---

name: sync-and-converters description: "Guide SwanLab local sync, offline run validation, crash-tolerant upload behavior, and TensorBoard/W&B/MLflow conversion." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

SwanLab sync and converters

Use this sub-skill when the task is about uploading an existing SwanLab local/offline run, validating a run directory before upload, interpreting incomplete or legacy local records, or converting/synchronizing logs from TensorBoard, Weights & Biases, or MLflow into SwanLab.

Route here for

• swanlab.sync(...) or the swanlab sync CLI.
• Checking whether a local/offline run directory is readable and structurally plausible before cloud sync.
• Explaining corrupted-record, missing-finish, crash-marked, duplicate-step, or old run-file-version behavior during sync.
• Choosing among sync_wandb, sync_tensorboardX, sync_tensorboard_torch, sync_mlflow, WandbConverter, WandbLocalConverter, TFBConverter, MLFlowConverter, and the swanlab convert CLI.
• Troubleshooting converter optional dependencies, monkey-patch ordering, source path confusion, or upload failures.

Route elsewhere

• Creating the original SwanLab training run, logging scalars/config, or using finish: use the experiment-tracking sub-skill.
• Login, API-key storage, default mode selection, and host configuration: use the settings-and-modes sub-skill.
• Fetching/exporting cloud metadata with swanlab.Api or swanlab api: use the open-api-and-cli sub-skill.
• Media object constructors and rich chart object details: use the media-and-custom-charts sub-skill.

Operating references

1. Read references/sync-and-conversion.md for the sync/converter decision matrix, API and CLI shapes, local run expectations, and conversion path checklist.
2. Read references/troubleshooting.md when a user reports authentication, run-directory, record-integrity, old-version, optional-dependency, path-selection, or network upload failures.
3. Use scripts/check_sync_guards.py for no-network parser and path-guard assertions before trusting advice that depends on local sync/converter surfaces.

Safety defaults

• Do not start a network upload unless the user explicitly asks to sync and has provided or confirmed credentials/host context.
• Validate the directory first; for sync, the target is the run directory containing the SwanLab run record file, not a TensorBoard log root, W&B root, or MLflow tracking URI.
• Treat missing finish records as a crash-tolerant sync case, not as proof that all metrics are complete.
• Treat converter dependencies as optional until the selected converter path requires them.