Pith. sign in

REVIEW 3 major objections 6 minor 11 references

Automatic mixed precision for optimizing gained time with constrained loss mean-squared-error based on model partition to sequential sub-graphs

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A first-order Taylor expansion of loss under quantization noise, combined with measured time gains per sequential subgraph, lets integer programming pick the mixed-precision configuration that maximizes speed for a fixed loss-error budget.

desk verdict A genuinely useful hardware-aware mixed-precision selector with honest assumptions, but the 8B headline result needs a direct check of the additive loss-MSE constraint. read the letter →

arxiv 2505.13060 v1 pith:F4L7FWJD submitted 2025-05-19 cs.LG

classification cs.LG
keywords mixedprecisionpost-trainingquantizationlossMSEsensitivityintegerprogrammingsequentialsubgraphsLLMinferenceGaudi2time-to-first-token
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

The paper's aim is to make post-training mixed-precision selection automatic and hardware-aware: given a model, a calibration set, and a loss-error budget, choose which layers run in which floating-point format. It claims that two quantities can be treated as additive: the mean-squared error of the loss contributed by quantizing a layer, estimated from a first-order Taylor expansion of the loss as a function of weight and activation quantization noise, and the inference-time gain of a sequential subgraph, measured directly by running the model with each candidate configuration. With these values, an integer program selects the configuration that maximizes time, memory, or MAC-based gain while keeping loss MSE below a threshold. On Llama-3.2-1B and Llama-3.1-8B on Gaudi 2, the paper reports that this selection dominates random and prefix baselines across accuracy-versus-latency curves; for the 8B model it achieves accuracy loss below 0.1% at 450 ms time-to-first-token where baselines need about 600 ms.

What carries the argument

The carrying object is the per-layer sensitivity scalar $s_\ell = \frac{1}{R}\sum_r \|z^r_\ell \odot \dot z^r_\ell\|^2$, computed with high-precision forward and backward passes on a small calibration set. It converts quantization into an additive loss-MSE model through $d_{\ell,f}=s_\ell \alpha_f$ with $\alpha_f = 2^{-2m_f}/12$, and the group MSE is the sum of per-layer terms under an independence assumption. The second mechanism is graph partitioning: the computation DAG is split into single-entry/single-exit subgraphs whose execution is strictly sequential, and each subgraph's time gain for every candidate configuration is measured directly rather than predicted from MAC counts; the paper shows that per-layer time-gain sums badly miss the measured time gain of a parallel attention subgraph. The integer program then maximizes $\sum_{j,p} c_{j,p} i_{j,p}$ subject to $\sum_{j,p} d_{j,p} i_{j,p} \le \tau^2 \mathbb{E}[g^2]$ and exactly one selected configuration per group.

What would settle it

Quantize one early layer alone and measure end-to-end loss MSE, then one late layer alone, then both together; compare the joint measurement with the sum of the two single-layer measurements. If the joint error is not close to the sum, or if it reverses the ordering of two candidate configurations, the additive constraint that the integer program relies on fails.

Watch

Extended reading notes

Core claim

The central discovery is that mixed-precision configuration can be framed as an integer program with additive linear objectives and a single additive constraint. Each layer's quantization sensitivity $s_\ell$ is the average squared norm of the element-wise product of its extended input and its loss gradient; multiplying by $\alpha_f = 2^{-2m_f}/12$, the variance of a uniformly distributed mantissa-rounding error, gives that layer's loss-MSE contribution for format $f$. Summing these contributions over layers gives the predicted loss MSE of any configuration, and summing measured time-to-first-token gains over single-entry/single-exit sequential subgraphs gives the predicted time gain. The paper's validation on two Llama models shows the additive predictions match measurements well enough that the empirical-time variant, IP-ET, produces accuracy-versus-latency curves that dominate the random and prefix baselines on Gaudi 2.

Load-bearing premise

The load-bearing premise is that quantization noise in different layers is statistically independent, so total loss MSE is the plain sum of per-layer contributions; if errors amplify or cancel across layers, the integer program's constraint no longer tracks real quality.

