Start with the stage
Prefill processes input tokens; generation produces output tokens. A report can measure either stage or both together. In llama-bench, pp denotes prompt processing, tg denotes text generation, and pg combines them. The number beside the test name records its token workload. A high pp result therefore does not establish an equally high output rate.
Keep the full test label and context depth. The llama-bench tool can prefill its cache to a specified depth before timing a test, and its measurements exclude tokenization and sampling. These are useful engine measurements, but they do not represent every step in an interactive application.
Latency answers a different question
Time to first token measures the delay before an answer starts. End-to-end latency measures the whole request. The vLLM benchmark client also reports time per output token and inter-token latency, which describe how output arrives. Check whether a timing comes from the client or the engine before comparing it with another report.
Even similarly named latency metrics need care. vLLM distinguishes gaps between streamed output events from a per-request average over output tokens. They can differ when an event contains several tokens or when statistics use different weighting. Preserve the source's metric definition instead of automatically converting every latency into a tokens-per-second figure.
Compare the same workload
Start with a supported GPU page, open the explorer, and narrow the card count, exact model, quantization and host settings. Inspect prompt length, output length, context depth and concurrency in each record. Keep a single-request generation result separate from server throughput across simultaneous requests.
Open the original report before drawing a conclusion. Check whether the figure is a mean, median, peak, range or approximate chart reading. Missing settings are a limit on the comparison, not permission to assume defaults. A speed measurement also does not establish an accuracy score; use an associated evaluation only when the source documents that relationship.