Repository navigation
v0.23.0 - Progressive precision and formula-transfer screen
Pre-release
Progressive precision and formula-transfer research: a bounded negative screen
This release implements a resident four-bit FFN base with physically loadable
two-bit increments, adaptive six/eight-bit acquisition and cross-token cache
state. It also tests a numerical-analysis hypothesis suggested by the project
owner's formula-to-other-domains method. No llama.cpp code or flags were changed.
Measured results
The single screen finishes in 93.543 seconds: 81.182 seconds of supervised
worker time plus 12.362 seconds of independent analysis. Two known documents,
four prefix/four output tokens each, six conditions, one repeat. All 12 scored
candidate outputs and six warmups match the dense target. Small denominators
remain explicit; this is not general quality or larger-model qualification.
| Draft | Accepted proposals | Scored seconds | Increment H2D GiB |
|---|---|---|---|
| Four-bit base | 6/16 | 2.958 | 0 |
| Six-bit streaming | 5/20 | 8.574 | 8.3441 |
| Eight-bit streaming | 8/8 | 6.841 | 8.6133 |
| Adaptive precision | 8/8 | 6.886 | 8.6133 |
| Adaptive + value-ranked cache | 8/8 | 6.877 | 7.9884 |
| Adaptive + fixed cache | 8/8 | 6.823 | 7.5366 |
Adaptive precision acquires all eight bits on 448/448 scored FFN executions.
Value-ranked retention saves 7.25% of traffic, below the registered 10% threshold;
fixed retention saves 12.5% and attains the access-pattern cache bound. Value
ranking uses 5.99% more traffic than that fixed control. Its 0.13% time reduction
versus streaming does not pass the 5% timing check or establish a reliable gain.
The resident dense reference takes 0.271 seconds for the same eight tokens.
Eight-bit draft proposals work in this tiny sample, but six bits fails its
registered improvement over four bits. All three acquisition/representation
hypotheses fail their thresholds. The frozen decision is stop expansion of this
configuration, not that novel loading is impossible. No long suite or native
pivot follows.
Cross-domain hypothesis and memory
For successive four/six/eight-bit outputs, Richardson extrapolation assumes a
stable leading error term. The registered formula y6 + (y6-y4)/3 instead raises
mean local relative-L2 from 0.0498965 to 0.0794774: 59.28% worse, with 0/28
layers improved. Increment norms shrink about fourfold, but mean directional
cosine is only 0.0714. Similar magnitudes do not supply the directional premise
needed to cancel error. H4 fails; it was a local diagnostic, not a generation mode.
Peak extra CUDA is 990.635 MiB under the new 1024 MiB allowance; total sampled
GPU peaks at 8967.098 MiB. Target and draft memory, CPU backing weights,
refinement store, pinned staging, scratch, KV, construction and copies are charged
separately. Traffic savings are not VRAM savings. The implementation computes
eight-bit FFNs with nonresident increments, but does not establish a larger-model
capacity frontier against optimized quantization or a long-session speedup.
Provenance and validation
- Full results, failure explanations and reproduction commands
- Frozen protocol
- Cited inspirations and exact departures, including PMPD, supplied by the owner before freeze; Any-Precision, BitStack, AnyBCQ, DecDEC and numerical-analysis antecedents. Their reported gains are not claimed for this runtime.
- Source freeze 3e0f585
- Compact receipts, reviews and tests
- Bounded delivery #38; broader milestone stays open.
Pre-inference validation: 402 CPU tests passed. Review required all-phase CUDA
accounting, independent reconstruction of archived bit planes and a forced
speculative-rejection test. A second independent audit reproduces the complete
summary. Numerical packing comparisons pass the unchanged tolerance; they are
not a claim that eight-bit quantization equals FP32. Publication requires a passing
final clean-tip suite; its separate
test receipt
and commit-bound validation record
are release assets, not part of the frozen experimental archive.
The raw archive includes all receipts, constructed representations, source snapshot
and model licensing metadata. Every member was decompressed and SHA-checked:
refinement-screen-v023-raw.zip: 1,259,987,941 bytes, 169 members,
SHA256 da6b5cf78d59d52d1663e172c2e3055b278bd51b678e4868cf46f6fb082fcc17.
This is a research prerelease, not a deployable runtime. Next hypotheses may test
statistically calibrated error prediction or high-precision prefill with explicit
KV controls; neither has been measured here. Historical failures, protocols,
stock 32B baselines and the native-pivot boundary remain unchanged.