DEDUPED Hot AI repos, turned into use cases Get 5 a day

How to compare FP32 and quantized state accuracy on long tasks with STEP Quant

Run the evaluation suite to score FP32 against STEP Quant@6 on long-generation benchmarks.

  • Research
  • Intermediate
  • half day
  • From the README
  • Any agent

Who it is forA model evaluator who needs accuracy evidence before deciding whether a quantized recurrent state is safe to deploy.

The problemMemory savings mean little if long-generation accuracy drops, and the comparison has to be run on the same model.

The outcomeAverage accuracy scores for FP32 and STEP Quant@6 on the long-generation benchmark set, to compare against the paper table.

What you need first

  • Model checkpoints for a supported model, such as Qwen3.8-27B, placed on disk
  • A CUDA toolkit and NVIDIA GPUs (the Qwen example uses four GPUs; the paper used NVIDIA A800 GPUs)
  • Python with the tested stack: SGLang 0.5.12, PyTorch 2.11.0, Triton 3.7.1
  • A clone of the repository, run from its root

Steps

  1. Complete the calibration step so the stepquant6 plan exists at artifacts/reproduction/plans/qwen-stepquant6.pt.
  2. Stop any manually started server, since the evaluation suite starts its own.
  3. Run .venv-eval/bin/python -m stepquant.evaluation.suite --models qwen --benchmarks long --formats fp32 stepquant6 --server-python .venv-serving/bin/python --devices 0,1,2,3 --tp 4. (inferred, check the README)
  4. Open REPRODUCTION.md for the zero-shot short-task protocol if you also want short-task scores. (inferred, check the README)

Prompt for your agent

In the STEPQuant repo at <repo_path>, run the evaluation suite on the long benchmarks for the fp32 and stepquant6 formats of the Qwen model. Report the average accuracy for each format and state whether the stepquant6 score is within 1 point of the fp32 score.

Caveat. The README notes that the 4 and 6 bit budgets are nominal because FP16 pivots and scales add storage, so reported accuracy should be read with that in mind.

Share: X LinkedIn Hacker News Email Copy link

Disclaimer

Deduped lists open-source repositories that belong to other people. We do not own, maintain, audit, or vouch for any of them, and a repo appearing here is not an endorsement or a safety check. The use cases, steps, and prompts are written by AI and may be incomplete, outdated, or just wrong. Where a use case is marked "Seen in the wild", we link to public posts where someone described something similar, but we have not verified those posts and the use case itself is still our AI's own write-up. Use all of this at your own risk. Only run code, commands, or agents on systems and accounts you are authorized to use, and read the code and every command before you run it, especially anything that touches credentials, files, or the network. Nothing here is professional, security, legal, or financial advice. All trademarks and project names belong to their respective owners, and their use here does not imply affiliation. Content is provided as is, without warranties of any kind. Found a mistake, a safety problem, or something you want removed? Email [email protected] and we will look at it promptly. Full disclaimer page