| --- |
| license: apache-2.0 |
| pretty_name: CapyCTL bench results |
| language: |
| - en |
| tags: |
| - benchmark |
| - llm-inference |
| - inference-engines |
| - vllm |
| - sglang |
| - llama.cpp |
| - tensorfold |
| - capyctl |
| size_categories: |
| - 10K<n<100K |
| configs: |
| - config_name: points |
| default: true |
| data_files: data/points.parquet |
| - config_name: entries |
| data_files: data/entries.parquet |
| --- |
| |
| # CapyCTL bench results |
|
|
| Every point measured by [capyctl-bench](https://github.com/edurdias/capyctl-recipes/tree/main/tools/capyctl-bench) |
| for the 38 recipes and 12 comparisons in |
| [capyctl-recipes](https://github.com/edurdias/capyctl-recipes), in one table: decode speed, time to first token, |
| prompt processing and memory against prompt length and concurrent streams for chat models, and latency, throughput, |
| calibration and accuracy for decision models. vLLM, SGLang, TensorFold, llama.cpp and a custom engine, on an NVIDIA GB10, |
| an RTX 4090 Laptop GPU and two RTX A6000 cards. |
|
|
| Every request went through [CapyCTL](https://github.com/edurdias/capyctl)'s inference endpoint, not to the engine |
| directly. Built from capyctl-recipes at |
| [`1829ae0`](https://github.com/edurdias/capyctl-recipes/commit/1829ae0b0e6c4731016965eafdeea0ee0bc2a225). The reports themselves, with their charts, |
| are in the [benchmarks Space](https://huggingface.co/spaces/capyctl/benchmarks). |
|
|
| ```python |
| from datasets import load_dataset |
| |
| points = load_dataset("capyctl/bench-results", "points", split="train") |
| entries = load_dataset("capyctl/bench-results", "entries", split="train") |
| ``` |
|
|
| ## Configs |
|
|
| ### `points` (11,268 rows) |
|
|
| The `data.csv` of each of the 53 capyctl-bench reports, stacked, with the recipe or comparison it belongs to. |
|
|
| | Column | Meaning | |
| |---|---| |
| | `entry` | The recipe or comparison directory, such as `vllm/gpt-oss-120b-gb10`. Joins to `entries`. | |
| | `report` | The report's directory inside `entry`: `bench` for the main one, others for extra studies (`bench/overhead`, `bench-sweetspot`). | |
| | `kind` | `recipe` or `comparison`. | |
| | `report_title` | The report's title. | |
| | `series` | One results file: an engine, version and setting. Compare series only within one report. | |
| | `section` | `context` (one stream, prompt length growing), `concurrency` (streams or users), `latency`, `cache`, `accuracy` and `reliability` (decision models). | |
| | `point`, `x` | The point on the x axis: a prompt size such as `32k`, a stream count, a question count and state size, or a quality set. | |
| | `metric`, `unit` | What was measured, such as `decode_tps` in `tok/s` or `ttft_p50_s` in `s`. | |
| | `median`, `min`, `max`, `n` | The median over `n` samples and its range. Empty where nothing was measured (`n` 0). | |
|
|
| The chat metrics: |
|
|
| | Metric | Definition | |
| |---|---| |
| | `ttft_s`, `ttft_p50_s`, `ttft_p95_s` | Request sent to the first token chunk. | |
| | `decode_tps` | (completion tokens - 1) / (last token chunk - first token chunk): the decode rate one stream sees, drafts included. | |
| | `tpot_ms` | The inverse of `decode_tps`. | |
| | `prompt_tps` | Prompt tokens / time to first token. Includes queueing and the first decode step, so it understates raw prefill at short prompts. | |
| | `total_s` | Request sent to the end of the stream. | |
| | `output_bytes_per_s`, `output_chars_per_s` | Content and reasoning over the generation window; compares engines whose tokenizers differ. | |
| | `aggregate_tps` | Concurrency: a round's completion tokens / (first request sent to last stream finished), median over rounds. | |
| | `per_stream_tps` | Concurrency: median `decode_tps` over every request at that stream count. | |
| | `draft_acceptance` | Accepted / drafted tokens, when the engine reports draft counts. | |
| | `memory_peak_gib` | The highest memory sample taken while the request or round ran (machine RAM in use on unified memory). | |
|
|
| The decision-model metrics (latency, `requests_per_s`, `questions_per_s`, `accuracy`, `macro_f1`, `brier`, `nll`, `ece`, |
| `kl`, and the reliability bins) are defined in |
| [Measure a decision model](https://github.com/edurdias/capyctl-recipes/tree/main/tools/capyctl-bench#measure-a-decision-model). |
|
|
| ### `entries` (50 rows) |
|
|
| One row per recipe or comparison, from the tables of the capyctl-recipes README: `engine`, `model`, `hardware`, the |
| headline `result`, the date it was `measured`, the number of `reports`, the GitHub `url` and the main `report_url`. |
| The native TP2 and concurrency comparisons used their own harness, so they have no rows in `points`; their evidence is |
| on GitHub. |
|
|
| ## Reading the numbers |
|
|
| - Compare series within one report. A report puts series side by side only when they ran on the same machine with the |
| same prompts and settings; across reports, prompts, settings, CapyCTL and engine versions can differ. |
| - Each recipe's README names the exact CapyCTL, engine, model and drafter revisions, the deployment file and the |
| settings. Read it before quoting a number. |
| - One machine per result, on the date given. These are measurements, not a qualification of an engine on your |
| hardware. |
| - Each recipe states its thinking setting and sampling parameters. Reasoning tokens count as output tokens, and many |
| runs cap replies with `max_tokens`. |
|
|
| ## License |
|
|
| Apache-2.0, as capyctl-recipes. The model names refer to their own repositories and licenses. |
|
|
