{"id":"90ec4192-384b-4b0e-9268-a1ae41966e77","arxiv_id":"2607.24377","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A closed-form MXFP4 scale boundary Qmax=7.25 plus pre-normalization softmax quantization recovers nearly all FP16 video quality lost by naive MXFP4 attention.","lead":"MXAttention makes 4-bit MXFP4 attention nearly lossless for video diffusion by fixing two numerical bugs: a bad shared-scale choice and broken softmax row sums. It matters because MXFP4 is landing in mainstream accelerators, and this is a calibration-free way to use it without wrecking generation quality.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Theorem 1's \"distribution-independence\" holds only for the block-maximum distribution; the objective D(x) (Eq. 21) silently bakes in a uniform within-block density, and the empirical validation never sweeps the softmax-exponential tiles where that assumption is most wrong.","rationale":"The reader's weakest_assumption — that D(x) may be the wrong proxy — is the right place to push, and my analysis sharpens it rather than replacing it: the specific mechanism by which D(x) can mislead is its implicit uniform within-block density, which is an internal, checkable property rather than a vague \"maybe the loss landscape differs\" worry. The paper is otherwise careful: the periodicity argument (Lemma 1) is exact and correctly proved, the ∆(q) piecewise analysis in Table 6 is verifiable and consistent, PNQ's Proposition 1 is a genuine construction-level fix, and the fully-4-bit HunyuanVideo results plus the ablations provide real independent support. The residual hybrid-precision issue on Wan2.2 (fully-4-bit Imaging gap closure is ~64%, not 95%; Subject/Imaging deviate from FP16 by >0.01 in Table 3) is real but was already priced into the reader's CONDITIONAL verdict. The proposed test directly settles whether the uniform weighting is inert: if the P̃ sweep also lands on 7.25, the concern collapses and the paper's empirical story is substantially strengthened; if not, the theorem's scope needs honest qualification. Either outcome is cheap to obtain with infrastructure the authors already have. This does not warrant moving the verdict off CONDITIONAL; it identifies which condition matters most.","tokens_in":25864,"tokens_out":2888,"duration_ms":105055,"concrete_test":"Extend the §5.5.1 per-block oracle sweep (Q_max ∈ [6.00, 8.00], step 0.05) to the unnormalized softmax exponential tiles P̃ across the same Wan2.2 (40 blocks) and HunyuanVideo attention blocks, using realized tensor MSE. Separately, recompute the analytical optimum with D_w(x)=∫₀ˣ(v−Π(v))²w(v)dv / (x³·norm) using the empirical within-block density w of P̃ and of Q/K/V. If either the P̃ oracle mode or the weighted-D minimizer departs from 7.25 by ≥0.1, the uniform weighting in Eq. (21) is load-bearing and the \"data-free universal boundary\" claim should be scoped to Q/K/V-style tensors, with the softmax path requiring its own justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has a mathematical half (Q_max*=7.25 globally minimizes J(q) for any block-max law) and an empirical half (near-FP16 quality). The math is internally sound *given the objective*, but the objective is not as data-free as advertised. E(x)=∫₀ˣ(v−Π(v))²dv weights every magnitude in [0,x] equally — i.e., it assumes elements are uniformly distributed within the block up to the block max. Theorem 1 removes dependence on the distribution of M, but the uniform within-block weighting remains a distributional choice, and it is exactly the choice that determines the 7.25 root of Eq. (37). For Q/K/V after Hadamard rotation, within-block values are roughly Gaussian and concentrated well below M, so the uniform weighting overweights the tail region near x where the clipping term lives. For the unnormalized softmax tiles P̃ that PNQ quantizes, values lie in (0,1] with mass piled near 0 and at the max — maximally non-uniform. The paper's own evidence hints the uniform proxy is not inert: in the §5.5.1 MSE oracle sweep (realized tensor MSE, not the analytical objective), 7.25 is the modal optimum in only 5 of 6 model–tensor pairs, and HunyuanVideo Q prefers ≤7.15 in 45% of blocks while K concentrates at 7.20 (62.5%). Critically, the sweep covers only Q, K, and V — the softmax exponential tile, the one tensor whose within-block distribution most violates the uniform weighting, is never swept. If a within-block-weighted D(x) (or the empirical P̃ MSE) prefers a different boundary, Theorem 1 remains true but vacuous for the softmax path: the \"universal\" constant would be an artifact of the weighting in Eq. (21), and the fixed-boundary claim would rest on the same kind of proxy the paper criticizes calibration for avoiding. This is a correctness-risk concern about what the theorem establishes, not a consensus disagreement.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript proposes MXAttention, a data-free post-training quantization framework for MXFP4 attention in video diffusion models, with two components. (1) Universal Optimal Scaling (UOS): exploiting the periodized density induced by power-of-two E8M0 scales, the authors show that the derivative sign of a global quantization-error objective J(q)=E[D(X_q)], with D(x)=E(x)/x³ the normalized integrated squared E2M1 projection error, is independent of the block-maximum distribution (Lemma 1, Eqs. 31–32), and solve the resulting grid-dependent condition exactly, obtaining the closed-form boundary Q_max=29/4=7.25 with a complete global sign chart (Table 6). (2) Pre-Normalization Quantization (PNQ): quantizing the unnormalized softmax exponential tile before both the row-sum and output-accumulator updates in the FlashAttention loop, so the induced attention weights are row-normalized by construction (Proposition 1). Experiments on Wan2.2-14B and HunyuanVideo-13B show MXAttention recovering at least 95% of the VBench Imaging Quality gap between direct OCP MXFP4 and FP16 and staying within 0.01 of FP16 on all reported VBench metrics, competitive with NVFP4-based baselines. Code is integrated into MindIE-SD.","tokens_in":26288,"tokens_out":8897,"duration_ms":302286,"significance":"If the results hold, this is a useful contribution at the intersection of low-precision attention and open-standard formats. The UOS result is a genuinely parameter-free derivation: the constant 7.25 falls out of the E2M1 grid and a stated analytical objective via exact piecewise integration, not from fitting to VBench; the periodization argument (Lemma 1) is correct and cleanly removes the block-maximum law from the derivative sign. PNQ's row-sum identity is immediate but addresses a real and under-discussed failure mode of quantized online softmax, and the mechanism diagnostics (Fig. 7, Table 4) are more honest than typical. The end-to-end evidence spans two large, architecturally distinct video models, and the implementation is publicly available in MindIE-SD. Given native MXFP4 support across Blackwell, MI350, and Ascend 950, a calibration-free recipe with near-FP16 quality has clear practical value.","major_comments":[{"comment":"The 'distribution-independent' claim of Theorem 1 is proved only with respect to the block-maximum law. The objective itself, E(x)=∫₀ˣ(v−Π(v))²dv, weights all magnitudes in [0,x] uniformly, which is an implicit within-block distributional assumption, and it is this weighting that determines the 7.25 root of Eq. (37). This is not an internal inconsistency — the theorem is correct as stated — but the abstract/§4.1 language will be read as stronger than what is proved. The empirical oracle sweep (Table 4) validates realized tensor MSE for Q, K, V only; the unnormalized softmax tile P̃, whose within-block distribution (mass near 0 plus a spike at the block max) most violates the uniform weighting, is never swept, and HunyuanVideo Q already shows 45% of blocks preferring ≤7.15. Requested: (a) scope the 'data-free / distribution-independent' phrasing explicitly; (b) add a P̃-tile MSE sweep or","section":"§4.1, Eqs. (21)–(23), Theorem 1; §5.5.1, Table 4"},{"comment":"The headline Wan2.2 results use a hybrid-precision policy (Block 0 and denoising steps 38–39 retained in high precision), disclosed only in §5.1. Under fully 4-bit attention (Table 3), MXAttention on Wan2.2 reaches Imaging Quality 0.6842 versus OCP 