Editorial extensions

If this is right

  • The IP-ET strategy yields accuracy-versus-latency curves that dominate random and prefix baselines on the tested 1B and 8B Llama models, with the 8B model keeping accuracy loss below 0.1% at 450 ms time-to-first-token versus roughly 600 ms for the baselines.
  • The same additive machinery works with alternative objectives: IP-TT maximizes MAC-based theoretical time gain and IP-M maximizes memory gain, and both usually outperform the baselines as well.
  • The loss-MSE estimate and time-gain estimate are validated on selected configurations, so the method can serve as a fast post-training quantization pipeline that needs only a small calibration set and one backward pass per sample.
  • Because the sensitivity metric is additive and the integer program is inexpensive, the approach extends naturally to many layers, formats, and model sizes without retraining.

Reading between the lines

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

  • The independence assumption behind additive loss MSE is the first thing to stress-test: if quantization errors in early layers are amplified or canceled by later layers, the ordering of configurations under the constraint could diverge from real accuracy.
  • The empirical time-gain model is hardware-specific; transferring the method to another accelerator would require re-measuring subgraph timings, though the loss-MSE side should carry over.
  • The paper's own LAMBADA-8B observation that IP-ET wins on perplexity while a baseline wins on accuracy suggests that loss MSE is a proxy tied to perplexity; users targeting one benchmark may need a task-aware threshold or a different surrogate.
  • The calibration cost of measuring every subgraph and configuration could grow for models with many parallel branches, so adaptive or sampled subgraph timing is a natural extension.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes an automatic mixed-precision (MP) selection method for post-training quantization of LLMs. The method computes a per-layer sensitivity metric from a first-order Taylor expansion of the loss with respect to quantization noise in weights and activations, then assumes the resulting loss-MSE components are additive across layers. It partitions the model graph into sequential sub-graphs, measures empirical time gains for each sub-graph and each MP configuration, and solves an integer program that maximizes a performance metric (empirical time, theoretical MAC-based time, or memory) subject to a loss-MSE threshold. Experiments on Intel Gaudi 2 with Llama-3.2-1B and Llama-3.1-8B compare the proposed strategies (IP-ET, IP-TT, IP-M) with Random and Prefix baselines, reporting accuracy-versus-latency improvements, including the headline claim that IP-ET achieves accuracy loss below 0.1% at 450 ms TTFT on 8B whereas baselines require about 600 ms.

Significance. If the claims hold, the paper makes a useful engineering contribution: it gives a concrete, hardware-aware way to select MP configurations using an additive loss-MSE proxy and measured per-subgraph time gains, and it directly validates the additivity assumptions on the 1B model. The use of 10 seeds, multiple tasks, and a public evaluation harness strengthens the empirical part. The main weaknesses are that the loss-MSE independence assumption and the time-gain additivity are validated only on the 1B model and only on configurations selected by the proposed method itself, while the headline 8B result depends on these assumptions. The method's value is also constrained by the absence of a public code release and by the task-dependent correlation between the loss-MSE proxy and final accuracy.

