W4A4 Quantization for Inference on Wan2.2-I2V-A14B

W4A4 Quantization for Inference on Wan2.2-I2V-A14B ICME 2026 Low-Bit-width Large-Model Quantization Challenge (Sub-Challenge 1: HiF4 / MXFP4) Yidong Chen∗ , Chengyu Shi∗ , Jiahao Liu∗ ∗ Tsinghua University

arXiv:2606.29337v1 [cs.CV] 28 Jun 2026

ICME 2026 Low-Bit-width Large-Model Quantization Challenge team submission

Abstract—We summarize our submission to Sub-Challenge 1: W4A4 Quantization for Inference (HiF4 / MXFP4) of the ICME 2026 Low-Bit-width Large-Model Quantization Challenge. The sub-challenge targets 4-bit weight and 4-bit activation (W4A4) inference on Wan-AI/Wan2.2-I2V-A14B under HiF4 or MXFP4 numerical formats. We adapt two complementary ideas from LLM quantization—MixQ-style mixed precision for sparse activation outliers and SmoothQuant-style per-channel smoothing—together with block-wise HiF4 packing for Wan2.2 feed-forward (FFN) linear layers. Calibration on representative OpenS2V-5M batches identifies heavy-tailed activation channels; smoothing rebalances dynamic range before W4A4 rounding; and a dual-branch GEMM preserves outlier columns in higher precision while the bulk of channels use strict W4A4. On official VBench I2V metrics, our pipeline stays within 2–3.5% of FP16 on most quality axes and improves motion smoothness, outperforming a native HiFloat4 baseline that degrades roughly 5% relative to FP16 across all reported scores. Index Terms—W4A4, Quantization for inference, Image-tovideo, HiF4, MXFP4, MixQ, SmoothQuant

I. I NTRODUCTION Image-to-video (I2V) diffusion models such as Wan2.2I2V-A14B [1] deliver strong generative quality but impose substantial memory and compute costs at inference time. Sub-Challenge 1 of the ICME 2026 quantization grand challenge therefore focuses on W4A4 inference: quantizing both weights and activations of linear layers to 4-bit HiF4/MXFP4compatible formats while preserving perceptual quality measured by OpenS2V-5M prompts and VBench I2V metrics. Organizers allow a limited number of Transformer blocks to remain in high precision (at most five layers for MXFP4 and two for HiF4); we target the main Wan2.2 track and recommend generation at 720×1280×61 frames. Direct W4A4 rounding on video diffusion transformers is difficult for the same reason it is difficult in large language models (LLMs): activation tensors contain heavy-tailed outliers—a tiny fraction of channels carries magnitudes far larger than the bulk distribution, which inflates per-tensor scales and destroys 4-bit signal-to-noise ratio (SNR), especially inside FFN projections. Prior LLM systems address this along two Code: https://github.com/shch-y/icme.

Fig. 1. End-to-end W4A4 inference pipeline: offline calibration, SmoothQuant fold, MixQ column split, block-packed weights, and dual-branch GEMM before VBench evaluation.

largely orthogonal axes. SmoothQuant [2] migrates quantization difficulty from activations to weights via an exact perchannel fold, making W8A8 (and, with care, W4A4) feasible without retraining. MixQ [3] splits computation into a highthroughput low-bit bulk path and a narrow high-precision outlier path, using locality-based prediction in LLMs to avoid expensive runtime outlier detection. Our submission transfers these insights to Wan2.2 I2V under a fixed inference contract: we apply SmoothQuant-style folding first, then MixQ-style column splitting on FFN linear maps, and finally block-wise HiF4 packing for weights. Fig. 1 summarizes the end-to-end stack evaluated in this report. II. BACKGROUND : O UTLIERS , S MOOTH Q UANT, AND M IX Q A. Why activations dominate W4A4 error For a linear layer Y = XW , weight tensors in Wan2.2 are comparatively smooth, whereas FFN activations exhibit sharp per-channel maxima. Fig. 2 shows calibration statistics from high_noise_model.block_00.ffn.0: activation column maxima reach above 6.0 while weight column maxima stay below 0.6, and the |x| histogram has a long tail beyond |x|>1.5. Under per-tensor or coarse per-channel W4A4, such tails set the quantizer step size and compress the majority of small-valued channels into a few bins. B. SmoothQuant recap SmoothQuant [2] introduces a positive per-channel scale s such that

Y = X ⊘s s⊙W , (1)

Fig. 4. Schematic per-channel maxima before/after SmoothQuant fold (α=0.5): activation tails shrink while corresponding weight columns grow.

Fig. 2. Wan2.2 FFN activation/weight statistics (block_00.ffn.0): heavytailed |x| and sparse high-magnitude columns motivate mixed-precision and smoothing before W4A4.

Fig. 5. MixQ-style workflow adapted to Wan2.2 W4A4: score channels, split O/J , run bulk W4A4 GEMM and a narrow high-precision branch, then sum.

