- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
QUASAR studies attention acceleration for video Diffusion Transformers. It scores KV blocks using mean-pooled Q/K representations, computes attention exactly on a Top-K subset, and approximates the remaining numerator and denominator using K/V block summaries before joint normalization. Recovery weights can be shared across an entire query block, shared within smaller groups, or computed per query. These choices vary the extent of query sharing while keeping the exact selected set and KV summaries fixed. Selection can operate over all legal blocks or over candidates supplied by another method; blocks outside the candidate set receive no Recovery.
The paper analyzes candidate coverage and attention-output error and evaluates implementations across models, backends, and hardware. On Wan2.1 at 480p, it reports an AVG5 of 0.8301, compared with 0.8347 for dense attention, at 84.77% sparsity and a 1.54× E2E speedup. At 720p, the fastest per-query configuration achieves a 2.41× speedup with an AVG5 of 0.8215, compared with 0.8268 for dense attention. Ablations show that summary-based recovery improves over dropping unselected blocks in the evaluated settings.
Critical strengths
-
The ablations provide good evidence for the role of recovery. In Table 3, under the same selection configuration and at 89.84% sparsity, per-query Recovery raises AVG5 from 0.7926 to 0.8270 while E2E latency changes from 167.832 s to 167.957 s. Similar gains occur at 84.77% sparsity. Table 5 holds Q/K/V and the exact selected set fixed and shows lower attention-output error with recovery; the appendix also reports recovery gains with another seed. These results help attribute the quality improvement to recovery rather than changes in selection.
-
The implementation is fairly complete, and the reported acceleration includes the main online costs. The paper describes summary computation, online-softmax state merging, and the kernel implementation of per-query recovery. E2E timing includes candidate generation, selection, layout conversion, Recovery, and VAE decoding, making the reported acceleration practically relevant. Complete attention-call breakdowns, execution tests across backends, and validation on a second model strengthen the evidence beyond isolated kernel timings in a single configuration.
Critical weaknesses
-
The distinction from existing compensation methods is not sufficiently established. Exact computation on selected blocks combined with summary compensation for the remaining blocks is directly related to PISA’s exact-plus-approximate framework and SVG-EAR’s centroid compensation. Showing that recovery improves over dropping unselected blocks does not establish what the method adds over existing compensation approaches. Fixed-block summaries, recovery granularity, and execution choices may offer advantages, but their differences from prior methods and the significance of those differences require a clearer comparison.
-
Multi-granularity primarily refines the accuracy–cost trade-off. The different granularities change the scope of query sharing while retaining the same recovery formula and KV summaries. The paper theoretically explains how reducing query sharing improves the approximation and provides supporting experiments. These results establish a useful refinement, but do not establish a major methodological contribution distinct from the recovery mechanism. Its contribution to the overall novelty is therefore limited.
-
The practical case for candidate reuse remains unclear. Candidate generation adds overhead, and blocks outside the candidate set receive no Recovery. In the representative Wan comparisons, candidate modes are slower and yield lower quality than Full-Domain. The evidence therefore primarily establishes interface compatibility. The paper should explain when candidate reuse is preferable in practice, since scoring fewer blocks does not by itself improve overall inference performance.
-
The theory explains the method but contributes less to the overall novelty. The analysis identifies errors arising from query sharing, within-block variation, and normalization, and the paper numerically validates its bounds. These results are useful. However, the tighter bounds mainly support offline analysis, while the practical guidance provided by the input-summary-based bounds remains unclear. I regard the theory as supporting the method’s rationale, but do not yet see a sufficiently significant new analytical result to change my assessment of the contribution.
Questions for authors
-
What are the main mathematical and implementation differences between Recovery, SVG-EAR’s centroid compensation, and PISA’s approximation? Do these differences reduce overhead, broaden applicability, or improve the quality–latency trade-off under comparable settings? Please focus on the distinction that most clearly establishes the paper’s added value.
-
In the representative Wan comparisons, candidate modes are slower and yield lower quality. Under what deployment conditions are they preferable to Full-Domain? If their practical value depends on candidate indices already being available at no additional generation cost, please clarify where that assumption applies.
-
For the input-summary-based bounds that do not require exact attention information, what do the existing results establish about their practical tightness? Can these bounds help determine when finer-grained recovery is worthwhile, or is their current role primarily to provide a qualitative explanation?
