Pith. sign in

REVIEW 3 major objections 4 minor 42 references

FPQVAR: Floating Point Quantization for Visual Autoregressive Model with FPGA Hardware Co-design

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FPQVAR claims that low-bit floating-point quantization, with dual-format and Hadamard-based outlier suppression, brings 4-bit and 6-bit visual autoregressive image generation close to full-precision quality while running efficiently on…

desk verdict A well-executed PTQ + FPGA co-design paper that delivers a big FP4 gain on VAR (FID 3.58 vs 10.83) but with a calibration-set caveat that keeps me from fully trusting the headline number. read the letter →

arxiv 2505.16335 v1 pith:NDDQLT2K submitted 2025-05-22 cs.CV cs.AI

Renjie Wei , Songqiang Xu , Qingyu Guo , Meng Li This is my paper
classification cs.CVcs.AI
keywords floating-pointquantizationvisualautoregressivemodelpost-trainingHadamardtransformationoutlierchannelsFPGAacceleratorimagegenerationlow-bitinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to make visual autoregressive (VAR) image generation cheap enough to run on edge hardware by quantizing weights and activations to 4 or 6 bits in floating point rather than integer. The central claim is that the usual obstacles—highly imbalanced activations in one MLP layer and time-varying outlier channels in attention and MLP projections—can be removed with two algorithmic fixes: dual-format quantization, which uses separate FP grids for positive and negative values, and a group-wise Hadamard transform paired with a learned per-layer smoothing factor. On the 256x256 ImageNet benchmark, a 6-bit quantized model reaches FID 2.03 versus 1.98 for the full-precision model, and a 4-bit model improves FID from 10.83 to 3.58 over the previous quantization baseline. A custom FPGA accelerator for the 4-bit model reaches 1.1 images per second, 3.1x faster than an integer-based accelerator. The paper concludes that low-bit floating-point quantization plus transform-based outlier suppression, rather than keeping sensitive layers in high precision, is the viable path to low-cost VAR inference.

What carries the argument

The load-bearing objects are three algorithmic components and one hardware component. Dual Format Quantization (DFQ) splits an activation tensor into negative and positive parts and quantizes each with its own FP grid and scale, with the grid pair chosen offline by a coarse grid search that minimizes layer-wise MSE. Group-wise Hadamard Transformation (GHT) multiplies activations by a block-diagonal $128\times128$ Hadamard matrix to reduce time-varying outlier channels while keeping the rotation cheap enough to run online in groups. GHT-Aware Learnable Transformation (GALT) learns one scalar-per-channel vector $\lambda$ for each QKV/FC1 layer by minimizing the sum of per-step quantization errors over a calibration set, using a straight-through estimator, and then absorbs $\lambda^{-1}$ into the following weight offline. On hardware, the mechanism is a LUT-based FP4 quantizer and multiplier that converts the low-bit FP multiply into table lookups, plus a two-level pipeline that overlaps the condition-MLP of the next block with the FC2 matrix multiplication of the current block.

What would settle it

Re-run the FP4 recipe after replacing the 100-image calibration set with 100 images from a single ImageNet class, then evaluate FID on the full validation set; if FID degrades by more than about 1 point, the calibration assumption is the load-bearing part of the method.

Watch

Extended reading notes

Core claim

FPQVAR claims that a post-training floating-point quantization pipeline can compress a visual autoregressive transformer to 4-bit weights and activations with only a modest quality drop, and to 6-bit with essentially none. The paper isolates three distributional problems that break integer quantization: low-bit integer grids are uniform while network values are roughly Gaussian; the FC2 layer's input is strongly imbalanced, with 97.6% of values in $[-0.17,0]$; and the QKV/FC1 input activations contain outlier channels whose positions and magnitudes drift across the ten autoregressive steps. The proposed remedies are Dual Format Quantization (separate FP grids and scales for negative and positive activations), Group-wise Hadamard Transformation (a block-diagonal Hadamard matrix applied within 128-element groups to spread outliers), and GHT-Aware Learnable Transformation (a per-layer vector $\lambda$ optimized by straight-through gradient descent to minimize per-step quantization error, then fused into the preceding MLP weights so inference cost is unchanged). On hardware, the paper replaces DSP-heavy low-bit FP multiply-accumulate with LUT-based quantizers and multipliers, and pipelines rotation, quantization, and matrix multiplication so the group-wise transform adds no serial latency. The reported results are a 4-bit model at FID 3.58 (down from 10.83 for the previous baseline), a 6-bit model at FID 2.03 (versus 1.98 for FP16), and an FPGA accelerator reaching 1.1 images/s at 3.1x the speed of an integer-based accelerator.

