Model training units
This topic introduces Alibaba Cloud Bailian model training units (MTUs), including specifications, billing method selection, the subscription purchase process, and the submission configuration of RL training jobs.
Overview
A model training unit (MTU) is the unit of measure and sale for model training compute, available in six specifications from Type I to Type VI. Jobs billed by training unit run on exclusive training resources: training is faster and queue-free.
Training units are currently available for reinforcement learning (RL) training: RL training supports only training unit billing (billing by token is not supported). For the full workflow, see RL training overview.
Two billing methods involve training units:
- Model training unit · subscription: purchase training units monthly. The fee is paid once at purchase; training then draws down the purchased quota with no usage-based charges.
- Model training unit · pay-as-you-go: no pre-purchase required. Select the specification and instance count when you submit an RL training job, and pay for the actual training runtime.
Training units are an invitation-only feature. Contact your Alibaba Cloud account manager to activate it. For billing rules and prices, see Model training unit billing.
Billing method selection
Billing method | Suitable for | Description |
|---|---|---|
Model training unit · subscription | Long-term, stable RL training | Purchase training units in advance for the strongest resource certainty, with no usage-based charges during training. Purchase first. |
Model training unit · pay-as-you-go | On-demand RL training | No pre-purchase. Select the specification and instance count when you submit the job, billed by the actual training runtime. |
Training unit specifications
Training units come in six specifications. Each specification targets a different compute tier and set of supported models:
Specification | Description |
|---|---|
Type I training unit (MTU1) | Mainstream compute, suitable for fine-tuning small and mid-sized models and small-scale training experiments |
Type II training unit (MTU2) | High-performance compute, suitable for full-parameter training of large models |
Type III training unit (MTU3) | Ultra-large-scale compute, suitable for pre-training and full-parameter training of ultra-large models |
Type IV training unit (MTU4) | - |
Type V training unit (MTU5) | - |
Type VI training unit (MTU6) | - |
For subscription and pay-as-you-go prices of each specification, see Model training unit billing.
NoteThe set of models supported by each specification is subject to the console display. For RL training recommendations, see RL training overview. Contact your account manager for specific capacity calculations.
Billing and usage essentials
ImportantFor the complete billing rules, specification prices, and cancellation and refund policies, see Model training unit billing.
- The subscription fee is paid once at purchase; during training, usage draws down the purchased quota, and unused remaining duration expires automatically at expiration. Pay-as-you-go is billed only for the actual training runtime (minimum billing granularity: 1 minute). Job queueing, data preprocessing, and training artifact saving are not counted toward billable duration.
- Reserve duration for long-running jobs: the remaining subscription duration must exceed the estimated training duration. Purchase sufficient duration for your training needs to avoid jobs being suspended due to expiration.
- Start small: when trying a new model or configuration, first verify the pipeline with a small dataset and a small instance count, then switch to the full dataset and target instance count. This avoids wasting large-scale resources.
- Early cancellation is charged at 1.2 times the hourly unit price. When training units expire, running jobs are suspended and training checkpoints are saved, so training progress is not lost.
- Pay-as-you-go fees are deducted from your account balance, and overdue accounts suspend training jobs. Keep the balance sufficient.
Purchasing training units (subscription)
- Go to the Model Fine-tuning page.
- Click Resource pool management (subscription) in the upper-right corner to open the Training resource management drawer.
- Switch to the Purchase training units (subscription) tab, select a training unit specification, set the number of training units (instances) and the duration (months). The fee is displayed in real time.
- Click Buy now. The training units take effect immediately after payment.
Purchase and management rules:
- Each specification can be purchased only once within its validity period, and can be purchased again after it expires.
- Within the validity period, purchased training units can be used by multiple jobs in sequence. The total number of training units occupied by jobs running at the same time cannot exceed the number purchased.
- The instance ID (instance_id) of subscription training units cannot currently be queried through the console: it is required when you submit a subscription training job, so contact your account manager to obtain it. For scaling, renewal, cancellation, and other management operations, also contact your account manager.
Submitting an RL training job
For the full submission workflow of RL training jobs, see RL training overview. When using training unit billing, fill in resources in the submission configuration according to the billing method:
# Model training unit · subscription: instance_id is required (contact your account manager)
"resources": {
"charge_type": "mtu_prepaid",
"mtu_spec_code": "MTU4",
"mtu_capacity": 24,
"instance_id": "sfm_trainingunit_public_cn-xxxx"
}
# Model training unit · pay-as-you-go: no instance_id needed
"resources": {
"charge_type": "mtu_postpaid",
"mtu_spec_code": "MTU4",
"mtu_capacity": 24
}
Configuration notes:
- mtu_capacity (the number of training units occupied by this job) does not have to use up the purchased capacity, but it must follow the capacity step of the selected specification and not exceed the remaining instance capacity (subscription) or the available capacity (pay-as-you-go).
- For subscription jobs, the platform validates the instance validity, remaining duration, and capacity when you submit. If validation fails, the job cannot be submitted. Adjust according to the prompts or contact your account manager.
- After submission, you can view the billing method, priority, and status of the job in the job list on the Model Fine-tuning page. For training fee details, click Training fee records in the upper-right corner.
FAQ
Q: How do I activate training units?
A: Training units are an invitation-only feature. Contact your Alibaba Cloud account manager to activate it. After activation, you can purchase training units in the console or submit RL training jobs billed by training units.
Q: Subscription or pay-as-you-go?
A: See Billing method selection. For long-term, stable RL training, subscription is recommended (strongest resource certainty, with no usage-based charges during training). For on-demand training, pay-as-you-go is recommended (no upfront investment, billed by actual duration).
Q: What happens to running jobs when my subscription training units expire?
A: The jobs are suspended and training checkpoints are saved, so training progress is not lost. Contact your account manager to renew in a timely manner. For the expiration rules, see Subscription billing rules.
Q: Where can I find the instance ID of my subscription training units?
A: The instance ID cannot currently be queried through the console. Contact your account manager to obtain it.
Related documentation
- RL training overview — the full RL training workflow and recommended specifications and instance counts
- Model training unit billing — billing rules, specification prices, and refund policies
- Model tuning overview — overview of training billing methods and the job list
- Bill query and cost management — bills and usage summaries