Skip to main content
Training guide chapters All documentation

Select a reference and a trainable student

Match provider access, student architecture, checkpoint revision and training method.

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.

Model compatibility checks

Check compatibility before allocating an expensive training run.
  1. Reference access Provider plan and available teacher
  2. Student Architecture and pinned weights
  3. Method SFT or compatible adapter training
  4. Runtime Precision, memory and dependencies

Select within your provider plan

Use reference identifiers from the active catalogue and your job access details. The session accepts an optional reference_model_id from the same paid provider and source product. It preserves the purchased student and invoice. Once preparation has generated data or begun attempts, the reference cannot be changed in that session; do not mix teacher data silently.

  • The reference must remain available and compatible with the purchased method and student.
  • Existing-customer-only restrictions remain in force.
  • A model display name is not necessarily its API identifier.
Example · python 
import os
from fusionpointai_training import TrainingClient

client = TrainingClient(api_key=os.environ["FUSIONPOINTAI_API_KEY"])
# Set this to a reference ID included in your paid provider plan.
reference_id = int(os.environ["FUSIONPOINTAI_REFERENCE_MODEL_ID"])
session = client.open_session(
    os.environ["FUSIONPOINTAI_JOB_UUID"],
    runtime="customer_datacenter",
    reference_model_id=reference_id,
)
# Change the reference only BEFORE preparation or generation attempts.
print(session["reference_model_id"])

Load a pinned local student model

TrainingProject accepts an existing local model path through the registered Transformers causal-language-model adapter. Model files are loaded with remote code disabled. Model architecture, tokenizer, context length, precision and available device memory are checked for the selected method before training. A repository identifier must resolve to a pinned revision inside the project root.

  • SFT updates eligible model parameters; LoRA and QLoRA train adapters.
  • A compatible tokenizer and model configuration must accompany local weights.
  • An inference-only API model is not a trainable student checkpoint.
Example · python 
# Continue with the client and session from the quickstart.
state = client.get_session(session["id"])
configuration = state["configuration"]
print(configuration["student_checkpoint"]["repository"])
print(configuration["student_checkpoint"]["revision"])
print(configuration["method"])
print(configuration["sdk_training"])
# These values come from the approved job, not local overrides.
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.

Check adapter and runtime compatibility

The registered Transformers causal-language-model adapter checks the actual model, method, precision, hardware, distribution and export combination. It does not support every architecture or GPU configuration. Unsupported combinations fail compatibility checks; no hardware or model capability is inferred from a configuration label.

  • Use only a model layout the Transformers adapter can load with remote code disabled.
  • QLoRA requires compatible CUDA and bitsandbytes support.
  • Compatibility checks are not a benchmark or a guarantee of successful inference on another runtime.