Repository navigation
bgz-tensor: symmetric signed-i8 probe + Fisher-Z error decomposition (lab, no wire change) - #1757
Conversation
Isolated mathematical tests for a proposed -127..127 code space with -128 reserved as an invalid sentinel, measured against the existing FamilyGamma codec, which is not changed: - the code space: 255 values, one sentinel, one zero, odd negation; conventional i8 has no negation for -128 - saturating add over all 256x256 pairs: odd and commutative, sentinel propagates, valid inputs never produce it; conventional saturating_add breaks oddness on 16256 pairs - round-half-to-even gives symmetric ties; floor(x+0.5) does not - FamilyGamma: build() never stores -128, but encode() emits -128 for any input below the fitted range (cosine 0 in an all-positive family) and decodes it as a value; truncation shifts codes half a bin toward the mid code on both sides - four schemes over 13 synthetic distributions: symmetric keeps exact zeros and matches affine on centred data; affine beats it by about 2x RMSE on shifted and skewed data Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Splits FamilyGamma's round-trip error into stages that telescope exactly to the codec total (asserted per pair): input/range calibration, rounding to a code, the codec's as-i8 truncation, f32 decode, and vstanh in place of f32::tanh on the same z. On five seeded synthetic families (k=256, d=64): truncation is the largest and the only biased stage (mean |e| 2.0e-3 against 1.0e-3 for rounding in the centred family); the tanh kernel contributes about 3e-9 mean, 1.2e-7 max (<= 2 ULP from f32::tanh). A gamma fitted on 64 rows and applied to held-out rows clamps 80 pairs (27 to -128) and calibration then dominates the max error (3.6e-2). A family whose cosines all exceed the 0.9999 input clamp collapses to one z. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
… RUN Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
 |
…xed it vml_vsneg_is_not_onnx_neg_on_positive_zero pinned vsneg(+0) == +0 so it would fail once ndarray fixed the sign of zero. ndarray #368 did, and the test fails on main (Rust Tests run 40047762268). The replacement pins that vsneg and Neg both give -0.0 for +0.0 in the SIMD body and the tail. Neg stays inline: vsneg is f32-only and writes into a view. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
|
The I've put the fix in this PR: the test is replaced by Generated by Claude Code |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
…dex) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Research only. Nothing here changes
FamilyGamma,FisherZTableor any stored format.What lands
crates/bgz-tensor/tests/symmetric_i8_quantization_probe.rs: exhaustive tests for a proposed −127..127 code space with −128 reserved as a sentinel (all 256 codes, all 256×256 add pairs, tie handling), measurements of the existing FamilyGamma contract, and a comparison of four schemes over 13 seeded distributions.crates/bgz-tensor/examples/fisher_z_error_decomposition.rs: splits FamilyGamma's round-trip error into calibration, rounding, truncation, f32 decode and tanh-kernel stages that sum exactly to the total (asserted per pair)..claude/board/entries/2026-10-24-tanh-precision-symmetric-i8.md: findings, the wire-contract analysis, and a real-data manifest marked NOT RUN.Findings
build()never stores it, butencode()emits it for any input below the fitted range (cosine 0 in an all-positive family; 27 of 32,640 held-out pairs) and gives no signal. The format has no version byte and is written to safetensors by thinking-engine, so a sentinel would need a new versioned format.as i8truncation is the largest in-range error and the only biased stage: codes shift half a bin toward the mid code, about twice the rounding error.lookup_i8's 0 for an unknown centroid is not neutral there.vstanh(ndarray fix(simd,vml): simd_exp_f32 and vstanh within 1 ULP; tanh keeps ±0 and subnormals ndarray#369) stays within 2 ULP off32::tanhon every decoded z.Tests
cargo test --release -p bgz-tensor --test symmetric_i8_quantization_probe: 8/8 pass. Each of three disable runs turns the gate red: round half away from zero, no sentinel propagation, and an add clamped to −128.-D warningsclean and fmt clean.vstanhsignature, so it does not depend on ndarray#369 merging.NOT RUN: the real GGUF and safetensors families (manifest D1–D5).
Generated by Claude Code