Skip to content

PRIOR_ART.md

Latest commit

 

History

History
188 lines (162 loc) · 10.3 KB

PRIOR_ART.md

File metadata and controls

188 lines (162 loc) · 10.3 KB
Raw

Prior-art audit

Last searched: 2026-08-02

This is a living claim boundary, not proof that no related work exists. Repeat the search before any paper or novelty statement.

Persistent recurrent-state quantization

  • Q-Mamba evaluates W8A8H4, including 4-bit persistent Mamba2 state caches with decoupled state/channel scaling and selectivity reconstruction. RecurQuant cannot claim the first 4-bit or sub-8-bit persistent SSM-state quantizer.
  • Quamba2 and its official implementation load and store calibrated Mamba cached states at INT8 in sequence kernels. Quamba2 also searches mixed W4A8/A16 blocks, although that search is not the same as RecurQuant's per-recurrent-layer cache layout. RecurQuant cannot claim the first INT8 SSM cache, broad sensitivity-guided SSM quantization, or a first packed recurrent kernel.
  • Nemotron 3 Super studies recurrent rounding accumulation and uses FP16 stochastic rounding in its released deployment path.
  • Nemotron 3 Ultra evaluates FP16, INT8, and FP8 recurrent states with block scaling, stochastic rounding, and periodic quantized-state checkpointing plus cached-activation replay. It reports a fixed checkpoint period CC = 8 on Nemotron 3 Super; those checkpoint experiments were emulated while optimized 8-bit kernels were still under development. Fixed checkpoint/replay is prior art and a mandatory Experiment 010 stability baseline.

Existing Gated DeltaNet systems work

  • When Good Enough Is Optimal uses low- precision approximations inside quantized Gated DeltaNet computation. Its focus differs from persistent-cache storage, but it precludes a broad "first quantized Gated DeltaNet" claim.
  • KVBuffer buffers Gated DeltaNet updates to reduce state writes.
  • ReplaySSM applies checkpoint plus cached-input replay to Mamba-2 and Qwen3.5 Gated DeltaNet; its Gated DeltaNet algorithm buffers the post-correction update, normalized key, and log decay as (u, k, g). Its implementation is public. RecurQuant cannot claim the first Gated DeltaNet update buffer, checkpoint/replay cache, reduced state-write frequency, or Qwen3.5 replay implementation.
  • HOLA combines Gated DeltaNet state with a bounded exact KV cache and uses committed update magnitude as an importance signal. Update residuals and hybrid state/KV memory are therefore not novel.
  • AVMP dynamically rebalances KV and SSM memory pools.

Transition- and observability-weighted error

  • Classical finite-wordlength state-space analysis already propagates quantization-noise covariance to output noise through an observability Gramian. In particular, Hilaire, Menard, and Sentieys (EUSIPCO 2007) writes the state-noise contribution as tr(Psi_X * W_o). With diagonal error covariance, this is the same broad algebraic family as weighting each row's distortion by a diagonal observability term. RecurQuant cannot claim to have invented observability-weighted quantization error.
  • WriteSAE derives the Gated DeltaNet cache- write perturbation transition and its later-query/logit effect. It is an interpretability method rather than a quantizer, but it precludes claiming that the transition or downstream-influence principle is new.
  • TQS-PTQ treats quantized rollout error as a dynamical-system trajectory-sensitivity problem and uses the score for mixed-precision allocation. It targets offline time-series model tensors, not causal Gated DeltaNet cache rows.
  • Nemotron 3 Super and Ultra propagate recurrent-cache errors through products of state transitions. They do not use CORA's online row allocator, but they preclude a broad "first transition-aware recurrent-cache quantizer" claim.

Mixed-precision cache allocation

  • KVTuner, KVmix, and Quantize What Counts already use layer-wise sensitivity or loss signals to allocate cache precision under hardware or memory constraints.
  • RateQuant, PM-KVQ, and SpectrumKV cover fixed-average-bit, progressive, or per-token mixed-precision KV policies.
  • MixKVQ combines query relevance and intrinsic quantization difficulty for budgeted mixed-precision key-cache channels. It is the closest published analogue to multiplying a read- importance signal by physical quantization benefit.
  • Kitty dynamically promotes a fixed quota of key-cache channels using runtime sensitivity and magnitude, with Triton kernels. It targets transformer KV caches rather than recurrent matrices.
  • OuroMamba updates outlier-channel choices at every timestep and assigns higher precision dynamically for Vision-Mamba activations. It precludes a broad "first online dynamic mixed-precision SSM channel" claim.

These works target transformer KV caches rather than Gated DeltaNet recurrent matrices, but they preclude broad claims that sensitivity-guided, per-layer, dynamic, or equal-budget cache precision allocation is new.

Runtime adaptation and residual correction

  • Runtime-Certified Bounded-Error Quantized Attention computes online per-head, per-step error bounds for a tiered quantized KV cache and uses them for adaptive precision selection and deterministic fallback. It targets attention rather than a Gated DeltaNet state, but it precludes a broad first claim for runtime-checked adaptive cache precision or fallback.
  • Don't Waste Bits! uses a learned token-level controller to choose among 2-bit, 4-bit, 8-bit, and FP16 KV-cache precision during decode. Dynamic per-token precision is prior art.
  • GEAR combines low-bit KV-cache quantization with low-rank and sparse reconstruction-error correction.
  • TurboQuant combines randomized rotation and scalar quantization with a quantized residual correction for inner products.

These methods do not establish Experiment 010's result, but they prohibit broad claims that online risk selection, adaptive fallback, rotation, or residual correction is new.

Rotation and Hadamard codecs

  • QuIP# uses incoherence processing with randomized Hadamard transforms for low-bit language-model quantization.
  • QuaRot uses computational invariance and Hadamard rotations to remove activation outliers.
  • SpinQuant learns rotations for improved LLM quantization, and TurboQuant develops fast randomized transforms for outlier suppression.
  • MambaQuant applies rotation-based outlier suppression to an architecture adjacent to RecurQuant's recurrent state setting.

These methods make the transform principle established prior art. RHT-CQER-32's narrow hypothesis is whether a fixed, reproducible right-side transform composes usefully with physical causal row selection for one Gated DeltaNet cache layout. A positive result cannot support a broad first, rotation, Hadamard, or outlier-suppression novelty claim.

Architecture evidence

  • Gated DeltaNet-2 separates channel-wise erase and write gates. That distinction motivates analyzing the two state axes separately but does not establish precision allocation. Its official code uses the NVIDIA Source Code License-NC and is not copied into this Apache-2.0 repository.
  • Sparse Delta Memory changes the architecture to use sparse reads and writes over a larger state. RecurQuant instead studies post-training storage of an existing model's state.

Audited gap, not a novelty claim

The scan did not locate a published implementation combining all four:

  1. sub-8-bit persistent Gated DeltaNet cache storage;
  2. a causal transition-derived diagonal future-read score;
  3. physical row-distortion allocation with confirmation-gated admissions; and
  4. a packed end-to-end runtime kernel with measured quality and latency.

That negative search is not proof of firstness, and the current RecurQuant alpha does not satisfy item 4: it physically stores the cache in packed form but materializes one recurrent state for each layer call and has no fused recurrence kernel or latency result.

Experiment 008 also failed its frozen development gate: CORA-C2 and raw CORA were both worse on macro excess NLL than CQER-32. Exact-combination novelty would not rescue a method that has not demonstrated the required quality.

Experiment 010's StateLease-H5 protocol therefore treats checkpoint/replay, (u, k, g) buffering, rotations, dynamic precision, and residual correction as prior art. Its frozen contribution question is narrower: whether choosing between legal c4 and c5 handoff boundaries by direct local handoff distortion improves reference-aligned trajectory drift and excess NLL for a physically packed RHT-CQER Gated DeltaNet checkpoint at exactly 3,454,664 resident bytes. No Experiment 010 quality result exists at protocol freeze.

The frozen v0.2 protocol compares uniform INT4, one uniform INT8 reference, three same-byte random layer placements, MSE-selected placement, nearest and stochastic rounding, and the prespecified layer-0 policy. It does not compare against Q-Mamba's DSQ method, Nemotron checkpoint/replay, ReplaySSM, KVBuffer, per-head/per-block allocation, or fused-kernel baselines. A later claim of novelty or general superiority would require those comparisons, multiple models and tasks, a repeated prior-art search, and measured end-to-end latency.