0.5452 and FP16 0.7085 — roughly 85% of the gap, not the ≥95% quoted in the abstract. For an attention-quantization paper, needing a specific block and the final steps in high precision is material information. The abstract should disclose that the 95% figure is conditional on the hybrid policy, and the paper should explain why Wan2.2 requires it while HunyuanVideo (fully 4-bit, exceeding FP16) does not.","section":"§5.1 (Implementation Details), Table 2; Abstract"},{"comment":"The claims of 'negligible algorithmic overhead', fusion-friendliness, and competitiveness with NVFP4-based baselines are supported only by qualitative argument; the paper contains no latency, throughput, or kernel-timing measurements, and the claim that rotation and quantization can run on orthogonal compute units 'eliminating resource contention' is hardware-specific and unmeasured. Since the motivation of the work is attention acceleration, at least end-to-end and attention-kernel runtime numbers on one platform (or a clear softening of these claims) should be provided.","section":"§5.6"}],"minor_comments":[{"comment":"No prompt counts, seed counts, or variability estimates are given. The '<0.01 absolute degradation' parity claim and small margins (e.g., 0.0031 Imaging Quality below FP16 on Wan2.2; ~0.007 over NVFP4) are within plausible evaluation noise for VBench-style metrics. Please state the number of prompts/seeds and, ideally, report variance or a paired-difference confidence interval; the mechanism studies use only two prompts.","section":"§5.1, Table 2"},{"comment":"Two issues. (i) The 'w/o UOS' row does not state which boundary replaces 7.25 (presumably exact OCP); please make this explicit. (ii) The full configuration is not best on every metric: removing UOS raises Wan2.2 Aesthetic Quality (0.6176 vs 0.5994) and HunyuanVideo Imaging Quality (0.6423 vs 0.6380). The text's 'complementary gains' / 'strongest overall balance' phrasing should acknowledge these small regressions.","section":"§5.4, Table 3"},{"comment":"Since Appendix G establishes that OAS at B=32 is equivalent to the ceiling boundary q=7, an ablation row using q=7 (or a q∈{7.0, 7.25, 7.5} comparison) would isolate how much of the UOS gain is specific to 7.25 versus any boundary near 7; as Table 4 shows 7.20 is also frequent, this would strengthen the case for the exact constant.","section":"Appendix G; §5.4"},{"comment":"The row-sum diagnostic is computed on the normalized probability matrix P, while PNQ quantizes the unnormalized tile P̃. The connection is mentioned but the quantitative transfer (how row-sum error on materialized P relates to the online-softmax mismatch of Eq. (18)) deserves one clarifying sentence.","section":"§5.5.2, Figure 7"},{"comment":"Typographical: missing spaces at 'proposeMXAttention', 'First,Universal Optimal Scaling (UOS)leverages', 'we proposeMXAttention' (Conclusion). Also 'claim: less than 0.01 absolute degradation on all reported VBench metrics' is verified against Table 2 but only under the Wan2.2 hybrid policy — see Major Comment 2.","section":"Abstract/Introduction"},{"comment":"Clarify that e_elem^max=2 is the exponent of the grid value 4=2² (the wording 'corresponding to the grid value 4' is terse). In Figure 2, define the 'first harmonic', 'two harmonics', and 'concentrated' wrapped densities precisely (e.g., Fourier coefficients used) so the figure is reproducible.","section":"§2, Eq. (11); Figure 2"}],"recommendation":"minor_revision","confidential_remarks":"The mathematical core is sound and I verified the key calculations (the Δ(q) branches on [6,8] and the root at 29/4 check out exactly). The main weaknesses are in the empirical periphery: no runtime numbers in an efficiency-motivated paper, a hybrid-precision caveat on the headline Wan2.2 result that belongs in the abstract, and no MSE sweep on the softmax-exponential tile where the objective's implicit uniform within-block weighting is least defensible. All are fixable within the manuscript's scope. The authors are affiliated with a hardware vendor whose accelerator line supports MXFP4; the claims are nonetheless platform-neutral and the strongest baseline compared against is NVFP4-based, so I see no conflict affecting the technical assessment."