major comments (3)
  1. [§2.2, Eq. (23)] The loss-MSE additivity in Eq. (23) rests on the assumption that quantization noise is statistically independent across layers. This is not innocuous for sequential layers: the noisy extended input z_l in Eq. (14) contains propagated activation quantization error from earlier layers, so the cross terms in E[(sum_l (z_l)^T grad g)^2] need not vanish. The validation in Fig. 3a covers only the 1B model and only the configurations produced by IP-ET for thresholds up to 0.7%; these are exactly the configurations the additive model judges safe, so the comparison does not probe configurations where the approximation could break down. Since the IP constraint in Eq. (5) is the mechanism that controls quality, the 8B accuracy-latency result is not fully supported without either direct 8B validation of Eq. (23) or an argument that cross-layer correlations are negligible for these models.
  2. [§3.2, Fig. 3b] The empirical-time-gain additivity is validated only on the 1B model, while the headline 8B result in Fig. 5b relies on IP-ET, which selects configurations using per-group time-gain sums. If the additive model misorders configurations on the 8B model, the IP-selected configuration may be suboptimal even though the final TTFT is measured directly. Please provide a validation of predicted versus measured total time gain for 8B configurations, preferably including configurations not selected by IP-ET.
  3. [§3.4, Table 1] The paper treats loss-MSE as a proxy for accuracy, but Table 1 and Appendix C.1 show that this proxy can be misleading: on 8B LAMBADA, IP-ET has lower perplexity difference (0.922 vs. 1.075) but worse accuracy difference (-0.229 vs. -0.029) than Prefix. This does not invalidate the method, but it means the accuracy-latency comparison is task-dependent. Please report the correlation between the loss-MSE/constraint level and actual accuracy degradation, or at least discuss why the proxy is expected to track accuracy on the averaged tasks used for the headline claim.
minor comments (6)
  1. [Abstract / Introduction] There are several typographical errors, e.g., 'weighta optimizer' in the Introduction and 'ℓ∈Llin SLBGEMM' around Eq. (17).
  2. [§3.2, Fig. 3 caption] The text says Figure 3a and 3b correspond to empirical time gain and loss MSE 'respectively,' but the caption lists loss MSE in (a) and TTFT reduction in (b); the references should be aligned.
  3. [Algorithm 2] The notation 'next[vertex]' is not defined in the algorithm or the text; please define the successor function explicitly.
  4. [Appendix C.3] The second sentence refers to memory gain as 'defined in Sec. 2.3.2,' but memory gain is defined in Sec. 2.3.3.
  5. [Figure 7 caption] The caption ends with 'Random, Prefix' and omits a closing parenthesis; also, the caption for Fig. 9a/9b refers to IP-TT, but the figure appears to show IP-M results.
  6. [§2.3.2] The per-MAC time-gain parameter δ_{T,f} is introduced without specifying how it is obtained or calibrated; if it is a fixed hardware parameter, please state this explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IP tradeoff uses measured per-group gains and gradient-derived sensitivities, with validation against external accuracy and latency benchmarks.

full rationale

The derivation chain is self-contained in the relevant sense. The IP objective (Eq. 5) maximizes a sum of measured per-group time gains c_j,p (Sec. 2.3.1) subject to a loss-MSE constraint d_j,p built from per-layer gradient sensitivities s_l and the analytic quantization variance alpha_f (Eqs. 19-23); neither c_j,p nor d_j,p is fitted to the final accuracy/TTFT curves. The additivity of loss MSE is an explicit modeling assumption (Eq. 23), empirically checked against measured loss MSE on the 1B model (Fig. 3a), and the additivity of time gain is checked against measured TTFT (Fig. 3b); validating an assumption is not circular even if the assumption is debatable. The headline 8B result (IP-ET at 450 ms TTFT with accuracy loss below 0.1%) is an external measurement of accuracy on HellaSwag, LAMBADA, Winogrande, and PIQA versus measured TTFT, not a quantity encoded in the IP inputs. The only overlapping-author citation, Lee et al. 2025, appears in the opening survey list as context for FP8-on-Gaudi and is not used to justify the IP formulation or the sensitivity metric; it is therefore not load-bearing, and it is externally falsifiable as a separate hardware study. The independence-of-quantization-noise assumption behind Eq. (23) and the absence of a direct 8B loss-MSE validation are legitimate correctness/robustness concerns, but an unvalidated or imperfect assumption is not a circular step. No fitted parameter is renamed as a prediction, and no load-bearing claim reduces to a self-citation.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. It rests on four domain assumptions about quantization noise, additivity, and proxy validity, plus an unreported hardware constant for the theoretical time metric. The paper checks some of these indirectly through measured loss MSE and latency, but the checks are narrower than the claimed generality.

