What is the difference between data parallelism and model parallelism?
Deployment and LLMOps · Pro 2
Short answer
Data parallelism runs full copies of the model on several GPUs, each handling different batches in training or different requests in serving. Model parallelism splits one model across GPUs because it is too big or too slow for one: tensor parallelism splits work inside each layer, pipeline parallelism splits groups of layers.
Deeper answer
Data parallelism is simple and scales throughput, but each copy must fit. In training it also needs gradients and optimiser states, which is why sharded data parallelism (ZeRO, FSDP) spreads those across GPUs.
Tensor parallelism needs fast links because GPUs exchange results every layer; pipeline parallelism tolerates slower links but adds per-request delay, using micro-batches to keep stages busy in training. Large training runs combine all of these, often with context or sequence parallelism for long inputs.
What a strong answer covers
- Data parallelism: full copies, different work
- Model parallelism: split one model
- Tensor versus pipeline
- Notes communication or memory trade-offs
Answer it out loud in the mock interview about 1 min
Follow-ups to expect
- What does ZeRO or FSDP shard?
- Which would you use to cut latency for one large model?
Common mistakes
- Calling replicas model parallelism
- Ignoring communication cost
Learn it
Running models on many GPUs: replicas, tensor and pipeline parallelism
- When would you serve predictions in batch, and when online?
- How would you set up CI/CD for an LLM application?
- Why does experiment tracking matter, and what do you record?
- Your model scored well offline but performed badly after launch. What could explain it?
- Explain Kubernetes pods, deployments and services, and why you would use Kubernetes for AI apps.
- What is the KV cache, and why can it cause out-of-memory errors when serving an LLM?
- How would you load test an LLM feature and set a latency budget?
- Your model provider has an outage during peak hours. What should already be in place?
- How do you manage API keys and secrets across environments?
- How do you detect and handle model drift in production?
- A bank wants your AI system but says no data may leave its network. What are the options?
- Why does an LLM reply have a wait before the first word and then a different speed after it?
- What is quantisation, and what are the trade-offs?
- When would you use a semantic cache for an LLM app, and what can go wrong?
- How would you serve an LLM feature as an API for a web app?
- How would you serve a model that is too big for one GPU, and how would you scale it to more users?
- How would you add single sign-on to an enterprise AI assistant?
- How do you version prompts and models so you can roll back safely?
- What would you capture in traces for an LLM application?
- Why use Docker for an AI application?
- What does an LLM gateway do, and when do you need one?
- Why use vLLM to serve an open-weight model instead of a simple Hugging Face script?