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: two real MXFP4 failure modes, a fixed Q_max=7.25 from power-of-two periodicity, and a simple pre-normalization placement that restores row sums. On Wan2.2 and HunyuanVideo they close most of the Imaging Quality gap vs naive OCP MXFP4 and sit next to strong NVFP4 baselines.\n\nWhat is new is UOS. They show the ceiling-based scale family makes the sign of J'(q) depend only on the E2M1 grid via D(q)-D(q/2), not on the law of the block max, and pin the root at 7.25 with an exact piecewise Δ(q). That derivation is careful (periodized density, Leibniz, full sign chart in the appendix). PNQ is less deep but correct and under-used: same quantized exp tile for ℓ and the output accumulator, so row sums are one by construction. Ablations separate the pieces; mechanism plots on row-sum bias are useful. Code is public.\n\nSoft spots, in proportion. The stress-test is partly right: “distribution-independent” means independent of the block-maximum law under a fixed analytical objective. D(x) integrates squared projection error uniformly on [0,x], so within-block shape is baked in. Softmax exp tiles (mass near 0 and at 1) are never swept in the MSE oracle—only Q/K/V—and Hunyuan optima already drift slightly (7.20 vs 7.25). That does not break Theorem 1; it means the constant is universal for that proxy, not for every downstream loss. Main Wan2.2 table still uses hybrid precision on Block 0 and last steps; fully 4-bit ablations are weaker but still show large gains. No uncertainty on VBench. None of this looks load-bearing against the reported recovery.\n\nWho it is for: people shipping MXFP4 attention on video diffusion or writing low-precision FlashAttention kernels. Not a new learning principle. Math, tables, and citations look solid enough for a serious referee. I would engage—read the appendices, try the constant on your own softmax path, and push them on within-block weighting and a fully 4-bit main table. Send to peer review.","headline":"Clean systems paper: closed-form MXFP4 scale boundary plus a softmax placement fix that nearly restores FP16 video quality—math is sound for what it claims, with one real caveat on the error proxy.","tokens_in":27281,"tokens_out":579,"would_cite":true,"duration_ms":17014,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A fixed MXFP4 scaling boundary of 7.25 plus consistent softmax-path quantization recovers nearly full-precision video attention quality without calibration.","keywords":["MXFP4","attention quantization","post-training quantization","video diffusion","FlashAttention","microscaling","softmax normalization","data-free PTQ"],"falsifier":"Re-run the per-block scale sweep and end-to-end VBench comparison while optimizing a different objective (for example attention-output MSE or softmax KL) instead of the paper’s D(x); if the best fixed boundary moves well away from 7.25 or Imaging Quality no longer recovers ~95% of the FP16 gap, the central claim fails.","tokens_in":26904,"feed_emoji":"🎬","tokens_out":1031,"duration_ms":20320,"temperature":0.7,"pith_summary":"Video diffusion models spend heavily on attention, and packing that math into the open MXFP4 4-bit format usually hurts image quality. The authors argue the damage comes from two concrete numerical failures: power-of-two block scales force a bad trade-off between clipping large values and under-resolving small ones, and quantizing only part of the online softmax path breaks the rule that attention weights must sum to one. MXAttention fixes both without data or training. Universal Optimal Scaling derives a single shared-scale boundary, Qmax = 7.25, that minimizes a data-free quantization-error objective for any block-maximum distribution, because power-of-two scaling makes the optimum grid-determined rather than data-determined. Pre-Normalization Quantization then feeds the same quantized softmax exponentials into both the row-sum and the output update, so normalized weights still sum to one by construction. On two large text-to-video models the method closes at least 95% of the