free parameters (1)
  • delta_T,f (per-MAC time gain for numerical format f relative to BF16)
    Appears in Eq. (24) for the theoretical time gain metric c_TT; its value is not reported and appears hardware-calibrated, so IP-TT results cannot be reproduced without recovering it.
assumptions (5)
  • domain assumption Quantization noise is small enough that a first-order Taylor expansion of the loss in the quantization errors is accurate (Eq. 17).
    The entire loss-MSE metric is derived from this expansion in Sec. 2.2; if the perturbation is large, second-order terms dominate and the predicted MSE can be wrong.
  • domain assumption Quantization noise is statistically independent across layers and elements, making loss MSE additive across layers (Eq. 23).
    This additivity is what lets the IP solve a group-wise knapsack problem; violations create correlated errors that the sum ignores.
  • domain assumption Per-element quantization error follows a scaled uniform distribution with variance |z|^2 times 2^-2m_f over 12 (Eqs. 15-16).
    The model treats FP8 and BF16 as relative-precision formats and ignores per-tensor scale effects, exponent overflows, and clipping.
  • domain assumption Loss MSE is correlated with task accuracy, so constraining it preserves accuracy (Sec. 3, first paragraph).
    The IP optimizes loss MSE rather than accuracy because accuracy is non-differentiable and non-additive; the paper notes the correlation is assumed.
  • domain assumption Execution time of the identified sequential subgraphs is additive, so measured per-group gains sum to total gain (Sec. 2.3.1).
    The whole empirical time-gain model depends on the partition producing truly sequential components; parallel, fused, or reordered kernels can break this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic mixed precision for optimizing gained time with constrained loss mean-squared-error based on model partition to sequential sub-graphs." pith.science (2026). https://pith.science/paper/F4L7FWJD

@misc{pith2026250513060,
  author       = {Pith},
  title        = {Pith review of: Automatic mixed precision for optimizing gained time with constrained loss mean-squared-error based on model partition to sequential sub-graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4L7FWJD}},
  note         = {Machine review of arXiv:2505.13060}
}
read the original abstract

Quantization is essential for Neural Network (NN) compression, reducing model size and computational demands by using lower bit-width data types, though aggressive reduction often hampers accuracy. Mixed Precision (MP) mitigates this tradeoff by varying the numerical precision across network layers. This study focuses on automatically selecting an optimal MP configuration within Post-Training Quantization (PTQ) for inference. The first key contribution is a novel sensitivity metric derived from a first-order Taylor series expansion of the loss function as a function of quantization errors in weights and activations. This metric, based on the Mean Square Error (MSE) of the loss, is efficiently calculated per layer using high-precision forward and backward passes over a small calibration dataset. The metric is additive across layers, with low calibration memory overhead as weight optimization is unnecessary. The second contribution is an accurate hardware-aware method for predicting MP time gain by modeling it as additive for sequential sub-graphs. An algorithm partitions the model graph into sequential subgraphs, measuring time gain for each configuration using a few samples. After calibrating per-layer sensitivity and time gain, an Integer Programming (IP) problem is formulated to maximize time gain while keeping loss MSE below a set threshold. Memory gain and theoretical time gain based on Multiply and Accumulate (MAC) operations are also considered. Rigorous experiments on the Intel Gaudi 2 accelerator validate the approach on several Large Language Models (LLMs).

Figures

Figures reproduced from arXiv: 2505.13060 by the authors.

