Speed and quality measurement tool to compare inference engines honestly.
Summary • Quick start • What crit-eval solves •
What crit-eval measures • Commands •
Limitations • Documentation
Summary
crit-eval runs one fixed suite against every inference engine, so results compare.
Quick start
python -m pip install "crit-core @ git+https://github.com/crit-labs/crit#subdirectory=core/python"
python -m pip install "crit-eval[reference] @ git+https://github.com/crit-labs/crit-eval"
Serve your engine behind an OpenAI-compatible API, then measure it:
crit-eval probe --out device.json
crit-eval fetch --model MODEL_DIR --data EVAL_DATA_DIR --packages PACKAGES
crit-eval run --url http://127.0.0.1:8000/v1 --model MODEL_DIR --data EVAL_DATA_DIR --device device.json \
--profile hub --out result.json
crit-eval status result.json
fetch gets the checkpoint's reference package from PACKAGES (a folder or an
https base URL) when it has one; without it,
crit-eval reference scores the reference once on your GPU (the reference extra installs torch and
transformers for that). Inside crit the same commands are crit eval .... An engine without a server
plugs in through a Python adapter instead of --url.
What crit-eval solves
There is no shared way to measure a bespoke inference engine.
Different bespoke inference engines report vastly different types of numbers. An honest comparison requires downloading and running the engines yourself.
crit-evalfinally makes apples-to-apples comparison possible.
Physics-based evaluation lets you see what performance is left on the table.
crit-evalmeasures your GPU's memory bandwidth and matrix-multiply rates and turns them into a limit for each score. Its evaluation report can tell your coding agent how much performance can still be squeezed out of your inference engine.
Speed that can't hide a quality loss.
Many inference engines push decode and prefill speed with no regard for quality.
crit-evaltreats quality as a first-class KPI. It measures not only inference speed but also quality regressions.
A hardware report rounds out the evaluation.
The same model on the same inference engine but on two different machines can also lead to different inference speeds, even if the GPU is the same. We make this transparent by also reporting the runner's hardware configuration.
What crit-eval measures
| KPI or record | In the result | What it records |
|---|---|---|
| decode_greedy | status.kpis.decode_greedy |
Greedy decode tokens per second over six workloads (chat, code, math, tool calls, a long document, a 14K-token agent session), with run-to-run noise |
| decode_sampled | status.kpis.decode_sampled |
Decode tokens per second at the model's own sampling, the headline decode, and tokens per round for speculative engines; a case whose three seeds disagree is repeated at further seeds (up to 9) until its standard error is within 3%, and each case's error is reported |
| prefill | status.kpis.prefill |
Prompt tokens per second on the two long prompts |
| ttft | status.kpis.ttft |
Milliseconds to the first token over every workload |
| long-context decode and ttft | status.kpis.context |
The same scores along a curve of prompts up to the required context (262,144 tokens by default) |
| perplexity | status.kpis.perplexity |
The engine's perplexity over the reference's: the worst quality set (prompt and served tokens, each long length) against its limit; per set in quality.sets |
| kl_divergence | status.kpis.kl_divergence |
Mean and p99 KL divergence of the engine's token distributions from the reference's: the worst set against its limits; per set in quality.sets |
| limits and gaps | headroom |
Each KPI's physical limit on this machine, from the probe's best rate per format (cuBLASLt GEMM or raw tensor cores), and what each term (model pass, draft steps, verify pass, draft width, drafter acceptance, prefill, first token) would give at its floor |
| prefill by activation format | headroom.activations |
The prefill and TTFT ceilings with the quantized matrices' prompt activations at BF16, FP16 accumulation, FP8 and the stored format, which of them the quality evidence rules out, and the reachable limit: the fastest shown to pass |
| prefill by term | crit-eval status --kernels |
With a kernel trace, each prefill term (matmul per weight format, attention, recurrence, other kernels, host) against its floor, a gap per term, and a speculative round split into verify pass and draft steps |
| time to first token by part | headroom.ttft |
Each case's time to first token split into the engine's own prefill and first token and the rest (server, tokenizer, HTTP) |
| quality noise and attribution | quality.sets.*.budget_spread, quality.attribution |
Each set's share of the drift budget with its spread over its positions; with --attribute-drift, each matrix group's share of the budget |
| free device memory | context.device_memory |
Device memory left free with the engine built for the required context, beside the capacity gate |
| gates | status.gates |
Repeatable output, request history, prompt cache, retrieval, exactness, sampling and quality, each passed or not, with every KPI against its limit (status.kpis) |
| retrieval | context.retrieval |
Facts at the start, middle and end of long prompts, found or not |
| exactness | exactness |
Whether speculative output equals plain greedy, case by case |
| sampling | sampling_check |
Whether sampled tokens follow the model's distribution |
| request history | history |
Whether the same greedy requests give the same tokens after a long request history |
| prompt cache | followup |
Follow-up turns answered from a prompt cache: the same answer, and their prefill and ttft |
| held-out prompts | heldout |
Speed on ten prompts in other languages and styles, never used for tuning |
| hardware | host, device |
OS, CPU, RAM and GPUs (name, VRAM, driver, PCIe link), and the measured GPU bandwidth and matmul rates, PCIe transfer rates, host memory bandwidth and CPU throughput |
