Teams usually look at EUVYRA when they want tracking, training compute, serving and AI governance in one installation that they run themselves.
Where your W&B work goes in EUVYRA
| In W&B | In EUVYRA |
|---|---|
wandb.init, wandb.log, config and summary | euvyra.init, run.log, run.config and run.summary, with the git commit, diff hash and system load recorded automatically |
| Media: images, audio, video, HTML, histograms, plots, tables, 3D objects, molecules | The same kinds, including GLB, OBJ, STL and PLY meshes, point clouds with boxes, and PDB, SDF, MOL2 and XYZ molecules |
wandb.watch | run.watch for gradient and parameter histograms of PyTorch models |
| Sweeps with grid, random and Bayesian search and Hyperband | The same parameter format and early termination settings; the Bayesian search uses a Tree-structured Parzen Estimator |
| Artifacts with versions, aliases, lineage and references | Versioned artifacts with aliases and lineage; references to S3, GCS, Azure, HTTPS and file systems with checksums verified on download |
| Reports with co-editing | Reports with real-time co-editing, charts, run tables, media and lineage graphs |
| Custom charts with Vega | Vega-Lite charts in chart views, reports and logged media |
| Launch | Jobs on Kubernetes or Docker GPU hosts through a pull agent, with queues, quotas, cost per run and multi-node training |
| Weave | Built-in tracing, evaluations with CI gates, prompts and annotation queues |
Offline mode and wandb sync | mode="offline" and euvyra sync |
What EUVYRA adds: an OpenAI-compatible gateway that serves only approved releases, approvals bound to people's functions, AI Act risk classification, controls checked against evidence, and incident clocks for the AI Act, NIS2, DORA and the GDPR. Platform overview
How to move
The importer runs on a machine that has W&B credentials. Data flows from W&B to that machine and from there to EUVYRA, so the EUVYRA server never contacts W&B and its outbound allowlist stays closed.
Install the SDK with the W&B extra and sign in
pip install "euvyra[wandb]" euvyra login --host https://euvyra.example.eu --api-key euv_...Import a project
Runs keep their configuration, summary, full metric history with steps and timestamps, tags, notes, group, job type, start time, runtime and final state. They are tagged
imported-from-wandb, and runs imported before are skipped, so you can run it again.euvyra import wandb acme/chest-xray --project chest-xrayPort your training code
The calls map almost one to one.
# before: wandb.init(project="chest-xray", config=cfg); wandb.log({"loss": loss}) import euvyra with euvyra.init(project="chest-xray", config=cfg) as run: run.watch(model, log="gradients") for epoch in range(cfg["epochs"]): run.log({"loss": train_one_epoch()}, step=epoch)Bring your sweeps
Sweep files in the W&B format work as they are.
euvyra sweep create sweep.yaml --project chest-xray euvyra agent swp_... --count 10 -- python train.py
What is different, and what does not carry over
- The importer copies runs and their metrics. Media, artifacts, tables, reports, sweeps and Weave traces stay in W&B; log again what you still need.
- Tables and confusion matrices exist, but there are no interactive queries over tables.
- The Bayesian sweep method uses a Tree-structured Parzen Estimator rather than a Gaussian process, so the same sweep can suggest different configurations.
- Moving a report block while someone types in it loses that person's edit.
Questions about moving from W&B
Do we need to keep W&B running during the import?
Only access to it: the importer reads through W&B's public API with your credentials. Runs imported before are skipped, so you can import again later to pick up new runs.
Can we use our framework integrations?
EUVYRA has its own callbacks for Hugging Face Transformers, PyTorch Lightning, Keras 3, XGBoost, LightGBM and scikit-learn, tested against the real libraries.
Can we try EUVYRA without moving everything?
Yes. Import one project, run both side by side for a while, and switch the logging calls when you are satisfied.