Fig. 3. SmoothQuant folding preserves Y = XW in floating point while rebalancing outliers between X and W .

where ⊘ and ⊙ apply along input channels. During calibration, activation and weight magnitudes are collected and combined as act scaleα j sj = , α ∈ (0, 1). (2) weight scalej1−α Larger α pushes more dynamic range into W ; smaller α leaves more in X. At inference, s is absorbed into folded weights while activations are divided by s on the fly, reducing the activation range exactly before HiF4/MXFP4 rounding. Fig. 3 and Fig. 4 illustrate the fold and the intended per-channel rebalancing. C. MixQ-style mixed precision recap Mixed-precision quantization [3] keeps a small outlier set in high precision and quantizes the remainder to INT4/INT8 or floating 4-bit formats. For activation A with outlier channel set O, a generic mixed GEMM can be written as X Ci,j ≈ SiA SjW (Aq )i,k (Wq )j,k k∈O /

+

X

(3)

(Ahp )i,k (Whp )j,k ,

k∈O

where Aq , Wq are low-bit tensors with per-group scales S A , S W . MixQ’s original LLM system further exploits token-

level locality: outlier channels are predictable across consecutive decode steps, enabling online prediction and a quantization-ahead-of-detection fast path that avoids atomic outlier scans on most tokens [3]. Video diffusion inference differs: tokens correspond to spatio-temporal tokens across denoising steps, and the challenge mandates a fixed HiF4/MXFP4 W4A4 kernel rather than a custom INT8/FP16 outlier micro-kernel. We therefore retain the outlier split and dual-branch structure of MixQ, but calibrate a static top-k column set per FFN layer from OpenS2V5M batches, analogous to ahead-of-time outlier selection yet guided by the same maxt |xt,j | score used in MixQ’s channel view. III. M ETHOD A. Per-layer calibration protocol We collect activation and weight statistics from a small set of OpenS2V-5M calibration prompts at the organizerrecommended resolution. For each targeted FFN linear map we record per-input-channel activation maxima, weight column ranges, and optional block-wise extrema for HiF4 packing. All offline steps run once; deployed inference executes only folded weights, fixed column masks O, and W4A4/HiF4 kernels. B. SmoothQuant-style folding for Wan2.2 FFN We apply Eq. (2) layer-wise with α tuned on calibration clips (typically 0.5–0.7 for FFN up-projections where activation tails are strongest). Folded weights are re-packed into block HiF4 tiles; runtime activations are divided by the same s immediately before 4-bit activation quantize. This step is applied before MixQ splitting so that both bulk and outlier

Fig. 8. Block layouts explored for HiF4/MXFP4 weight packing after SmoothQuant fold.

Fig. 6. FFN statistics at block_29.ffn.2: outlier structure persists in deep layers, supporting a shared SmoothQuant+MixQ recipe.

latter matches GEMM memory order. Per-block exponents are stored alongside 4-bit payloads as required by HiF4/MXFP4 simulation kernels. E. Design rationale: why combine SmoothQuant and MixQ?

Fig. 7. Sorted per-channel scorej on an FFN projection: a few dominant columns justify a MixQ-style split.

branches see reduced activation tails. Deeper blocks exhibit similar but attenuated tails (Fig. 6); the same recipe is applied uniformly unless a layer is kept in FP16 under organizer rules. C. MixQ-style column split and dual-branch GEMM For X ∈ RT ×din we rank input channels by

F. Implementation notes

scorej = max xt,j

(4)

and form O as the top-k columns (Fig. 7). Let J {1, . . . , din } \ O. The layer output is

=

(q)

(5)

t

(hp)

Y = X:,J WJ ,: + X:,O WO,: ,

SmoothQuant and MixQ attack different failure modes. SmoothQuant is global and linear: it rebalances every channel but cannot fully eliminate the heaviest tails that remain after fold (Fig. 2, bottom-left). MixQ is selective: it spends highprecision budget only where tails survive smoothing. In LLM serving, MixQ additionally removes detection overhead via online prediction; in our challenge setting, static O keeps the inference graph simple while still capturing the same outlier energy that MixQ was designed to preserve. Applying smoothing before splitting also shrinks the high-precision branch: fewer columns exceed the W4A4 representable range after fold, so k can stay small. In short, SmoothQuant stabilizes the W4A4 bulk path, MixQ protects residual high-magnitude columns with small |O|, and block-wise packing provides finer scales for local weight outliers. The full stack is used for the final OpenS2V-5M submission.

where (·)(q) denotes HiF4/MXFP4 W4A4 and (·)(hp) is FP16/BF16. k is chosen to respect the challenge’s highprecision budget: a handful of columns per FFN map adds negligible FLOPs but prevents the largest tails from collapsing under 4-bit activations. Conceptually, Eq. (5) is the Wan2.2 specialization of Eq. (3) with a static O.

