To maximize throughput, what two distinct computational operations will the server perform in parallel during the next single processing step (at time T+1)?

Google

This diagram illustrates the sixth iteration in a continuous batching process, which occurs after request x₂ has completed. New requests, x₄ and x₅, arrive in the system. The scheduler dynamically adjusts the batch by using the resources freed up by the completed request x₂ to accommodate the new request x₄. In a single computational step, the system concurrently performs the prefilling phase for x₄ while also executing a single decoding step for the ongoing requests x₁ and x₃. This highlights a key efficiency of continuous batching: the immediate reuse of resources to interleave the processing of new and existing requests, thereby maximizing throughput.

Example of Reusing a Completed Slot in Continuous Batching (Iteration 6)

An inference system is processing a batch of requests using a dynamic scheduling method. At a specific moment, one request (Request A) completes its generation. The system still has two ongoing requests (Request B and Request C) that require further processing. At the same time, a new request (Request D) arrives. Given this state, which of the following actions by the system's scheduler represents the most efficient use of computational resources in the very next step?

Inference Server Task Scheduling Analysis

An LLM inference server is processing a batch containing requests A, B, and C. Request B completes, and its resources are freed. A new request, D, arrives and is immediately added to the batch. In the next single computational step, what two distinct types of operations will the server perform concurrently on the active requests (A, C, and D) to maximize efficiency?

Learn Before

Related