cs.CLNov 9, 2026

How Far Does Second-Order Linear Response Predict Quantization Mismatch? Pre-Registered Measurements on Five Language Models

Authors: Dezhen Lu, Zixiang Ni

Organizations: University of Tübingen · Xi’an Jiaotong University

Abstract

Serving systems choose a weight format for each model and deployment. A low-cost screen predicts the resulting KL divergence from the unquantized model's Fisher information using 12Δ⊤FΔ\tfrac12Δ^\top FΔ, the expansion underlying Hessian-aware and second-order post-training quantizers. We assess the reliability of this screen in GPU batches with channel-wise int8, per-tensor fp8 and group-wise int4 on Qwen3-8B, Qwen3-4B, OLMo-2-7B, Qwen3-32B and Qwen3-30B-A3B. Relative predictions hold across a 16-fold range of perturbation scales, int4 group sizes and rounding modes, while absolute predictions are heavy-tailed. A trimmed average brings 10 of 12 cells on Qwen3-4B and OLMo-2 within 15% of measured KL. The MoE model exhibits nonquadratic response at small perturbations, with int8 KL scaling as α0.63α^{0.63}. Computing per-sequence gradients in bf16 can cause fourfold prediction errors. We trace these errors to arithmetic in the lower half of the backward pass and eliminate them by using fp32 there. Batch-invariant kernels run at 0.55 times the throughput, while changes in reduction order introduce a floor of 0.17-0.22 nats per sequence. We build a screener with one streamed fp32 Fisher pass, trimmed tensor-block predictions with a band, a flag for aligned tensors and per-layer format allocation. On the unseen Mistral-7B-v0.3 family, all six predictions fall within the band, all three KL budgets are met, and vLLM throughput is predicted within 0.4%.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 25, 2026

Deterministic LLM Inference Across GPU Kernels: Power-of-Two INT8 Quantization Scales and the Limits of Tolerance-Based Conformance

Conformance suites for quantized GEMM kernels ask whether two implementations agree within a tolerance. We measure what such a suite can detect. Injecting nine faults into a reference INT8 pipeline over 8,232 layer--fault--regime cells of Qwen3-1.7B, we find that every one of five epilogue faults -- scale precision, double rounding, multiplication order, output truncation, fused ordering -- moves the output by at most a single bfloat16 spacing, and by exactly one whenever it moves it at all, across 5,880 cells. A tolerance of one spacing is therefore blind to the entire class by construction: four of the five faults are detected by no check in the suite, and the fifth only under power-of-two scales. Faults that violate the accumulator's exactness preconditions, or that break operand sharing, are detected without exception, and a null fault never fires. What a tolerance-based suite of this shape establishes is therefore narrower than interchangeability: that the preconditions hold, that operands are shared, and that differences stay within one spacing. The power-of-two constraint that exposes the one detected fault is also deployable. Requantizing every weight scale to its nearest power of two makes CUTLASS and Triton agree bitwise at every linear layer (196/196 and 252/252, against 8/196 and 10/252 under the checkpoints' own scales) and yields byte-identical generated token sequences at 1.7B, 8B and 14B (8/8 prompts, against 0/8 at all three). Observed perplexity point estimates are +0.32%, -0.28% and +0.48%; the 90% intervals cover zero at the two smaller sizes but not at 14B, reaching +0.71% and +0.76%. A previously reported +157% perplexity for this intervention was an artifact of a probe that rewrote scales without requantizing the weights; separating the effects attributes 99.8% of it to the resulting weight--scale mismatch rather than to the power-of-two constraint itself.
Aug 13, 2026

The Integer Alibi: Localizing Cross-Kernel Divergence in INT8-Quantized LLM Inference

Two GPU kernels implementing the same scaled INT8 GEMM interface are usually treated as interchangeable. We test that assumption: holding the checkpoint, prompts, hardware, inference engine, decoding, and quantization configuration fixed, we swap only the INT8 linear kernel (CUTLASS versus Triton) inside vLLM. At 1.7B each arm reproduces itself bit-for-bit across cold restarts, yet the arms agree on no sequence in any end-to-end comparison we ran (0/8, 0/16, and 0/64). What makes this more than a benchmark discrepancy is an integer alibi: for shared INT8 operands under a verified no-overflow bound, the INT32 dot product is exact and order-independent, so the accumulator cannot be the source of any difference. Feeding both kernels identical operands from every linear layer of Qwen3-1.7B and 8B (196 and 252 layers), we find bit-identical outputs under power-of-two scales, confirming a pinned prediction list 196/196 and 252/252 (pre-registered at 1.7B, pinned but not blind at 8B), and observed differences of at most one bfloat16 spacing under the checkpoints' real scales. This localizes the divergence to scale application and output rounding after the exact accumulator. Applied as a probe checkpoint, the same intervention restores end-to-end bitwise agreement (8/8 and 16/16 sequences). Cross-implementation FP8 GEMM shows a different signature: both the prevalence and the magnitude of differences grow with reduction depth, while the INT8 fraction stays at parts per million and within one spacing over a 64x range of K. Teacher-forced replay ties layers to tokens: flips concentrate at small logit margins, which predict flip risk with ROC-AUC 0.94 on 16,384 positions. We will release the pre-registration, per-layer predictions, manifests with kernel-selection evidence, and a conformance procedure that turns these controls into a concrete check for kernel interchangeability.
Nov 3, 2026

Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs

A team serving a 9B-20B open-weight language model on vLLM or SGLang must choose a quantization recipe, and needs two numbers: the speedup over BF16 and the accuracy given up. We assemble every public measurement of these numbers that we could verify and find that the record supplies them only in pieces. Only two checkpoints, both INT4 builds of Gemma 3 12B, have both a vLLM timing and an accuracy score, and that score covers a single MMLU subject, one that the AWQ build was calibrated on. The batch-1 timings predate current engine releases, AMD and Intel Gaudi have no timing in this band at all, and FP4 has none against BF16. The timings that do exist follow one rule. Decoding a small batch is limited by memory bandwidth, so 4-bit weights roughly double speed (2.12x for Qwen2.5-14B on an A100), but the gain vanishes once a large batch makes the GPU compute-bound (1.02x for Gemma 3 12B at batch size 256), where FP8 is the recipe that can still help. Many published accuracy deltas, including every HumanEval delta, lie within benchmark sampling noise. We close with a seven-item reporting checklist for future measurements.