Load-bearing premise

The 100-image calibration set used to choose the dual-format grids and the per-layer smoothing factors, together with the layer-wise mean-squared-error loss, must represent the true generation distribution; otherwise the reported FID/IS gains will not transfer outside that calibration set.

Editorial extensions

If this is right

  • At 6-bit weights and activations (W6A6), FPQVAR matches the FP16 teacher on ImageNet 256 (FID 2.03 vs 1.98), so full-precision quality is available at a fraction of the memory and compute.
  • At 4-bit (W4A4), FPQVAR reduces FID from 10.83 (previous baseline) to 3.58, making 4-bit VAR image generation practical rather than degenerate.
  • The KV cache can be quantized to FP6-E2M3 with negligible quality loss, further reducing memory during generation.
  • The FPGA accelerator reaches 1.1 images/s on VCK190, 3.1x faster than the integer baseline and 2.8x more energy-efficient than an A6000 GPU running FP16.
  • Because GALT's smoothing factor is absorbed into the preceding MLP weights offline, the algorithmic quality gains add zero inference-time cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The step-dependent outlier-channel finding implies that any static per-tensor or per-channel smoothing scheme will be fundamentally limited for multi-step autoregressive generation; transform-based methods that are step-invariant (like GHT) or explicitly optimize across steps (like GALT) are likely to be needed in other sequential generative models, including next-token AR image models and diffusi
  • The DFQ idea of independent negative and positive FP grids could transfer to any layer with a GeLU-like asymmetric activation (e.g., SwiGLU or SiLU) in LLMs or diffusion transformers, offering a cheap alternative to full asymmetric quantization.
  • The LUT-based FP multiplier design suggests a general recipe for low-bit FP inference on FPGAs without DSP-heavy MACs; a natural test is applying it to 4-bit FP quantization of a diffusion transformer or LLM decoder, where the same outlier phenomena appear.
  • A testable extension: run the same DFQ+GHT+GALT pipeline on a larger or different VAR backbone (e.g., VAR-d30 versus a smaller variant) to see whether the FID gains scale with model capacity, and whether the learned smoothing factors transfer across random seeds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes FPQVAR, a post-training floating-point quantization framework for visual autoregressive (VAR) image generation, with a co-designed FPGA implementation. The algorithm part makes three contributions: Dual Format Quantization (DFQ) for the imbalanced FC2 activations, Group-wise Hadamard Transformation (GHT) to reduce time-varying outlier channels at lower hardware cost than full Hadamard rotation, and GHT-Aware Learnable Transformation (GALT), which learns per-channel smoothing factors on a small calibration set. On ImageNet 256x256 with W4A4, FPQVAR reports FID 3.58/IS 241.5 versus FID 10.83/IS 175.9 for LiteVAR, and with W6A6 it reports FID 2.03 versus 1.98 for FP16. The hardware section describes LUT-based FP4 quantizers/multipliers and a two-level pipeline on AMD-Xilinx VCK190, achieving 1.1 image/s and 3.1x speedup over an integer baseline accelerator.

Significance. If the results hold, this is a strong contribution. The paper identifies concrete, well-motivated failure modes of VAR quantization and shows internally consistent ablations (DFQ improves IS from 167.0 to 189.5; adding GHT and GALT reaches 241.5 IS and 3.58 FID). The hardware is implemented and measured on FPGA rather than simulated, and the code is released. The main risk is the calibration dependence of GALT and the absence of uncertainty estimates; both are addressable with additional experiments.

