HighVRAM cited benchmark reference snapshot
cited-4d82916d23e704c3

These are externally reported results, not HighVRAM measurements. Sources use different models, runtimes, workload shapes and timing definitions. Unknown settings stay unknown; rows are not a normalized ranking. Contract validation does not verify execution or reproduce the results.

This snapshot contains 30 metric-observation records from 5 cited sources. Reported metric counts: 13 prompt-processing/proxy, 13 generation/decode and 4 TTFT observations. Observation rows are not independent runs. Prompt-processing and generation rows may be separate microbenchmarks; do not assume their token lengths describe a single paired request. Repetition counts and exact execution dates may be unreported.

JSON is the lossless reference. Null means not reported or not established by the cited source; it is never zero. CSV empty numeric fields have the same meaning. CSV source strings that could be interpreted as formulas are prefixed with an apostrophe; JSON preserves their original text.

Prompt processing, generation, TTFT and throughput whose phase is unspecified are separate columns. Memory values are in the source's reported GB, with no GB/GiB conversion implied. TTFT is in milliseconds. Source dates are publication dates, not necessarily measurement dates. Access timestamps are recorded per source row. Reported GPU power limits are not measured power consumption. Model-file size is not peak memory. Source qualifications, including mixed or 'mostly' quantization, are retained in each record's notes.

Each record has a citation URL and table/row locator. Source IDs derive from citation URLs. Test-profile IDs derive conservatively from source, hardware, model, runtime, workload and notes; they are identifiers, not proof of comparability. These exports are observations attributed to sources, not raw logs or reproduced runs.

For Rapid-MLX, prompt throughput is prompt tokens / streamed TTFT including serving overhead; it is not isolated engine prefill. Corrected decode is (completion tokens - 1) / (total time - TTFT). TTFT ends at the first visible streamed content, reasoning or tool delta. These three metrics describe the same three-request median group. Values preserve the published precision; extra digits from CSV recalculation are not extra measurement precision. The JSON source_metric_definitions map supplies these source-specific definitions. Neither this route nor the MLX-LM CLI is normalized against the other sources.

This snapshot attributes extracted factual observations to their original sources. It does not grant a license to the original sources or claim ownership over them. Existing source rights remain unchanged.