4 comments
Response of the Review from Reviewer japD (Part 2/3)
Q3.1 For the input-summary-based bounds that do not require exact attention information, what do the existing results establish about their practical tightness?
A3.1 We answer your question through the following logical chain: “What tightness conclusions do existing results establish?”—“A review of the input-summary-based bound”—“How tight is this bound, and where is the margin?”—“How should the bound be tightened based on the margin, and what optimization direction follows?”
Existing results and additions in this round. Main text §3.3 Theorem 2 and Eqs. (10)–(11) provide the output-error decomposition and hierarchical bounds; K.13 gives the input-summary analysis; Appendix J Table 29(b) reports bound tightness. With Q/K/V, the Exact set, and granularity fixed, we quantified the margin from substituting Eq. (87) into Eq. (85).
The table reports “normalized upper bound − actual attention error” for G1, G2, and per-query, using the same Dense denominator across resolutions.
| Upper Bound | 480p Gap | 720p Gap |
|---|---|---|
| Input Summary | 6.954824–6.959285 | 7.995186–8.006168 |
| HCR | 0.230101–0.242398 | 0.330051–0.333757 |
| Pairwise | 0.007868–0.009571 | 0.026900–0.029461 |
The input summary bound is 37.52–47.33 times the actual error; the Pairwise bound, which preserves more complete output information, is only 5.09%–14.17% higher.
Appendix K.13 of the paper provides a complete derivation of the input bounds, namely Eqs. (84), (85), and (87); all subsequent tightness analyses are based on these formulas.
Where is the margin? On the same query/head set , take the square sum as per Eq. (81), take the square root, and divide by . The two margins are:
The margin gained by substituting the input statistics for the actual logit range:
\sqrt{\sum_{i\in\mathcal I}[\overline D_{V,D}\tanh(\epsilon_i/2)]^2} -\sqrt{\sum_{i\in\mathcal I}[\overline D_{V,D}\tanh(h_i/4)]^2} }{\|O_{\mathcal I}^{\Omega}\|_F} \approx 0.002856\text{–}0.004623.$$ The margin from the actual logit range to the output error: $$\frac{\sqrt{\sum_{i\in\mathcal I}[\overline D_{V,D}\tanh(h_i/4)]^2}-\Vert\widehat O_{\mathcal I}-O_{\mathcal I}^{\Omega}\Vert_F}{\Vert O_{\mathcal I}^{\Omega}\Vert_F}\approx 6.951134\sim 8.002457.$$ The range comes from two captures and three granularities. **The main margin lies in the transformation from actual logit range to output error**, including the slack from replacing $D_{V,D}$ with $\overline D_{V,D}$; replacing it with actual $h_i$ produces only a small reduction. This explains the tightening direction: Pairwise/HCR bounds preserve more output-distribution or Value-geometry information and lie closer to the true error; the main opportunity is output-side estimation rather than redefining input-summary terms. **Future work should prioritize tightening the output-side estimate in Eq. (85)** by using finer Value geometry and weight information from the input summary; Pairwise/HCR provide offline references. **Conclusion: The margin lies primarily on the output side; preserving Value geometry and error cancellation significantly tightens the bounds for the same output, providing direction for optimizing input bounds.**Response of the Review from Reviewer japD (Part 3/3)
Q3.2 Can these bounds help determine when finer-grained recovery is worthwhile, or is their current role primarily to provide a qualitative explanation?
A3.2 We answer four questions in sequence: (1) where the effect of Recovery stops; (2) how much upper-bound improvement the input bound predicts for each Recovery granularity and which positions are better suited to that granularity; (3) under the same budget, how many times the actual improvement from bound-based selection exceeds random selection; and (4) where the theoretical guidance is applied and how it determines the recovery positions. The following analysis is based on Eqs. (10)–(11) in Theorem 2 of the main text, Eq. (17) in Appendix F.2, and Eqs. (85) and (87) in Appendix K.13.
