feat: CLAM/XOR/holographic adaptive codecs + holograph crate import#200
Merged
Conversation
New module: bgz_tensor::adaptive_codec Uses ndarray's ClamTree LFD (Local Fractal Dimension) to classify weight rows by compressibility: - Top 10% LFD → passthrough BF16 (hardest to compress) - Next 20% → i8 Hadamard (moderate) - Bottom 70% → i4+i2 cascade (well-clustered) - KV projections: all non-passthrough promoted to i8 (GQA sensitivity) Measured on Qwen3-TTS: k_proj (was 25% at uniform k=64): now 97% argmax at adaptive i8 q_proj: 100% argmax gate_proj: 88% argmax (up from 69%) speaker_encoder.fc: 28% — genuinely pathological, needs GPTQ Remaining blocker: GPTQ Hessian-guided rounding for tensors where CLAM finds no cluster structure (uniform high LFD = everything equally anomalous). Marked TODO in code. 3 tests pass. Burn integration research completed (report in agent). https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
XOR between row fingerprint and archetype fingerprint identifies which dimensions flipped sign. Sign-flipped dims get i8, matching dims get i4. speaker_encoder.fc: 25% → 81% argmax (the hardest tensor) MLP down_proj: 69% → 81% The XOR delta IS the anomaly detector — no Hessian, no calibration. ~39% of dimensions flip sign → get 8-bit precision ~61% matching → get 4-bit precision Combined: ~5.5 effective bits/dim, targeted where error is largest. https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
XOR-bind + majority-vote bundle stores ALL cluster residuals in one fingerprint-sized holographic memory per archetype. 3.1:1 compression. Results: holographic retrieval works (cos 0.6-0.75) but argmax-poor (0-9%) — sign-bit-only residuals lose magnitude. Need gain-shape decomposition before holographic binding (TurboQuant on residuals). Comparison at k=64: speaker_encoder.fc: CLAM 28% → XOR 81% → Holo 6% (sign-only) k_proj: CLAM 97% → XOR 97% → Holo 9% Next: accumulate indexed phase + magnitude of residual distortion into the holographic memory via TurboQuant gain-shape split. https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
Adds holograph (AdaWorldAPI/RedisGraph) as dependency for VSA
slot binding — phase AND magnitude stored in separate recoverable
slots within the same holographic memory:
Memory = Base ⊕ (SlotPhase ⊕ sign) ⊕ (SlotMag ⊕ magnitude)
Retrieve: XOR out the slot key → get the specific component
Test confirms exact slot recovery on holograph BitpackedVector.
This is the foundation for using VSA bundle/bind/XOR as holographic
weight memory instead of traditional codebooks.
Three codec approaches compared:
CLAM-adaptive: 28-100% argmax (LFD-driven precision)
XOR-adaptive: 81-97% argmax (sign-flip per-dim precision)
Holographic: 0-9% argmax (sign-only, needs magnitude — fixed
by slot encoding in next iteration)
https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
Imported AdaWorldAPI/RedisGraph/holograph as crates/holograph/ — the VSA holographic memory library with: - BitpackedVector (10K/16K/32K bit widths) - XOR bind/unbind/bundle (slot encoding for recoverable attributes) - Resonance cleanup memory (noisy → clean retrieval) - HDR cascade (stacked popcount search) - DN-Tree (256-way hierarchical addressing) - 3D XYZ superposition (content/context/relation) bgz-tensor now uses local path dep instead of git dep. All 4 holographic_residual tests pass. https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1d7f64049a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let ms = bf16_to_f32(row.matched_scale_bf16); | ||
| let mut mi = 0; | ||
| for d in 0..self.n_cols { | ||
| if !row.flipped_indices.contains(&(d as u16)) { |
There was a problem hiding this comment.
In XorAdaptiveTensor::reconstruct_row, each dimension does row.flipped_indices.contains(...), which makes reconstruction O(n_cols * n_flipped) per row. For the target 16K/32K-width vectors, this becomes a major runtime bottleneck (e.g., large flip sets turn one row decode into hundreds of millions of comparisons), so reconstruct_all can become impractically slow in production-sized runs.