Figure 1
Figure 1. Measured empirical time gain c ET j,p of the Attention sub-graph in LLAMA-3.1-8B (in blue) compared to its prediction based on the summation of per-layer time gain measurements (in orange) and for the theoretical time gain c TT j,p (green) for any of its 2 5 MP configurations. The various configurations are ordered in ascending order of empirical time gain. Configurations are labeled as 5-bit binary words which repr… view at source ↗
Figure 2
Figure 2. Layer-wise quantization patterns across MP configurations (rows) and model layers (columns) for IP-ET (top), Prefix (middle), and Random (bottom). Yellow: FP8, purple: BF16. 3.1 Setup and compared MP strategies We evaluate MP quantization during the prefill stage of LLM inference using Intel’s Gaudi 2 ac￾celerator with F = 2 numerical formats (BF16 and FP8-E4M3: 4 exponent, 3 mantissa bits), the lm-evaluation-harnes… view at source ↗
Figure 3
Figure 3. Empirical validation of the additivity assumption on different MP configurations [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Theoretical loss MSE vs. empirical time gain on the 1B model across four tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Average accuracy difference [%] vs. TTFT across HellaSwag, LAMBADA, Winogrande, and PIQA. Comparing MP quantization strategies (IP-ET, Random and Prefix) Model Strategy LAMBADA LAMBADA HellaSwag Winogrande PIQA Tasks Avg. ppl diff ↓ [%] acc diff ↑ [%] acc diff ↑ [%] ac…
Figure 6
Figure 6. Figure 6: Single-entry/single-exit sub-graphs (V1–V4) identified in one Llama-3 transformer layer. Dashed blue regions denote the latency-additive sub-graphs used in 2.3.1; residual adds are omitted for clarity. The final LM-head forms an additional single-layer sub-graph that i…
Figure 8
Figure 8. Figure 8: Average accuracy difference [%] vs. time based on number of MACs [cycles], across Hel [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 7
Figure 7. Figure 7: Per task accuracy/perplexity difference vs. TTFT. Comparing layer selection strategies for [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 9
Figure 9. Figure 9: Average accuracy difference [%] vs. total memory across HellaSwag, LAMBADA, [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 3 canonical work pages

  1. [5]

    org/abs/2103.13630

    https://arxiv. org/abs/2103.13630. Yunhui Guo. A survey on methods and theories of quantized neural networks

  2. [7]

    Raghuraman Krishnamoorthi

    URLhttps://arxiv.org/abs/1712.05877. Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342,

  3. [8]

    Nilesh Prasad Pandey, Markus Nagel, Mart van Baalen, Yin Huang, Chirag Patel, and Tijmen Blankevoort

    https://arxiv.org/ abs/2106.08295. Nilesh Prasad Pandey, Markus Nagel, Mart van Baalen, Yin Huang, Chirag Patel, and Tijmen Blankevoort. A practical mixed precision algorithm for post-training quantization.arXiv preprint arXiv:2302.05397,

  4. [9]

    Babak Rokh, Ali Azarpeyvand, and Alireza Khanteymoori. A comprehensive survey on model quantization for deep neural networks in image classification.ACM Transactions on Intelligent Systems and Technology, 2023.https://arxiv.org/abs/2205.07877. 10 Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: an adver- sarial winograd s...

  5. [10]

    doi: 10.1145/3474381

    ISSN 0001-0782. doi: 10.1145/3474381. URLhttps://doi.org/10.1145/3474381. Olivia Weng. Neural network quantization for efficient inference: A survey

  6. [11]

    org/abs/2112.06126

    https://arxiv. org/abs/2112.06126. Michael Wu, Arnab Raha, Deepak A. Mathaikutty, Martin Langhammer, and Engin Tunali. Strum: Structured mixed precision for efficient deep learning hardware codesign,

  7. [2018]

    Intel Corporation

    https: //arxiv.org/abs/1808.04752. Intel Corporation. Intel neural compressor.https://github.com/intel/neural-compressor,

  8. [2019]

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi

    URLhttps://arxiv.org/abs/1810.05723. Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language. InThirty-Fourth AAAI Conference on Artificial Intelligence,

Show all 11 references
  1. [2021]

    EleutherAI

    URLhttps://arxiv.org/abs/2110.06554. EleutherAI. Language model evaluation harness. https://github.com/EleutherAI/ lm-evaluation-harness,

  2. [2024]

    Ron Banner, Yury Nahshan, and Daniel Soudry

    URLhttps://arxiv.org/abs/2411.11038. Ron Banner, Yury Nahshan, and Daniel Soudry. Post-training 4-bit quantization of convolution networks for rapid-deployment. InAdvances in Neural Information Processing Systems (NeurIPS),

  3. [2025]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi

    URL https: //arxiv.org/abs/2501.18953. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.