major comments (3)
  1. [Section IV-C, Algorithm 3, Table VII] The headline FP4 result depends on a calibration-set fit whose robustness is not demonstrated. GALT optimizes a per-channel vector λ∈R^{1×C} (C=1920 for VAR-d30) for each QKV and FC1 layer by minimizing layer-wise quantization MSE on 100 randomly selected images, and the best λ is selected by the lowest calibration loss over 50 epochs. Table VII shows that GALT is load-bearing: it moves FID from 5.18 (DFQ+GHT) to 3.58. The paper reports a single FID/IS evaluation (50k generated images, one seed) with no error bars and no sensitivity analysis to the choice of the 100 calibration images. Because tokens from the same image are not independent and the early time steps contain very few tokens, the effective number of independent calibration points is small relative to the number of fitted parameters. I ask the authors to add calibration-set sensitivity experiments (e.g., several independent draws of the 100 images, reporting mean and standard deviation of FID/IS) and/or to select λ on a validation split that is not used for reporting final metrics. Without this, the 3.58 FID could reflect overfitting to one calibration draw.
  2. [Equation (5)] The optimization objective for GALT is not dimensionally consistent as written. λ∈R^{1×C}, so the second quantized term Q(H_B^T λ^{-1} W^T) is the product of a C×C matrix with a 1×C vector, which is undefined; the first term Q(X_i λ H_B) is also ambiguous if λ is a row vector. Presumably the intended operation is channel-wise scaling, i.e., Q(X_i diag(λ) H_B) Q(H_B^T diag(λ^{-1}) W^T) (or an equivalent element-wise formulation). Please correct Eq. (5) and the absorption argument in Eq. (6), since this equation defines GALT.
  3. [Section V-B, Table III] The claim that FP6 FPQVAR is 'on par' with FP16 rests on a FID difference of 0.05 (2.03 vs 1.98) and an IS difference of 7.0 (295.0 vs 302.0). With a single evaluation seed and no confidence intervals, it is not possible to judge whether these differences are within noise. Please report at least two or three seeds (or bootstrapped FID confidence intervals) for the main configurations; this also bears on the FP4 comparison, although the gap there is much larger.
minor comments (4)
  1. [Section IV-C] The text refers to Fig. 11b for the calibration dataset construction, but Fig. 11 shows the GALT loss curves; the construction appears to be in Fig. 5b.
  2. [Table VII] The header 'Nomalized Throughput' should be 'Normalized Throughput'.
  3. [Footnote on code release] The GitHub URL reads 'FPQV AR' with a space; please ensure the link points to 'FPQVAR'.
  4. [Section V-A] The text states that all methods are evaluated under the same random seed; for completeness, state how many images are used for the calibration set and whether the 100 images are drawn from the training set.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: headline FID/IS results are evaluated on independently generated 50k images, not on the calibration objective used to fit DFQ formats and GALT smoothing factors.

full rationale

The paper's central claims (FID/IS improvements and accelerator throughput) are measured against quantities independent of any fitted input: IS/FID are evaluated over 50,000 freshly generated images, while the DFQ format choice and per-layer GALT smoothing factors λ are optimized to minimize layer-wise output quantization MSE on a calibration set of 100 images (Alg. 2 and Alg. 3, Sec. IV-C). The GALT objective in Eq. 5 is a quantization-error proxy, not the reported generation-quality metric, so the reported FID 3.58 vs 10.83 does not reduce by construction to the calibration loss. DFQ format search selects among a small set of FP4 grids by MSE, and GALT fits high-dimensional λ, but these are post-training optimization procedures whose final evaluation is external to the fitted objective. The only citations to overlapping-author prior work are [22] for SFU-inspired design and [37] for a 128-point Fast Hadamard transformation unit; these are hardware implementation inspirations and are not load-bearing premises for the quantization claims, nor do they import an unverified uniqueness theorem. Potential sensitivity of the 100-image calibration set is a robustness or overfitting concern, not circularity, because the test distribution is independent of the calibration distribution. No step in the derivation chain is equivalent by definition to its own inputs; the modest score reflects only the presence of minor overlapping-author citations that do not support the main results.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

DFQ, GHT, GALT are algorithmic procedures and hardware units (GHTU, FMU) rather than new scientific entities; they introduce no new particles, forces, or conserved quantities. The free parameters are calibration-time fitted values that directly influence the reported quality numbers.