Useful? React with 👍 / 👎.
Comment on lines
+65
to
+67
| let k = k.min(n).min(256); | ||
|
|
||
| let centroids = kmeans(data, k, n_cols, 10); |
There was a problem hiding this comment.
This clamps k with k.min(n).min(256) but never enforces k >= 1, so callers can pass k=0 and reach centroid-dependent code with no valid centroid, which later panics when indexing centroid-derived arrays for non-empty input. Adding an explicit validation/early return for k == 0 avoids a runtime crash on invalid-but-possible API input.
Useful? React with 👍 / 👎.
Comment on lines
+234
to
+237
| // Step 4: GPTQ compensation (if calibration data provided) | ||
| // For each column left-to-right, adjust remaining weights to minimize | ||
| // output error on calibration inputs. | ||
| // TODO: implement Hessian-guided rounding in a follow-up |
There was a problem hiding this comment.
The encode path advertises GPTQ compensation and accepts calibration_inputs, but the compensation stage is a TODO and no calibration-driven adjustment is performed. This means callers can supply calibration data and get identical output to None, which silently breaks the API contract implied by the function signature and comments and can mislead experiment results.
Useful? React with 👍 / 👎.
AdaWorldAPI
pushed a commit
that referenced
this pull request
Apr 19, 2026
Per user clarification (2026-04-19): REFINEMENT to prior IDEA CORRECTION-OF — the "no 10000-D VSA" ban is NOT workspace-wide. Three scopes legitimately preserve 10k until the coordinated rename PR: 1. Grammar prototype (role_keys + ContextChain, shipped at 10k in #210) 2. Quantum prototype (Vsa10kF32 holographic residual) 3. Ladybug-rs / bighorn imports (PRs #200-#203 cognitive stack) Elsewhere: strip 10k mentions. Files in-scope vs out-of-scope enumerated in the IDEAS entry. TECH_DEBT for the ladybug memory pathology: - Observed 700-1,100 MB runtime after #200-#203 imports at 10k - 16k rename WORSENS per-row cost 40 KB → 64 KB at f32 - Fix requires LanceDB mmap zero-copy + working-set cache policy, not wider substrate alone - Gate the 16k rename on peak-RAM measurement against Animal Farm D10 - Sparse-encoding candidate (Structured5x5 cells only) for common case https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
AdaWorldAPI
pushed a commit
that referenced
this pull request
Apr 19, 2026
CORRECTION-OF the 2026-04-19 "Ladybug 700-1100 MB memory blowup" entry. Per user: there is no 10,000 × 10,000 matrix we actually want. The blowup was a glitch — a dense 10k × 10k structure imported from outdated ladybug-rs / bighorn (PRs #200-#203) that ended up in the binary by accident. Math: - 10,000 × 10,000 × f32 = 400 MB (single allocation) - Plus cognitive-stack state → 700-1,100 MB total observed Fix: identify and DELETE the glitch allocation. Not a migration. Candidates: token-token distance matrix, co-occurrence matrix, dense attention matrix, K=10000 CLAM centroid table. High-probability locations: cognitive crate, CognitiveShader, BindSpace, CollapseGate, adaptive codecs imported from ladybug-rs without trimming. This invalidates: - "16k rename makes memory worse" — the per-row math was sound but irrelevant to this specific blowup. - Mmap zero-copy requirement — still good hygiene, not the fix here. - Sparse encoding dependency — still architecturally useful, unrelated to the glitch. 16k rename + f32 → BF16 migration proceed independently of this P0 deletion. https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Next: wire slot-encoded magnitude into holographic memory (phase+magnitude slots → full correction vector via XOR retrieval).
Test plan
cargo test -- adaptive_codec— 3 tests passcargo test -- xor_adaptive— 2 tests passcargo test -- holographic_residual— 4 tests pass (incl. holograph slot binding)https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh