AI 추론

A classifier can return a category score; a language model can generate tokens. Inference usually leaves the model parameters unchanged, although a surrounding system may separately save information or learn from feedback.

주요 시사점

• Measure the entire request path.
• Separate per-request latency from throughput.
• Retest quality after serving optimizations.

A request typically passes through input validation, preprocessing, the model, and output processing. A text service may tokenize a prompt, run the model repeatedly to generate tokens, and assemble the response. Retrieval and external tools can add more stages around the model. Their time and errors count toward the user experience. Measure latency and throughput separately. Latency is how long one request takes; throughput is how many requests the system finishes over a period. Batching requests may improve throughput while increasing the wait for an individual request. Streaming can make an answer begin sooner without reducing the time required to finish it. Hardware memory must accommodate more than the model weights. Working buffers, concurrent requests, and cached representations also consume memory. Longer inputs and outputs can change the serving cost, so test the actual workload distribution rather than one short demonstration prompt. An inference deployment needs limits, timeouts, and a usable response when the model cannot answer. Keep a versioned evaluation set and compare outputs after changing precision, batching, model versions, or preprocessing. An optimization is useful only if it preserves the quality required by the task.

1. In a constructed request, validation takes 20 ms, document retrieval 180 ms, model generation 900 ms, and formatting 30 ms.
2. If these stages run sequentially, the total is 1,130 ms. Halving formatting time saves only 15 ms.
3. Reducing retrieval to 100 ms saves 80 ms. Measure again under concurrent load because queueing can change the result.

These invented timings illustrate why optimizing a small stage may barely change the experience.