Google

An architectural model where the prefilling and decoding phases of inference are treated as separate stages of computation but are executed on the same hardware. This approach is a common foundation for advanced batching techniques that improve upon simpler strategies.

Aggregated Architecture for Prefilling and Decoding

Continuous batching is an iteration-based scheduling method, notably used in the Orca system, where the composition of a request batch is dynamically adjusted between computational steps. This flexibility allows for new input sequences to be added or completed sequences to be removed from the batch during any iteration. This adjustment can occur even if the processing for the entire batch is not yet finished, distinguishing it from static methods.

Continuous Batching for LLM Inference

Based on the architectural principle of separating the distinct computational phases of inference, propose a change to the team's batch processing logic to improve GPU utilization. Explain why your proposed change would be effective.

Optimizing Inference Throughput

An inference system processes user prompts by first computing the initial state for a whole batch of requests (the 'prefill' stage), and only then proceeds to generate responses token-by-token for that same batch (the 'decoding' stage). Describe one major efficiency benefit and one potential drawback of this approach where the entire batch must complete the first stage before the second stage begins.

Trade-offs in a Staged Inference Architecture

In a common architecture for language model inference, the initial processing of a user's prompt (prefilling) and the subsequent token-by-token generation of the response (decoding) are treated as distinct computational stages, even though they execute on the same hardware. What is the primary analytical reason for this architectural separation?

Learn Before

Related