105. Design a Distributed Training Platform

The brief

Design a Distributed Training Platform. Train 1 B to 400 B models on 8 to 10,240 H100s: DDP, ZeRO and FSDP, tensor and pipeline parallel over NVLink and EFA, checkpoints, hot spares, MFU and goodput. Work from the scoping questions below. State assumptions for any unspecified load, guarantee or target, then trace your design end to end. Explain one difficult case and a credible alternative; the worked example is a reference, not a required implementation.

• Set the scope: Model sizes? How many GPUs?
• Define the contract: Sequence and batch? Failure budget?
• Test the boundaries: Shared cluster? Long context? FP8?

Constraints

Explicit scope and guarantees

Resolve the scoping questions for a Distributed Training Platform. Separate stated behavior from assumptions, and identify what is outside your design.

Supported operating targets

Declare relevant volume, latency, freshness, quality or cost targets with units. Show calculations or an evaluation plan that can test them; unspecified targets are your assumptions, not hidden pass criteria.

Failure and boundary behavior

Explain how your guarantees hold in a difficult case relevant to this subject. Address: Long context? FP8?

What to cover

1. 01

   Scope and behavior contract

   Identify users, required behavior and exclusions. Answer: Model sizes? How many GPUs?

2. 02

   State and interfaces

   Define the information owned by the system and the inputs, outputs and errors at its boundaries. Resolve: Sequence and batch? Failure budget?

3. 03

   Capacity and operating targets

   Estimate the dominant workload and resource demand with units and explicit assumptions. For a learned system, also state how quality is measured and what data is available.

4. 04

   Architecture and central flow

   Draw or describe the responsibilities needed for a Distributed Training Platform. Trace a representative request, event or job from its input to a visible result; identify durable state owners.

5. 05

   Failure and boundary walkthrough

   Walk through a difficult case step by step, including detection and recovery. Consider: Shared cluster? Long context? FP8?

6. 06

   Tradeoffs and operations

   Compare a credible alternative using your chosen workload and guarantees. Explain a remaining risk, a signal to watch and when you would change the design.

Worked designs

Explore the architecture and decisions, then build on an example with Coach.

Official design

Design a Distributed Training Platform

Train 1 B to 400 B models on 8 to 10,240 H100s: DDP, ZeRO and FSDP, tensor and pipeline parallel over NVLink and EFA, checkpoints, hot spares, MFU and goodput: one-hour boards for junior, senior and staff, with the theory behind them.

Explore the design

Review rubric

AI feedback uses these criteria. Scores are practice feedback.