Skip to main content
Training guide chapters All documentation

Choose a runtime and export your model

Run on managed GPUs, your data center or a compatible notebook; understand exactly what is delivered.

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.

Choose where to train

The same SDK connects from each supported runtime; hardware is not created by selecting its name.
  1. Managed hosting Agreed FusionPointAI GPU environment
  2. Your data center Customer-controlled GPU and storage
  3. Notebook Compatible Colab-style runtime
  4. Local artifact Export and validate before inference

Choose a runtime you control

TrainingProject accepts fusionpointai_hosting, customer_datacenter or colab as runtime identifiers. The identifier does not create a GPU instance, grant notebook capacity or guarantee device compatibility. Install compatible dependencies, secure the key and check durable storage where local model training runs.

  • Notebook sessions may terminate and storage may be temporary.
  • Teacher API access requires network connectivity; local training uses the selected runtime.
  • Compute capacity and reference-model access are separate.
Example · text 
fusionpointai_hosting  -> agreed FusionPointAI compute environment
customer_datacenter   -> your own compatible compute environment
colab                 -> compatible notebook runtime

Choose the runtime already approved for your paid job.
Selecting a runtime does not allocate GPUs or notebook credits.
Example · bash 
# Run in a notebook terminal/shell after installing the SDK.
# Set FUSIONPOINTAI_API_KEY through the runtime's secret facility.
# The downloaded source includes the current Colab example.
curl --fail --silent --show-error \
  https://fusionpointai.com/downloads/fusionpointai-training.zip \
  --output fusionpointai-training.zip
unzip fusionpointai-training.zip -d fusionpointai-training-src
python fusionpointai-training-src/examples/colab_training.py \
  --job "$FUSIONPOINTAI_JOB_UUID" --output /content/customer-model-run-001
# Output must be new. Save your artifact before the notebook is terminated.
Example · python 
import os

supported_runtimes = {
    "fusionpointai_hosting",
    "customer_datacenter",
    "colab",
}
runtime = os.environ["FUSIONPOINTAI_TRAINING_RUNTIME"]
if runtime not in supported_runtimes:
    raise ValueError("Choose a supported training runtime")
print("Selected runtime:", runtime)
# Selection alone does not allocate compute or open a training session.
Example · python 
from importlib.metadata import PackageNotFoundError, version
from importlib.util import find_spec

for package in ("fusionpointai-training", "httpx", "torch", "transformers", "peft"):
    try:
        print(f"{package}=={version(package)}")
    except PackageNotFoundError:
        print(f"{package}: not installed")

for module in ("torch", "transformers", "peft", "safetensors"):
    print(f"{module} importable:", find_spec(module) is not None)
# The train extra includes optional model-training dependencies; inspect before use.

Inspect the local project artifact

A project run writes model or adapter files, tokenizer assets, training and evaluation reports, run metadata and an artifact manifest inside its owned output directory. Adapter exports require the exact base model and tokenizer. Export does not upload a local model archive or deployment package to the API.

  • Verify artifact paths, file sizes and SHA-256 values in the manifest.
  • Keep model and tokenizer revisions together.
  • Never include API keys, dataset secrets or unrelated files in a deliverable.
Example · bash 
cat ./support-project/runs/support-001/training-report.json
cat ./support-project/runs/support-001/evaluation-report.json
cat ./support-project/runs/support-001/artifact-manifest.json
sha256sum ./support-project/runs/support-001/exports/safetensors/artifact-manifest.json
# Compare the digest with the value returned by project.export().
# Adapter exports also require the matching base model and tokenizer.

Use the exported model independently

Load the exported model or adapter using a compatible inference runtime and run task acceptance checks before serving users. Normal inference need not call the teacher once the student is independently deployed. Export validates local files and writes reproducibility manifests; it does not deploy the model or guarantee that every format runs on every device.

  • Check tokenizer compatibility, precision and memory again at deployment.
  • Keep a rollback artifact and its evaluation report.
  • Do not overwrite a working deployed model before the replacement passes acceptance.