free parameters (5)
  • GALT smoothing factor λ = per-layer vectors, values not reported
    Learned per QKV projection and FC1 layer (Alg 3) on 100 calibration samples to minimize quantization MSE in Eq 5; directly shapes the final FP4 quality.
  • DFQ format pair for FC2 = FP4: negative grid from E1M2, positive grid from E2M1
    Searched offline on calibration set (Alg 2) to minimize FC2 input-activation quantization MSE; affects all FP4 results.
  • FP format selection (W/A) = FP6: W-E2M3/A-E3M2; FP4: W-E2M1/A-E2M1
    Chosen by the authors after preliminary experiments as the best uniform format; a model and hyperparameter choice tied to the VAR-d30 model.
  • GHT group size = 128
    Block-diagonal Hadamard group size set to 128 to match quantization group size; hardware efficiency choice, affects outlier reduction quality.
  • GALT optimization hyperparameters = 100 samples, 50 epochs, LR 0.01, AdamW
    Hand-picked; no sensitivity analysis is provided, so the reported gains could depend on these choices.
assumptions (5)
  • standard math Hadamard matrix orthogonality (H H^T = I) preserves the linear layer output when applied to both activation and weight.
    Invoked in Eq 4 and in GHT (Sec IV-B) to justify rotating activations and weights without changing the FP16 result.
  • domain assumption FP quantization rules from OCP MX Specification are a valid low-bit arithmetic model.
    Defines encoding in Eq 2 and FP4 formats in Table I; the paper does not validate these formats against other FP variants.
  • domain assumption Straight-Through Estimator gives a usable gradient through the nondifferentiable quantizer in GALT.
    Used in Sec IV-C, Alg 3 to optimize λ; standard in QAT/PTQ literature.
  • domain assumption Minimizing per-layer output MSE on calibration activations is a sufficient proxy for preserving end-to-end FID/IS.
    GALT's objective (Eq 5) optimizes quantization MSE, while success is measured by IS/FID on 50k generated images; no analysis links the two.
  • domain assumption Classifier-free guidance and the pretrained VAR-d30 weights are fixed and representative of the ImageNet conditional generation task.
    All experiments use one pretrained model and CFG; claims may not extend to other VAR scales or tokenizers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FPQVAR: Floating Point Quantization for Visual Autoregressive Model with FPGA Hardware Co-design." pith.science (2026). https://pith.science/paper/NDDQLT2K

@misc{pith2026250516335,
  author       = {Pith},
  title        = {Pith review of: FPQVAR: Floating Point Quantization for Visual Autoregressive Model with FPGA Hardware Co-design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDDQLT2K}},
  note         = {Machine review of arXiv:2505.16335}
}
read the original abstract

Visual autoregressive (VAR) modeling has marked a paradigm shift in image generation from next-token prediction to next-scale prediction. VAR predicts a set of tokens at each step from coarse to fine scale, leading to better image quality and faster inference speed compared to existing diffusion models. However, the large parameter size and computation cost hinder its deployment on edge devices. To reduce the memory and computation cost, we propose FPQVAR, an efficient post-training floating-point (FP) quantization framework for VAR featuring algorithm and hardware co-design. At the algorithm level, we first identify the challenges of quantizing VAR. To address them, we propose Dual Format Quantization for the highly imbalanced input activation. We further propose Group-wise Hadamard Transformation and GHT-Aware Learnable Transformation to address the time-varying outlier channels. At the hardware level, we design the first low-bit FP quantizer and multiplier with lookup tables on FPGA and propose the first FPGA-based VAR accelerator featuring low-bit FP computation and an elaborate two-level pipeline. Extensive experiments show that compared to the state-of-the-art quantization method, our proposed FPQVAR significantly improves Fr\'echet Inception Distance (FID) from 10.83 to 3.58, Inception Score (IS) from 175.9 to 241.5 under 4-bit quantization. FPQVAR also significantly improves the performance of 6-bit quantized VAR, bringing it on par with the FP16 model. Our accelerator on AMD-Xilinx VCK190 FPGA achieves a throughput of 1.1 image/s, which is 3.1x higher than the integer-based accelerator. It also demonstrates 3.6x and 2.8x higher energy efficiency compared to the integer-based accelerator and GPU baseline, respectively.