Imaging Quality gap to FP16, stays within 0.01 of FP16 on reported VBench scores, and matches strong proprietary 4-bit baselines with almost no extra work in a fused kernel.","feed_headline":"Fixed 7.25 scale restores 4-bit video attention quality","feed_subtitle":"Data-free MXFP4 rule plus consistent softmax math closes ~95% of the gap to FP16 on two video models.","key_machinery":"Universal Optimal Scaling (UOS): the log-periodic density of power-of-two-normalized block maxima makes the sign of the objective derivative depend only on the E2M1 grid difference D(q)−D(q/2), which vanishes uniquely at q★ = 7.25; Pre-Normalization Quantization (PNQ) then reuses one quantized exponential tile for both FlashAttention state updates so induced row sums equal one exactly.","core_discovery":"Under ceiling-based power-of-two MXFP4 scaling, the data-free objective that averages relative E2M1 projection error over the normalized block-maximum interval is globally minimized at the fixed boundary Qmax = 7.25 for any absolutely continuous block-maximum law; combined with quantizing unnormalized softmax exponentials before both the row-sum and output-accumulator updates, this yields MXFP4 attention that recovers nearly FP16 generation quality on video diffusion models without calibration or search.","pith_inferences":["The same log-periodic argument may yield closed-form optimal boundaries for other power-of-two microscaling formats beyond E2M1 MXFP4.","PNQ’s consistency principle is format-agnostic and could reduce row-sum bias in any low-bit online-softmax path, not only MXFP4.","If hardware vendors expose a programmable block-scale boundary, shipping 7.25 as the default for attention GEMMs would be a low-cost quality win."],"forward_implications":["MXFP4 attention can be deployed as pure post-training quantization with one fixed scale rule and no calibration set.","Fused FlashAttention-style kernels only need a changed scale-selection constant and reuse of the already-quantized softmax tile for the row sum.","Open MXFP4 can match or approach strong NVFP4 video-attention quality when scaling and softmax placement are handled this way.","The same UOS boundary applies across tensors, layers, and models because the optimum is grid-driven, not distribution-driven."],"fun_headline_variants":["Fixed 7.25 Qmax restores MXFP4 video attention near FP16","Data-free UOS at 7.25 plus PNQ closes 95% MXFP4 quality gap","Pre-norm softmax quant and 7.25 scale fix MXFP4 attention","MXAttention: data-free 7.25 scaling recovers FP16 video quality","Optimal power-of-two bound 7.25 enables calibration-free MXFP4"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method treats a simple relative squared projection error on the 4-bit number grid as the right stand-in for what actually matters in attention and final video quality; if a different error measure were the true target, 7.25 need not be optimal.","fun_headline_variants_meta":{"raw":{"variants":["Fixed 7.25 Qmax restores MXFP4 video attention near FP16","Data-free UOS at 7.25 plus PNQ closes 95% MXFP4 quality gap","Pre-norm softmax quant and 7.25 scale fix MXFP4 attention","MXAttention: data-free 7.25 scaling recovers FP16 video quality","Optimal power-of-two bound 7.25 enables calibration-free MXFP4"]},"model":"grok-4.5","effort":"low","cost_usd":0.004659,"raw_usage":{"total_tokens":1405,"prompt_tokens":834,"num_sources_used":0,"completion_tokens":97,"cost_in_usd_ticks":46588000,"prompt_tokens_details":{"text_tokens":834,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":474,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":834,"tokens_out":97,"duration_ms":10588,"temperature":1.0,"reasoning_tokens":474,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T16:24:57.120961+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the per-block scale sweep and end-to-end VBench comparison while optimizing a different objective (for example attention-output MSE or softmax KL) instead of the paper’s D(x); if the best fixed boundary moves well away from 7.25 or Imaging Quality no longer recovers ~95% of the FP16 gap, the central claim fails.","supporting_citations":[],"review_version":1}