Skip to main content
Training guide chapters All documentation

SDK versions, compatibility and migration

Identify the installed release, preserve existing jobs and artifacts, and migrate safely to project configuration.

Training guide chapters

Commands and code stay in English in every language. Replace uppercase placeholders with values from your own account and paid job; do not paste real keys into shared notebooks or logs.

Move from sessions to projects

Keep existing jobs and artifacts intact when adopting project configuration.
  1. Identify Package and supported interfaces
  2. Preserve Environment, reports and artifacts
  3. Map Compatible settings and new plan
  4. Validate Reload, evaluation and cleanup
  5. Adopt Accepted project artifact

Know which version you are using

SDK 0.2.0 includes TrainingProject and fpai-training while retaining TrainingClient and its session methods for existing integrations. Configuration schema 1.0 is separate from the package release version. Check the installed package metadata before running commands.

  • Use the package matching the examples and verify its checksum.
  • Compatibility depends on the selected model, runtime, method and export format.
  • Existing paid jobs remain accessible through TrainingClient.
Example · python 
from importlib.metadata import version
print(version("fusionpointai-training"))
# The installed package version is distinct from configuration schema 1.0.

Migrate without changing customer entitlements

Keep the existing environment, paid job identifiers, reports and original artifacts. Create a separate project configuration and a new owned output directory. Map only supported settings after checking the pinned model, tokenizer, data and hardware. Existing invoices, provider access periods and reference-plan restrictions remain authoritative.

  • Legacy session reports remain runtime-reported; migration does not turn them into independently verified evidence.
  • A session checkpoint is not automatically a compatible project recovery checkpoint.
  • Do not delete the old environment or replace production inference until export/reload and task acceptance pass.
Example · text 
MIGRATION FROM SDK 0.1.0 TO SDK 0.2.0
Keep TrainingClient and paid session workflows for existing job/session reporting.
Preserve existing job IDs, configuration, runtime reports and local artifacts.
Create a separate schema 1.0 project and a new owned run directory.
Map only compatible local model, data, runtime and evaluation settings.
Keep prior runtime reports labelled runtime-reported, not independently verified.
Existing adapters -> require exact base model/revision and tokenizer.
Do not import a session checkpoint as project recovery state.
Retain the old environment until the new artifact reloads and passes task checks.
Do not replace production inference as part of migration.

Check package and runtime compatibility

SDK 0.2.0 includes the documented project interfaces, schemas and examples. The Transformers causal-language-model adapter checks compatibility among the selected model, training method, runtime, hardware and export format. These checks do not constitute hardware benchmarks or guarantee inference results.

  • Verify installed package metadata and checksum.
  • Review actual reports and manifests for the local run.
  • Treat example thresholds as configuration values, not measured outcomes.
Example · text 
SDK 0.2.0 PACKAGE INFORMATION
Check installed package metadata and compare it with the package checksum.
This guide documents the Python and CLI project interfaces and their schemas.
Hardware compatibility depends on the selected model and runtime.
No GPU benchmark or inference guarantee is asserted by these examples.