Microscaling PTQ research harness
none
Fields
id
engines/mxptq-research-harnessRepo
microsoft/NeuralCompression
Languages
Python
License
MIT
First release
2023
Design goal
minimal-deps
Scheduling
none
Backends
none
Formats
mxfp4mxfp8mxint4mxint8int4int8
Notable features
Reimplements GPTQ, AWQ and SmoothQuant inside one harness so the algorithms are comparableReports WikiText-2-test and C4 perplexity plus 8-task zero-shot accuracy in a single tableMX formats at block size 128 (and a 16-wide ablation in the appendix)
Hardware caveats
Not a serving runtime. Results are fake-quant / simulated formats: nothing here was run on MXFP4 or MXFP8 hardware
Best for
apples-to-apples comparison of INT vs MX formats and of PTQ algorithm combinations
Avoid for
latency, or any claim about hardware MX kernels
notes
The single most useful source in this repo for the 'is that delta real or is that algorithm choice' question, because it holds model, harness, calibration set, group size and format fixed while varying only one thing. Its own finding that GPTQ and AWQ are synergistic but AWQ+SmoothQuant is not is a methodology warning in its own right. ||| MAINTENANCE: NOT VERIFIED on 2026-10-04 - the engine record's `repo` field (microsoft/NeuralCompression) returned HTTP 404 from the GitHub API, so no maintenance fact could be pulled for it and none is estimated here. See [[ev-maintenance-status-is-four-dated-facts]].
Notes
The single most useful source in this repo for the 'is that delta real or is that algorithm choice' question, because it holds model, harness, calibration set, group size and format fixed while varying only one thing. Its own finding that GPTQ and AWQ are synergistic but AWQ+SmoothQuant is not is a methodology warning in its own right. ||| MAINTENANCE: NOT VERIFIED on 2026-10-04 - the engine record's `repo` field (microsoft/NeuralCompression) returned HTTP 404 from the GitHub API, so no maintenance fact could be pulled for it and none is estimated here. See [[ev-maintenance-status-is-four-dated-facts]].
Cross-references
This record references no other records — it is a leaf. It is cited by 3.