Figures

Figures reproduced from arXiv: 2505.16335 by the authors.

Figure 1
Figure 1. VAR model architecture (a,b,c,d), its next-scale prediction (a), [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Weight distribution of QKV projection layer in the 9-th [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Activation distributions of the FC1 and FC2 layer and the corre [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Outlier channels (marked by red triangles) in VAR vary with the progression of time steps. Note that the number of tokens in the figure [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Activation distribution of the 9-th QKV projection layer at Step [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: (a) Quantization error of FC2 input activation with different [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Hardware architecture of our VAR accelerator. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: The proposed two-level pipeline. TABLE III: Quantitative comparison between different quantization methods on ImageNet 256 × 256 benchmark. IS ↑ FID ↓ Precision ↑ Recall ↑ FP16 302.0 1.98 0.82 0.60 INT6 RTN 253.6 2.99 0.76 0.63 INT6 LiteVAR 257.0 3.04 0.77 0.62 FP6 bas…
Figure 9
Figure 9. Figure 9: Qualitative comparison of 256 × 256 images generated by original FP16 VAR and different quantization methods. The class labels are (from top to bottom): goldfish (1), Maltese dog (153), palace (698), flowerpot (738). FP16 INT6 LiteVAR FP6 baseline FP6 FPQVAR (ours) INT…
Figure 10
Figure 10. Figure 10: Qualitative comparison of 512 × 512 images generated by original FP16 VAR and different quantization methods. The class labels are (from top to bottom): tiger shark (3), lion (291), barn (425), convertible (511) [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Loss curve of our GALT optimization for the (a) QKV [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 13 canonical work pages

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  2. [2]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

    arXiv 2010

  3. [3]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  4. [4]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4195–4205

    2023

  5. [5]

    Generative pretraining from pixels,

    M. Chen, A. Radford, R. Child, J. Wu, H. Jun, D. Luan, and I. Sutskever, “Generative pretraining from pixels,” inInternational conference on machine learning. PMLR, 2020, pp. 1691–1703

    2020

  6. [6]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 873–12 883

    2021

  7. [7]

    Janus: Decoupling visual encoding for unified multimodal understanding and generation,

    C. Wu, X. Chen, Z. Wu, Y . Ma, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, C. Ruanet al., “Janus: Decoupling visual encoding for unified multimodal understanding and generation,”arXiv preprint arXiv:2410.13848, 2024

    arXiv 2024

  8. [8]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling,

    X. Chen, Z. Wu, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, and C. Ruan, “Janus-pro: Unified multimodal understanding and generation with data and model scaling,”arXiv preprint arXiv:2501.17811, 2025

    arXiv 2025

Show all 42 references
  1. [9]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,”Advances in neural information processing systems, vol. 37, pp. 84 839–84 865, 2024

    2024

  2. [10]

    Smoothquant: Accurate and efficient post-training quantization for large language models,

    G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 38 087–38 099

    2023

  3. [11]

    Outlier suppression: Pushing the limit of low-bit transformer language models,

    X. Wei, Y . Zhang, X. Zhang, R. Gong, S. Zhang, Q. Zhang, F. Yu, and X. Liu, “Outlier suppression: Pushing the limit of low-bit transformer language models,”Advances in Neural Information Processing Systems, vol. 35, pp. 17 402–17 414, 2022

    2022

  4. [12]

    Quarot: Outlier-free 4-bit inference in rotated llms,

    S. Ashkboos, A. Mohtashami, M. Croci, B. Li, P. Cameron, M. Jaggi, D. Alistarh, T. Hoefler, and J. Hensman, “Quarot: Outlier-free 4-bit inference in rotated llms,”Advances in Neural Information Processing Systems, vol. 37, pp. 100 213–100 240, 2024

    2024

  5. [13]

    Spinquant: Llm quantization with learned rotations,

    Z. Liu, C. Zhao, I. Fedorov, B. Soran, D. Choudhary, R. Krishnamoorthi, V . Chandra, Y . Tian, and T. Blankevoort, “Spinquant: Llm quantization with learned rotations,”arXiv preprint arXiv:2405.16406, 2024

