inference/research
verifiedmedium · 0.75updated 2026-10-03

SpinQuant research harness (fake-quant + Triton INT4 kernels)

cuda

Fields

id
engines/spinquant-research-harness
Repo
facebookresearch/SpinQuant
Languages
PythonCUDA
License
MIT
First release
2024
Design goal
minimal-deps
Scheduling
none
Backends
cuda
Formats
int4spinquant
Notable features
Cayley SGD on the Stiefel manifold to learn rotation matrices R1,R2 (and online Hadamard R3,R4)Evaluated largely with fake quantization, plus research Triton INT4 GEMM kernelsTwo variants: SpinQuant_no_had (rotations merged into weights, zero forward-pass change) and SpinQuant_had (needs online Hadamard kernels)
Hardware caveats
NO hardware INT4 tensor-core path: W4A4KV4 accuracy is a fake-quant / research-kernel result, not a silicon resultSpinQuant_had needs an online Hadamard kernel, ~8% of network latency overhead per the paper
Best for
understanding what W4A4 costs when activations AND KV cache are quantized
Avoid for
production deployment as-is
notes
Recorded because the SpinQuant W4A4 quality deltas are harness numbers, and 'there is no INT4 tensor-core path for 4x4' is the whole point of the format record. The rotation matrices are ~0.26% of weight size and are folded into the checkpoint. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2025-02-14 (19.6 months before this record), NO GitHub release or tag found, 2 contributors (distinct committers including anonymous), open_issues_count 31 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 432 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Notes

Recorded because the SpinQuant W4A4 quality deltas are harness numbers, and 'there is no INT4 tensor-core path for 4x4' is the whole point of the format record. The rotation matrices are ~0.26% of weight size and are folded into the checkpoint. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2025-02-14 (19.6 months before this record), NO GitHub release or tag found, 2 contributors (distinct committers including anonymous), open_issues_count 31 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 432 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Cross-references

This record references no other records — it is a leaf. It is cited by 5.