| engine | engine |
What the engine reported: name, configuration, max context, drafting, bytes read per pass |
| model | model, model_sha256 |
Name, repository and revision, format, architecture and quantization |
| reference | reference |
The reference that computed the quality scores (version, transformers, torch, GPU) and the package they came from |
| run | crit_eval_version, protocol, profile, requirements, timing |
Which crit-eval, protocol and profile measured it, the drift budget it was held to, and how long each part took |
Commands
| Command | Does |
|---|---|
probe |
Measure the device's bandwidth, cuBLASLt GEMM rates per weight format and raw tensor-core rates per input and accumulator format (needs nvcc) |
context |
The max context for a model on a device: the model's positions and the device's memory |
prepare |
Freeze the suite as token ids for one model |
reference |
Score the prepared quality texts once on the checkpoint |
fetch / pack |
Get a checkpoint's reference package, or add one to a package folder |
pin |
Record an engine's plain greedy output of every speed case as the data's decode text |
run |
Evaluate one engine, through --url or an --adapter; with an adapter, --only quality checks quality in process (development evidence) and --attribute-drift attributes the drift to matrix groups |
status / compare |
Gates, goals, KPIs against their limits and gaps (--kernels joins a kernel trace to the prefill floor, --device and --data price a result again); results side by side |
ab |
Alternate two engines' speed runs in one session and report paired ratios |
kernels |
Trace every kernel of the speed cases (CUPTI, Linux), prefill on a fresh prompt, and rank them by GPU time per phase and prefill-floor term |
Limitations
| Area | Current status |
|---|---|
| Quality reference | Quality is measured against the checkpoint as stored, for example an NVFP4 quantization, not against the full-precision model. crit-eval reports the drift the engine adds, never the quantization loss of the checkpoint itself, so quality results of two different quantizations of a model do not compare. |
| What quality covers | Drift on fixed texts: six conversations (prose, code, math, tool calls and more) and one long text each at 1K, 4K and 32K tokens, as perplexity ratio and KL divergence. It is not a task benchmark: no accuracy on MMLU-style questions, coding tests or agent tasks. |
| Reference coverage | The reference runs the model through transformers' modeling code in FP32, so it needs an architecture transformers implements and a CUDA GPU with the machine to itself for 5 to 25 minutes. Checkpoints in exl2 or exl3 are refused. Reference packages for three checkpoints are published on Hugging Face (crit-labs/crit-eval-references) and fetched by default. |
| Rounding-sensitive models | Where a model amplifies rounding (MoE routing near-ties, sparse-attention ties, recurrent state), two valid FP32 runs of the reference disagree; quality is then judged against twice the reference's own spread instead of the budget, which is looser for those models. |
| Speed | One request at a time: crit-eval measures single-stream decode, prefill and time to first token, not throughput under concurrent requests or batching. A noisy sampled decode is repeated at up to nine seeds per case, so an engine whose speed swings with the drawn text takes longer to measure (--no-adaptive-sampled keeps three, an override the result records). |
| Hardware | NVIDIA GPUs only: the probe builds with nvcc and measures one GPU. Multi-GPU engines are not measured as such. PCIe, host memory and CPU rates are recorded, but no limit uses them yet, so engines that offload weights or compute on the CPU get GPU-only limits. On Windows the CPU probe uses AVX2, which under-reports an AVX-512 CPU. |
| Long context | The hub profile measures the long-context curve at its two ends and interpolates the points between; only the full profile measures every point. |
| Comparability | Results compare only under the same protocol series and profile, on prepared data of the same checkpoint. |
Documentation
| To… | Read |
|---|---|
Run crit-eval and read its results |
Guide |
| See every case, score, limit and gate | Measurements |
| Set the required context | Context |
| Plug in an engine without a server | Adapter |
| Build an engine to measure | crit · Building engines |
| See release history | Changelog |
1 maintainer
Metadata
1 maintainer
Release files for crit-eval 0.2.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | |
|---|---|---|---|---|
| crit_eval-0.2.0.tar.gz | 378.7 kB | With provenance | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| crit_eval-0.2.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 667.8 kB
Release files / crit_eval-0.2.0.tar.gz
| Download URL | crit_eval-0.2.0.tar.gz |
|---|---|
| Size | 378.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
b5e6c06cdb83a8b4dda9e5356356288e9e8184fb63d9d4a71e5622f45010f25b
|
|
BLAKE2b-256 checksum How to use checksums |
bf73c543cdaf8eb532c92c81e553fe7b8349de130dad635a11142d3d3fdb33e0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 27, 2026.
Transparency logRelease files / crit_eval-0.2.0-py3-none-any.whl
| Download URL | crit_eval-0.2.0-py3-none-any.whl |
|---|---|
| Size | 289.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
a580e7f41256d2fc940c4d4d71a827e20a9438c1645197696c787b458c90290c
|
|
BLAKE2b-256 checksum How to use checksums |
f6d0dfa6434a63009745cd068aad6258b5e0b33ea8d7466db58760600980c929
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 27, 2026.
Transparency logSecurity for crit-eval 0.2.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
All 2 files in this release have been uploaded with provenance metadata.
Attestation for crit_eval-0.2.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 27, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.
