Skip to the content.

ReQuant fixed-grid refinement evaluation

Status: modeled contract witness, not an observed model- or hardware-performance result.
Issue: #6253, child of #6221.

Pinned research and implementation envelope

Component Pin Provenance
Research artifact arXiv:2608.07019v1 PDF SHA-256 5505cab5060f170e5e7a03b07fce83682c7af3bb0c3cfbb9cdb5920c220d4beb; Yongge Ma et al., submitted 2026-08-07
Neutral contract requanteval/v1 internal/requanteval/contract.go
Recipe requant-arxiv-2608.07019v1-fixed-grid-v1 deterministic fixed-grid coordinate refinement modeled from Algorithm 1
Runtime fak-go-reference-cpu/v1 standard-library Go objective evaluator; not a production inference runtime
Initializer fixture-pinned, currently round-to-nearest@fixture-v1 carried in every result alongside artifact ID/version/digest/source

ReQuant is a post-processing method: it starts from an already quantized assignment, keeps the original quantization grid fixed, and accepts discrete coordinate changes that reduce a reconstruction objective. This leaf evaluates that defining seam against the same initial codes and grid. It does not create a fak artifact format, rank PTQ methods, implement a model kernel, or silently substitute another runtime.

Typed adjudication

requanteval.Evaluate always returns one of:

The v1 evaluator is deliberately bounded (2–256 finite ordered grid levels, 1–4096 coordinates, 1–1000 sweeps). The seed is captured for reproduction, although this deterministic recipe consumes no random numbers.

Named witness

The independently readable golden files under internal/requanteval/testdata/ are the reproduction artifact. Together they cover three cases on Windows and WSL/Linux:

  1. improves.json — a coupled two-coordinate quadratic where refinement starts from the same RTN assignment, remains on [-1,0,1], and lowers modeled reconstruction MSE from 0.164 to 0.144 (12.20%); the same pinned three-example linear probe exposes a counterexample—prediction MSE worsens from 0.08 to 0.7467, so lower reconstruction loss is not reported as a universal quality gain; the artifact records seed, grid, initial/refined codes, sweeps, conversion work, metrics, provenance, and claim-check result.
  2. stable.json — an initialization for which no strictly improving fixed-grid update is accepted.
  3. delegate.json — an unknown provider runtime produces a typed delegation and no fabricated metrics.

Run:

go test ./internal/requanteval -run 'TestEvaluateWitnessFixtures|TestTypedUnsupportedOutcomes|TestSameInitializationAndFixedGrid' -count=1

The golden read-back fails if recipe behavior, cost accounting, pins, evidence class, or adjudication drifts.

Evidence boundary and claim check

The fixture’s reconstruction MSE, synthetic prediction-quality MSE, and conversion-operation counts are observed outputs of the deterministic evaluator, but they describe a synthetic quadratic. Therefore the overall evidence classification is modeled, and its embedded claim-check verdict is not-yet for downstream model quality or hardware performance. hardware_envelope explicitly says that no wall time, throughput, GPU, or model-quality run was measured.

The paper reports broad model/task improvements and runtime overheads in its own experimental envelope (including RTN, GPTQ/GPTAQ and AWQ initializers), but fak has not independently reproduced those values. They are research provenance, not fak-observed results, and are intentionally not copied into a performance claim. A future observed claim must pin the weights/dataset/calibration recipe/runtime/device/software stack, compare the same quantized initialization and format, capture downstream quality plus conversion wall time, and pass fak claim-check against the tuned alternative.