AIAny
Icon for item

Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM

Shows that fully quantizing all 496 linear layers—including the recurrent Gated DeltaNet—of a hybrid 27B LLM to 4-bit NVFP4 W4A4 preserves benchmark accuracy while reducing model size to 17.5 GiB and improving prefill speed; includes a calibrated NVFP4 checkpoint.

Introduction

Most community recipes protected the recurrent Gated DeltaNet when applying aggressive 4-bit quantization to hybrid LLMs out of fear that recurrence would accumulate quantization noise. This paper overturns that assumption: applying NVFP4 W4A4 to every linear layer (496 total) of a Qwen3.8-27B hybrid model matches BF16 accuracy across multiple benchmarks while reducing memory and improving prefill throughput.

Key Findings
  • Full 4-bit quantization can match BF16 within seed noise on perplexity (4K/32K) and downstream tasks (MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench, RULER retrieval to 64K), so quantize-everything is a practical deployment recipe. This means smaller checkpoints and simpler serving stacks without substantial accuracy loss.
  • Quantized model size drops to ~17.5 GiB and prefill latency improves by ~14–19%, so inference cost and memory pressure are materially reduced for long-context serving.
  • The recurrent GDN gates are less sensitive than expected: nonlinearity parameterizations (softplus/exponential/sigmoid) compress GEMM error substantially, so gate errors translate to much smaller output errors.
  • A four-part mechanism explains robustness: NVFP4 block scaling localizes outliers; gate parameterizations dampen GEMM noise; the delta-rule recurrence overwrites and forgets injected impulses within hundreds of tokens; and per-token quantization cost washes out with longer contexts.
Who it's for

Great fit if you operate or research long-context hybrid LLMs and need lower-memory, faster inference with minimal accuracy regression—teams running large models on single high-memory GPUs, or anyone building compact NVFP4 checkpoints for production. Look elsewhere if your deployment hardware or kernels do not support NVFP4-style FP4 formats or if you require an end-to-end QAT solution; this work is calibration-only post-training quantization (no QAT/distillation).

Mechanistic notes

The paper also documents a serving pitfall and practical fixes: per-module-calibrated NVFP4 weights can be mis-scaled by fused GEMM kernels; shipping harmonized KV-cache scales and corrected fused-kernel handling are part of the released recipe. The authors provide a calibrated NVFP4 checkpoint for immediate evaluation and serving.

Information

  • Websitearxiv.org
  • AuthorsSergii Kozyrev, Davyd Maiboroda
  • Published date2026/09/03

More Items

Studies on-policy distillation (OPD) at the data-minimal limit by training on a single query, measuring state coverage and alignment dynamics, and showing OPD is often data-overfed but algorithm-starved.

Compresses KV cache for long-chain reasoning by keeping prompt tokens and evicting remaining entries uniformly at random per attention head; across four models and six reasoning tasks it matches the strongest prior evictor while delivering 32–43% higher vLLM throughput. Relies on prompt protection and redundancy across heads/text to retain reasoning traces; suitable when static memory budgets and higher serving throughput are priorities.

Compresses conversational histories and long documents into short sequences of continuous soft memory tokens that a frozen decoder can read directly without text reconstruction. Uses a small reader-matched writer that trains only a tiny adapter, achieving 4–16× compression and much faster write/read latencies.