Quantization

TL;DR for operators Long-context prefill creates an uncomfortable infrastructure choice. Lowering attention precision can reduce compute cost, but applying INT8 everywhere may degrade retrieval and question-answering quality. Sparse attention attacks the same cost problem differently by avoiding some interactions altogether. TileMix proposes another control surface: keep every legal attention interaction, but choose which regions of the attention-score computation run in FP16 and which run in INT8. The paper1 reports that this mixed routing can recover part of the quality lost under uniform INT8 while retaining much of the throughput advantage. On an NVIDIA A100 40GB, LLaMA 3.2 3B-Instruct at 4k tokens reaches 31.80 K tokens/s with the SpTrans75 configuration, versus 14.33 K tokens/s for FlashAttention and 29.80 K tokens/s for the matched all-INT8 path. ...

TL;DR for operators A deployment team can reach an awkward point after quantization: the checkpoint is small enough to ship, but quality is below target, and replacing the compression pipeline would mean another round of calibration, validation, packaging, and operational risk. The paper on Activation-Weighted Seeded Residual Coding, or AWSRC, asks whether some of that lost quality can instead be bought back with a small additional payload while leaving the existing low-bit reconstruction untouched.1 In its cleanest comparison, all tested residual codecs receive exactly 49,245,876 extra serialized bytes on the same RTN-SDQ parent. AWSRC reaches perplexity 7.04 and mean zero-shot accuracy 0.69—the best perplexity and a top accuracy result among the tested sparse, low-rank, learned-vector-quantized, and AWSRC repairs. Learned vector quantization has the best unrounded KL result, so the evidence supports a quality-per-byte advantage rather than dominance on every metric. ...

TL;DR for operators A small model can perform acceptably after task specialization and still lose more accuracy than expected when compressed. The paper examined here suggests that part of that loss may be shaped earlier, during fine-tuning, rather than being determined by the quantizer alone. Paper evidence: Muon-optimized models lost less accuracy after quantization than Adam-optimized models on six of eight benchmarks. On ARC-e, for example, accuracy fell by 0.55 percentage points after quantization with Muon versus 3.16 points with Adam. The integrated pipeline also produced a 2.86 GB quantized model from a 6.01 GB pre-quantized model while improving reported serving throughput and roughly halving per-token latency. ...

TL;DR for operators A team trying to move a transformer onto a phone, camera, robot, or embedded device has several levers: reduce the model, lower numerical precision, change the runtime, or move to a better accelerator. Hema Hariharan Samson’s survey of lightweight transformers1 suggests these choices cannot be evaluated independently. In the paper’s analyzed batch-size-1 setting, smaller workloads can become limited by how quickly the device moves model data rather than by raw arithmetic capacity; the paper reports roughly 60–75% hardware utilization in a favorable 15–40M-parameter range. Its broader comparisons similarly show that INT8 execution, operator fusion, optimized runtimes, and specialized accelerators change realized latency by very different amounts across devices. For an operator, the practical target is not the smallest model. It is the smallest accuracy loss that satisfies the product’s measured latency and energy budget on the actual deployment hardware. ...

TL;DR for operators Prox1 addresses a deployment problem that appears whenever sparsity itself requires computation: how much work should an inference system spend deciding which work to skip? Its answer is unusually specific. Use a cheap, input-sparse INT4 calculation to rank likely-important feed-forward channels, not to approximate their final values. Then recompute only the selected channels with the original model weights. The distinction matters empirically: at 70% effective FFN sparsity, removing the exact recomputation stage drops the aggregate downstream score from 68.6 to 44.3 on Qwen3-8B and from 74.8 to 56.7 on Qwen3-14B. ...

TL;DR for operators A lower model bit width is not automatically a speedup. A lower training loss is not automatically a reliable policy. GRINQH evaluates quantization through the mechanism it is meant to change: decoding-stage memory traffic, kernel throughput, end-to-end generation speed, and retained task accuracy.1 Kolmogorov regression evaluates diffusion policies through trajectory geometry, a PDE-based inference residual, rollout behavior, anomaly detection, and an external safety filter.2 The shared lesson is not that the two forms of “precision” are technically equivalent. They are not. The lesson is that fidelity and evidence should be allocated according to the actual failure structure of the system. A production evaluation should connect four things explicitly: the intervention, the mechanism it changes, the diagnostic that observes that change, and the operational outcome that justifies deployment. Composite scores are useful only when their weights reflect real business priorities and their components remain separately visible. Otherwise, they are merely spreadsheets wearing authority. The dashboard is not the system AI evaluation has developed an awkward habit: optimize a convenient number, improve that number, and declare the system improved. ...

TL;DR for operators Compression is usually sold as a tidy pipeline: pick a smaller architecture, prune some layers, quantize the result, then call procurement and explain why the GPU bill is still rude. This paper argues that the pipeline itself is the problem.1 The authors propose a joint compression framework for Llama-3.1-8B that searches architectural choices and quantization choices together. That means the system does not first decide “how much model” it wants and only afterward decide “how many bits” each part deserves. It treats width, depth, layer importance, weight precision, activation precision, and latency as interacting deployment variables. ...

TL;DR for operators Compression is usually sold as a clean engineering bargain: smaller model, lower memory, cheaper inference, acceptable accuracy loss. This paper asks the more operationally annoying question: after compression, does the model still know when it should hedge? The answer is: not reliably. Tong et al. benchmark compressed LLMs using conformal prediction, a framework that converts model probabilities into prediction sets with target coverage.1 In this setup, the important uncertainty metric is prediction set size: if the model needs to include more answer options to maintain coverage, it is less certain, even if its top-1 accuracy still looks respectable. ...

TL;DR for operators LoRA does not magically make LLM fine-tuning fit on phones, laptops, or small edge boxes. It reduces the number of trainable parameters. The paper’s useful contribution is showing that this is only the opening move. The real memory bill arrives from activations, checkpoint boundaries, vocabulary-sized output computations, and tokens that are being processed even though they do not contribute to the loss. Apparently the memory allocator did not attend the product strategy meeting. ...

TL;DR for operators The useful surprise in Tail-Aware HiFloat4 is not that a 4-bit video model gets worse. That part is not exactly a Nobel-level plot twist. The useful surprise is where it gets worse. The paper reports a W4A4 HiFloat4 post-training quantization pipeline for Wan2.2-I2V-A14B, and under matched generation settings the unweighted mean score drops from 0.6800 to 0.5880. But the collapse is concentrated: subject consistency falls from 0.9331 to 0.5324, while aesthetic quality is effectively unchanged, overall consistency is comparable, and motion smoothness drops only slightly from 0.9923 to 0.9803.1 ...