    2024 arXiv

  6. [14]

    Q-diffusion: Quantizing diffusion models,

    X. Li, Y . Liu, L. Lian, H. Yang, Z. Dong, D. Kang, S. Zhang, and K. Keutzer, “Q-diffusion: Quantizing diffusion models,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17 535–17 545

    2023

  7. [15]

    Temporal dynamic quanti- zation for diffusion models,

    J. So, J. Lee, D. Ahn, H. Kim, and E. Park, “Temporal dynamic quanti- zation for diffusion models,”Advances in neural information processing systems, vol. 36, pp. 48 686–48 698, 2023

    2023

  8. [16]

    Q-dit: Accurate post-training quantization for diffusion transformers,

    L. Chen, Y . Meng, C. Tang, X. Ma, J. Jiang, X. Wang, Z. Wang, and W. Zhu, “Q-dit: Accurate post-training quantization for diffusion transformers,”arXiv preprint arXiv:2406.17343, 2024

    2024 arXiv

  9. [17]

    Ptq4dit: Post-training quantization for diffusion transformers,

    J. Wu, H. Wang, Y . Shang, M. Shah, and Y . Yan, “Ptq4dit: Post-training quantization for diffusion transformers,”arXiv preprint arXiv:2405.16005, 2024

    2024 arXiv

  10. [18]

    Litevar: Compressing visual autoregressive modelling with efficient attention and quantization,

    R. Xie, T. Zhao, Z. Yuan, R. Wan, W. Gao, Z. Zhu, X. Ning, and Y . Wang, “Litevar: Compressing visual autoregressive modelling with efficient attention and quantization,”arXiv preprint arXiv:2411.17178, 2024

    2024 arXiv

  11. [19]

    Fp8 quantization: The power of the exponent,

    A. Kuzmin, M. Van Baalen, Y . Ren, M. Nagel, J. Peters, and T. Blankevoort, “Fp8 quantization: The power of the exponent,”Advances in Neural Information Processing Systems, vol. 35, pp. 14 651–14 662, 2022

    2022

  12. [20]

    Llm-fp4: 4-bit floating-point quantized transformers,

    S.-y. Liu, Z. Liu, X. Huang, P. Dong, and K.-T. Cheng, “Llm-fp4: 4-bit floating-point quantized transformers,”arXiv preprint arXiv:2310.16836, 2023

    2023 arXiv

  13. [21]

    Flightllm: Efficient large language model inference with a complete mapping flow on fpgas,

    S. Zeng, J. Liu, G. Dai, X. Yang, T. Fu, H. Wang, W. Ma, H. Sun, S. Li, Z. Huanget al., “Flightllm: Efficient large language model inference with a complete mapping flow on fpgas,” inProceedings of the 2024 ACM/SIGDA International Symposium on Field Programmable Gate Arrays, 2...

    2024

  14. [22]

    Hg-pipe: Vision transformer acceleration with hybrid-grained pipeline,

    Q. Guo, J. Wan, S. Xu, M. Li, and Y . Wang, “Hg-pipe: Vision transformer acceleration with hybrid-grained pipeline,” inProceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, 2024, pp. 1–9

    2024

  15. [23]

    Flightvgm: Efficient video generation model inference with online sparsification and hybrid precision on fpgas,

    J. Liu, S. Zeng, L. Ding, W. Soedarmadji, H. Zhou, Z. Wang, J. Li, J. Li, Y . Dai, K. Wenet al., “Flightvgm: Efficient video generation model inference with online sparsification and hybrid precision on fpgas,” in Proceedings of the 2025 ACM/SIGDA International Symposium on Fi...