1. What the sharing granularity controls. Theorem 2 Eqs. (10)–(11) and F.2 Eq. (17) provide the hierarchical decomposition and query-sharing terms. With KV summaries fixed, G1/G2/per-query alter only the sharing scope of query-sharing logits and recovery weights; they share one state across 128, 64, and 1 queries, respectively. Per-query removes the shared terms, while Eq. (87) retains intra-block key variation. Thus, granularity refinement targets query-sharing bias, and KV-summary granularity controls intra-block variation.
2. Deriving Recovery Priorities from Residual Output Bounds. Eq. (87) yields , and Eq. (85) yields the target-granularity residual bound . The common Exact-only starting point is constrained by Eq. (44) and the Value diameter:
\le D_{V,D}\le\overline D_{V,D}.$$ Therefore, for a complete query tile $G_r$, the following is computed at a given target granularity: $$\sum_{i\in G_r} \left[\overline D_{V,D}^{\,2} -\big(\overline D_{V,D}\tanh(\epsilon_i/2)\big)^2\right].$$ This score is the sum of unrecovered-bound squares minus target-granularity-bound squares. **A larger score means a larger reduction at that location.** It applies to G1, G2, and per-query, incorporating shared bias, intra-block variation, and Value scale; same-budget experiments verify the resulting gains. **3. Actual Benefits Under the Same Budget.** Starting from the same Exact-only baseline, we selected a specified number of 32 complete Q128 tiles to enable the target granularity and reported $$\frac{\text{Exact-only Rel-L2}-\text{Rel-L2 with Bound Selection}} {\text{Exact-only Rel-L2}-\mathbb E[\text{Randomly Selected Rel-L2}]}.$$ For the same target granularity and tile count, additional computation is fixed. **The ratio measures how many times the error improvement from bound-selected locations exceeds random allocation.** When the random average reduction is positive, a value above 1 indicates lower final error. **480p** | Number of restoration tiles (32 total) | G1 | G2 | Per-query | |---|---:|---:|---:| | 4 | 3.80× | 3.25× | 2.87× | | 5 | 3.12× | 3.72× | 3.26× | | 6 | 3.73× | 3.57× | 3.26× | | 8 | 3.18× | 3.35× | 3.15× | | 16 | 1.87× | 1.85× | 1.76× | **720p** | Number of tiles restored (32 total) | G1 | G2 | Per-query | |---|---:|---:|---:| | 4 | 2.58× | 2.91× | 2.71× | | 5 | 2.77× | 3.10× | 2.59× | | 6 | 2.50× | 2.83× | 2.66× | | 8 | 2.70× | 2.73× | 2.54× | | 16 | 1.46× | 1.52× | 1.45× | **4. Where the guidance is applied.** Calculate the Step 2 score from input summaries and enable the target granularity on the highest-scoring query tiles. Since Eq. (81)’s quadratic upper bound is additive across tiles, **this selection minimizes the overall quadratic upper bound** for a fixed granularity and tile count. This step determines the recovery priority based on the residual bound at the target granularity, and the true gain per tile is given by $$\sum_{i\in G_r}\left( \|O_i^M-O_i^{\Omega}\|_2^2 -\|\widehat O_i-O_i^{\Omega}\|_2^2\right)$$ The table reports Spearman correlation between input scores and true gains, testing whether locations with larger bound reductions improve more; true gains are used only for evaluation. | Recovery Granularity | 480p Spearman | 720p Spearman | |---|---:|---:| | G1 | 0.650 | 0.416 | | G2 | 0.629 | 0.513 | | Per-query | 0.730 | 0.485 | **Conclusion: The formulas in the main text provide residual bounds for each granularity, which are used to rank and select recovery positions; experiments with the same budget verify that this guidance is effective for G1, G2, and per-query.** **W4** R4: This weakness has already been addressed in Q3.1–Q3.2: Theorem 2/K.13 transforms the multi-granularity error decomposition into a criterion for determining recovery positions in the input summary, and its computational allocation role has been validated by 30 configurations. Thank you for your question, which has helped us clarify the practical guidance provided by our theory and identify directions for future research.Thank you for your rebuttal and the additional experiments. However, my concerns about the incremental novelty and the practical value of candidate reuse remain. The comparison with PISA-0th shows a modest gain, while candidate reuse has not demonstrated a clear quality–latency advantage over full-domain. I therefore maintain my original score.