All three stages are implemented as offline graph transforms plus inference-time hooks in the Wan2.2 high-noise and low-noise transformer stacks. Calibration reuses the same OpenS2V-5M prompts as generation; statistics are accumulated only on FFN nn.Linear inputs after SiLU/GELU gates. SmoothQuant scales are folded into master weights before HiF4 packing, and column masks O are saved as sparse index lists per layer. At inference, activations are divided inplace by s, quantized to HiF4, and routed through either the W4A4 bulk path or a narrow FP16 outlier path. This mirrors MixQ [3], but trades online token prediction for deterministic masks that are easy to validate under challenge rules. IV. E VALUATION AND R ESULTS

D. Block-wise HiF4 weight packing After folding and optional outlier column extraction on W , remaining weights are packed with block-wise scales [4]. We evaluate 32×32 tiles and 128×1 K-major stripes (Fig. 8); the former improves SNR on outlier-heavy tiles, while the

A. Protocol We follow the challenge OpenS2V-5M evaluation split and report VBench I2V metrics: aesthetic quality, I2V subject alignment, imaging quality, motion smoothness, and subject

TABLE I VB ENCH I2V METRICS VS . FP16. PARENTHESES : RELATIVE CHANGE (%). NATIVE H I F LOAT 4 ∼5% BELOW FP16 ON EACH METRIC .

format-level 4-bit quantization with outlier-aware structure rather than treating W4A4 as uniform per-tensor rounding. ACKNOWLEDGMENT

Model

Aesthetic I2V subj. Imaging Motion Subj. cons.

FP16 baseline

0.5445 0.5173

Native HiFloat4 Ours (W4A4)

0.9626 0.9145

0.7086 0.6732

0.9730 0.9243

0.9199 0.8739

−5.0%

−5.0%

−5.0%

−5.0%

−5.0%

0.5274

0.9375

0.6936

0.9769

0.8875

−3.1%

−2.6%

−2.1%

+0.4%

−3.5%

We thank the organizers of the ICME 2026 Low-Bit-width Large-Model Quantization Challenge. Our submission builds on Wan2.2, VBench, HiFloat4 tooling, SmoothQuant, MixQ, and OpenS2V-5M resources. R EFERENCES

consistency. VBench-I2V does not provide T2V overall consistency; we omit it. Table I lists FP16, a native HiFloat4 W4A4 reference, and our full pipeline; parenthetical percentages are relative change vs. FP16. B. Discussion Native HiFloat4 W4A4 without outlier handling or smoothing loses ∼5% on every axis relative to FP16 (Table I). Our combined recipe recovers roughly half of that gap on subject and imaging metrics and surpasses FP16 on motion smoothness (+0.4%), indicating that structured mixed precision plus smoothing is more effective than uniform 4-bit rounding for video diffusion FFNs. Compared with using SmoothQuant alone, adding MixQ-style splitting specifically addresses the residual column spikes visible in Fig. 2 (bottom-left) that survive global folding. Compared with MixQ-style splitting alone, SmoothQuant reduces the number of columns routed through the high-precision branch. This branch is narrow and deterministic; it is not equivalent to leaving whole Transformer blocks in FP16, but it still makes high-precision use explicit and easy to audit under the challenge constraints. Remaining gaps to FP16 concentrate on aesthetic and subject-consistency scores, which are sensitive to fine appearance details that still clip under 4-bit activations even after fold. We observe the largest activation tails in early high-noise FFN blocks; deeper layers (Fig. 6) show similar structure with slightly lower maxima, suggesting layer-wise α and k could further close the gap without exceeding the organizer highprecision budget. Future work could port token-local outlier prediction from full MixQ [3] to spatio-temporal diffusion tokens, and explore MXFP4 block shapes under the five-layer FP16 allowance. V. R EPRODUCIBILITY Code, calibration scripts, and evaluation commands are at https://github.com/shch-y/icme; Install.md documents environment setup and VBench reproduction. VI. C ONCLUSION We described a Wan2.2 I2V W4A4 inference stack that combines SmoothQuant-style folding, MixQ-style outlier splitting, and block-wise HiF4 packing. Calibration on OpenS2V5M shows heavy FFN activation tails; smoothing plus a narrow high-precision branch mitigates them enough to beat a native HiFloat4 baseline on all reported VBench I2V metrics while approaching FP16 quality. The key lesson is to combine

[1] Wan-AI, “Wan2.2: Open large-scale video generation models,” https: //github.com/Wan-Video/Wan2.2, 2025. [2] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” in International Conference on Machine Learning (PMLR), 2023, pp. 38087–38099. [3] Yidong Chen, Chen Zhang, Rongchao Dong, Haoyuan Zhang, Yonghua Zhang, Zhonghua Lu, and Jidong Zhai, “MixQ: Taming dynamic outliers in mixed-precision quantization by online prediction,” in Proc. Int. Conf. High Performance Computing, Networking, Storage and Analysis (SC), 2024. [4] Raghuraman Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepaper,” arXiv preprint arXiv:1806.08342, 2018.

Related documents

• 14 CFR Part 73 — Special Use Airspace#505710
• Ayala Ω Framework 4.4: Energy-Resolved HVP Inference-Pipeline Study for the Muon g−2 Hadronic Vacuum Polarization Discrepancy (Research Protocol)#456558
• WSi<sub>2</sub>N<sub>4</sub>/XSe and WSiGeN<sub>4</sub>/XSe (X=Al, Ga, and In) heterostructures for photocatalytic water splitting: a DFT investigation#642453
• CVE-2026-64972 — MEDIUM 4.8#476556
• CVE-2026-8584 — MEDIUM 4.2#190211