    2025

  16. [24]

    Pushing up to the limit of memory bandwidth and capacity utilization for efficient llm decoding on embedded fpga,

    J. Li, T. Li, G. Shen, D. Zhao, Q. Zhang, and Y . Zeng, “Pushing up to the limit of memory bandwidth and capacity utilization for efficient llm decoding on embedded fpga,”arXiv preprint arXiv:2502.10659, 2025

    2025 arXiv

  17. [25]

    Generating diverse high- fidelity images with vq-vae-2,

    A. Razavi, A. Van den Oord, and O. Vinyals, “Generating diverse high- fidelity images with vq-vae-2,”Advances in neural information processing systems, vol. 32, 2019

    2019

  18. [26]

    Autoregressive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregressive image generation using residual quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 11 523–11 532

    2022

  19. [27]

    Movq: Modulating quantized vectors for high-fidelity image generation,

    C. Zheng, T.-L. Vuong, J. Cai, and D. Phung, “Movq: Modulating quantized vectors for high-fidelity image generation,”Advances in Neural Information Processing Systems, vol. 35, pp. 23 412–23 425, 2022

    2022

  20. [28]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representation learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  21. [29]

    Hart: Efficient visual generation with hybrid autore- gressive transformer,

    H. Tang, Y . Wu, S. Yang, E. Xie, J. Chen, J. Chen, Z. Zhang, H. Cai, Y . Lu, and S. Han, “Hart: Efficient visual generation with hybrid autore- gressive transformer,”arXiv preprint arXiv:2410.10812, 2024

    2024 arXiv

  22. [30]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,

    J. Han, J. Liu, Y . Jiang, B. Yan, Y . Zhang, Z. Yuan, B. Peng, and X. Liu, “Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,”arXiv preprint arXiv:2412.04431, 2024

    2024 arXiv

  23. [31]

    Integer or floating point? new outlooks for low- bit quantization on large language models,

    Y . Zhang, L. Zhao, S. Cao, S. Zhang, W. Wang, T. Cao, F. Yang, M. Yang, S. Zhang, and N. Xu, “Integer or floating point? new outlooks for low- bit quantization on large language models,” in2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2024, pp. 1–6

    2024

  24. [32]

    Afpq: Asymmetric floating point quantization for llms,

    Y . Zhang, S. Zhang, S. Cao, D. Du, J. Wei, T. Cao, and N. Xu, “Afpq: Asymmetric floating point quantization for llms,”arXiv preprint arXiv:2311.01792, 2023

    2023 arXiv

  25. [33]

    Optimizing large language model training using fp4 quantiza- tion,

    R. Wang, Y . Gong, X. Liu, G. Zhao, Z. Yang, B. Guo, Z. Zha, and P. Cheng, “Optimizing large language model training using fp4 quantiza- tion,”arXiv preprint arXiv:2501.17116, 2025

    2025 arXiv

  26. [34]

    Microscaling data formats for deep learning,

    B. D. Rouhani, R. Zhao, A. More, M. Hall, A. Khodamoradi, S. Deng, D. Choudhary, M. Cornea, E. Dellinger, K. Denolfet al., “Microscaling data formats for deep learning,”arXiv preprint arXiv:2310.10537, 2023

    2023 arXiv

  27. [35]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  28. [36]

    Sda: Low-bit stable diffusion acceleration on edge fpgas,

    G. Yang, Y . Xie, Z. J. Xue, S.-E. Chang, Y . Li, P. Dong, J. Lei, W. Xie, Y . Wang, X. Linet al., “Sda: Low-bit stable diffusion acceleration on edge fpgas,” in2024 34th International Conference on Field-Programmable Logic and Applications (FPL). IEEE, 2024, pp. 264–273

    2024

  29. [37]

    Lightmamba: Efficient mamba acceleration on fpga with quantization and hardware co-design,

    R. Wei, S. Xu, L. Zhong, Z. Yang, Q. Guo, Y . Wang, R. Wang, and M. Li, “Lightmamba: Efficient mamba acceleration on fpga with quantization and hardware co-design,”arXiv preprint arXiv:2502.15260, 2025

    2025

  30. [38]

    Estimating or propagating gradients through stochastic neurons for conditional computation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,”arXiv preprint arXiv:1308.3432, 2013

    2013 arXiv

  31. [39]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,”Advances in neural information processing systems, vol. 29, 2016

    2016

  32. [40]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017

    2017

  33. [41]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  34. [42]

    Flashattention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “Flashattention-2: Faster attention with better parallelism and work partitioning,”arXiv preprint arXiv:2307.08691, 2023

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.