Quantize · Quantra

Lower precision, with the accuracy bill shown

Quantra searches per-layer FP8 and INT4 precision and sparsity, calibrates on representative data, and measures speed and accuracy on real GPUs. It keeps only configurations inside your accuracy budget.

Compression resultExample
Policy
FP8, 76 of 80 layers
Kept in BF16
4 attention layers
Accuracy change
−0.3% on eval set
Throughput
1.62× vs BF16
Budget
1.0%
Example. Values are illustrative.

Capabilities

What Quantra does

01

Per-layer mixed precision

FP8 and INT4 chosen layer by layer, not one setting for the whole model.

02

Sparsity search

Structured and unstructured sparsity tried and validated on hardware.

03

Calibration selection

Representative calibration samples picked automatically from your data.

04

Accuracy gate

Every configuration scored on your evaluation set before it can be kept.

05

Pareto report

Accuracy against throughput for every point tried, so you choose with evidence.

06

Portable artifacts

Compressed models for cloud, on-prem and edge through Edgyn.

Accuracy budget

Pick the point that meets your SLA

Every configuration Quantra tries is measured for speed and accuracy. Move the budget and see which ones would be kept. Example data.

1.0×1.4×1.8×2.2×0.0%1.0%2.0%3.0%Accuracy drop on your eval setbudgetBF16 baseline: accuracy −0.0%, 1.0× throughputFP8 all layers: accuracy −0.9%, 1.71× throughputFP8, 76 of 80 layers: accuracy −0.3%, 1.62× throughputFP8 + 2:4 sparsity: accuracy −1.6%, 1.94× throughputINT4 weights, FP8 act.: accuracy −1.2%, 1.88× throughputINT4 all layers: accuracy −2.8%, 2.15× throughputFP8, attention in BF16: accuracy −0.15%, 1.41× throughputINT4 MLP only: accuracy −0.7%, 1.77× throughputFP8 + KV cache FP8: accuracy −0.5%, 1.8× throughputINT8 smoothquant: accuracy −0.4%, 1.52× throughput

Best kept configuration

FP8 + KV cache FP8

Inside budget: 7 of 10

Best throughput inside budgetInside budgetOutside budget (rejected)

Example configurations for one 70B chat model on H100. Values are illustrative.

How it works

Inside a Quantra run

  1. 01Hour

    Calibrate

    Select calibration data and record baseline accuracy.

  2. 02Hours

    Search

    Propose per-layer precision and sparsity policies.

  3. 03Hours

    Measure

    Run each on the GPU for throughput and on the eval set for accuracy.

  4. 04Instant

    Choose

    Keep the fastest point inside the budget.

Specification

Inputs, outputs and hardware

ItemDetailStatus
InputModel, eval set, accuracy budgetPilot
PrecisionsFP8, INT8, INT4 weightsPilot
Sparsity2:4 structured, unstructuredNext
EngineTensorRT-LLM quantized enginesPilot
TargetsHopper, Blackwell, Jetson OrinPilot · Planned
OutputCompressed model, accuracy reportPilot

FAQ

Questions engineers ask

Who sets the accuracy budget?

You do, on your own evaluation set. We suggest a starting point, but the number and the eval set are yours.

What if my eval set is small?

Quantra reports the confidence of each accuracy measurement and flags results where the eval set is too small to tell.

Can some layers stay in full precision?

Yes. Quantra often keeps sensitive layers, such as some attention layers, in BF16, and you can pin layers yourself.

Does FP8 work on older GPUs?

FP8 throughput gains need Hopper or Blackwell tensor cores. On older GPUs Quantra uses INT8 and INT4 where they help.

Next step

See your model's accuracy curve

In the pilot Quantra produces a full accuracy-throughput report for one model on your evaluation set.