Response of the Review from Reviewer japD (Part 1/3)
Thank you for your detailed review. The core innovation of QUASAR is a portable, controllable-recovery sparse-attention operator introduced through Candidate integration, together with its adaptation to DiT. We address its mechanism, deployment value, and theoretical guidance through PISA-0th/SVG-EAR comparisons, Meta Comment C1/C2 Full-KV evidence, and input-bound analysis.
Q1.1 What are the main mathematical and implementation differences between Recovery, SVG-EAR’s centroid compensation, and PISA’s approximation?
A1.1 QUASAR combines three mechanisms: Candidate selection, independently controlled recovery coverage/granularity, and cross-backend recovery-state reuse.
① Controllable recovery granularity. PISA-0th and SVG-EAR use per-query compensation; QUASAR supports G1, G2, and per-query modes via g(i), adjusting accuracy and cost through group sharing or per-query weights.
② Independent candidate and recovery coverage. Exact selection remains within the Candidate, while Full-KV recovery covers all unselected valid blocks. With matching Exact set, granularity, summaries, and normalization, Candidate and Full-Domain produce identical outputs.
③ Recovery-state reuse across backends. Exact and Recovery states merge under the same scaling and integrate with Magi and FlexAttention; per-query recovery uses a query-block tiled kernel, while coarse recovery reuses group computation.
See Meta Comment C2 for related-work comparison; Theorem 2, F.2, and K.12 cover output error, granularity, and out-of-domain omissions. Quantitative effects appear in A3.2 and Meta Comment C1.
Q1.2 Do these differences reduce overhead, broaden applicability, or improve the quality–latency trade-off under comparable settings?
A1.2 The practical benefits are:
① Reduced execution overhead. With PISA-0th-aligned selection rules and Exact budget, QUASAR’s per-query invocation decreases from 12.280 to 11.422 ms. The table compares the full 96-prompt pipeline; matched settings are in Meta Comment C2.
② Extended Candidate Support. The same Recovery incorporates Draft and XAttention candidates; see A2.1.
③ Improved Quality–Latency Trade-off. Appendix Table 13(a) shows 720p per-query K=39 at AVG5 0.8215/428.712 s, outperforming G1 K=52 (0.8186/435.907 s) and K=64 (0.8210/452.561 s). Table 5 validates fixed-support control; generation shows the Pareto gain from fewer Exact blocks.
Q1.3 Please focus on the distinction that most clearly establishes the paper’s added value.
A1.3 The core distinction is transferability combined with controllable recovery. Decoupling candidates, Exact selection, Recovery, and backend lets different selection methods reuse one operator; at fixed sparsity, G1/G2/per-query adjust accuracy and cost independently.
Figure 3 shows competitive Full-Domain, Draft, and XAttention Pareto points across approximately 10%–45% Exact-attention computation.
W1 R1: This weakness has been addressed in Q1.1–Q1.3: C2 distinguishes between the zero-order endpoint in PISA-0th, the clustering/routing in SVG-EAR, and the combination of candidates, granularity, and state in QUASAR; Q1.2 provides an execution comparison with the same support.
W2 R2: This weakness has been addressed in Q1.1–Q1.2 and Q3.1–Q3.2: G1/G2/per-query make precision–cost a configurable axis, supported by fixed support error and 720p Pareto results.
Q2.1 In the representative Wan comparisons, candidate modes are slower and yield lower quality. Under what deployment conditions are they preferable to Full-Domain?
A2.1 For existing candidate workflows, the Candidate interface retains the original selection design and achieves quality close to Full-Domain through Full-KV recovery. Meta Comment C1 includes operator and E2E comparisons with candidate-generation costs.
Q2.2 If their practical value depends on candidate indices already being available at no additional generation cost, please clarify where that assumption applies.
A2.2 Candidate indices are available at the output of the methods in Figure 1. Draft/XAttention generates C within the same attention call, which QUASAR consumes; generation cost is included in total latency. Appendix Table 9 compares 75%-sparse providers:
In-domain recovery already improves AVG5 by 0.0104 while reducing E2E. Full-KV recovery further compensates outside the candidate domain, achieving quality close to Full-Domain (Meta Comment C1).
W3 R3: This weakness has been addressed in Q2.1–Q2.2: C1 and the provider comparison show that Candidate selection is retained while Full-KV recovery restores quality close to Full-Domain.