- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates the information deficit in parallel speculative diffusion when using anchor representations to predict future states. RELOAD first constructs a rough causal trajectory, then bundles and re-encodes the source states at each step, and uses a shared random innovation to reconstruct candidate trajectories. Role-LoRA adapts only the re-encoding of roles. Under conditions where candidate and target states share the same state-independent Gaussian covariance, as well as item-wise computation and prefix causality, the method preserves the distribution of a specified target sampling chain through acceptance and reflection corrections. Key experiments on DiT-XL/2, ImageNet-256, and DDPM-1000 demonstrate a latency reduction of approximately 10% compared to strict FREE with the same resources.
Critical strengths
- Mechanistic evidence is more robust than simple system comparisons. The B0/B1/B2 experiments—which fix weights while varying the source state—directly test the role of future state information. Ablations such as frozen re-encoding, serial/batch execution, and Role-LoRA further distinguish the sources of reception length and latency gains.
- The correctness argument addresses key dependencies. Appendix A explains why item-by-item re-encoding preserves prefix testability and argues for acceptance and correction given a history containing previous verification decisions. The paper also clearly distinguishes between distribution preservation under exact arithmetic and checks with finite precision and finite samples.
- Key speed conclusions are supported by rigorous paired evidence. Under the same eight-rank configuration, the average latency across 256 paired cases decreased from 6.092 seconds to 5.482 seconds, with 254 cases showing faster performance; paired confidence intervals are reported. The authors also disclose GPU-second costs and some adaptation results showing no significant differences, which aids in accurately understanding the gains.
Critical weaknesses
-
The generalizability of the empirical findings is limited.** Despite controlling for horizon, CFG, and validation rank, the main experiments were confined to a single model, dataset, and a fixed 1,000-step posterior variance DDPM setup. The target chain also explicitly ignored the model’s learned-variance output. Therefore, the empirical results support improvements specific to this particular random target chain and cannot yet be generalized to arbitrary diffusion samplers. The authors have already explained that ODEs and mismatched covariances require different validation constructs; I do not consider these unaddressed settings to be flaws in correctness. This limitation primarily affects the breadth of the contribution rather than the existence of a valid contribution.
-
New contributions should be evaluated based on gains relative to a strict FREE baseline with the same resources; the full acceleration relative to sequential TARGET should not be attributed solely to the new drafter. The primary new benefit is a latency reduction of approximately 10%; the 2.286-fold acceleration relative to the single-active-rank TARGET also accounts for the effect of multi-rank validation. Correspondingly, active-rank GPU-seconds increased from 12.535 for TARGET to 43.859 for RELOAD. The paper has disclosed these differences; therefore, the issue is not a lack of cost transparency, but rather that the research value should be framed as a low-latency improvement under specific resource configurations. Appendix C.3’s failure to report total GPU-hours for Stage I also makes it temporarily impossible to quantify the full training amortization when migrating from existing systems. These limitations lead me to lean toward marginal acceptance rather than a higher rating.
Questions for authors
- Within the specified scope of application, which samplers or architectural conditions are necessary for the causality and verification arguments presented in this paper, and which are merely choices made in the current implementation? Please briefly distinguish between them so that readers can assess the effort required for porting without needing to develop a new verifier.
- Appendix C.3 notes that the cost records for Stage I are incomplete. Is there a verifiable total cost or upper/lower bounds that can be provided to supplement this information? If these cannot be recovered, please retain this limitation and avoid drawing conclusions about the payback period for training from scratch or switching systems based on the 1.001 GPU-hours figure from Stage II.
Author Response to Reviewer 2evo
Thank you for your careful analysis of correctness, paired experiments, resource use, and training cost.
Verifier requirements and portability:
The verifier supports architectures satisfying these transition and causal conditions. DiT, resolution, and step count are implementation choices.
We choose fixed-posterior DDPM and ignore learned-variance outputs; ImageNet, GPU count, and LoRA rank are also implementation choices. Itemwise computation preserves causality. Shared innovations improve efficiency; Appendix A.1 also permits independent auxiliary coarse randomness under its causal conditions.
Porting requires target means/variances, drafter feature/weight adaptation, and causal packing. State-dependent covariances are outside the current proof and implementation. Extending to them requires separate validation of generating densities and rejection-residual correction, including covariance matching and causality. Mismatched covariances are not covered by the current reflection rule; deterministic ODE steps need a different acceleration interface. Section 3.4 now distinguishes proof assumptions, implementation choices, and porting work.
Training records and resource attribution:
Complete measured cost and verifiable bounds remain unavailable. Six recovered Stage-I runtime windows total 74.647 hours; segment timings and GPU-allocation records remain incomplete. Combining these runtimes with missing segments extrapolated from historical weighted mean update time gives:
Both base-training estimates assume continuous four-GPU allocation. Different observed segment speeds give scenarios of 162–167 GPU-hours for FREE and 377–451 for RELOAD; these are not rigorous bounds. The 1.001 GPU-hours covers only the Role-LoRA training loop, not full costs of training from scratch or switching from FREE. Appendix C.3 documents the gaps and estimation, distinguishing Role-LoRA adaptation, switching from FREE, and training from scratch, with associated evaluation costs.
The abstract and Section 4.2 emphasize 10.0% lower latency than strict FREE with the same eight active ranks: 6.092 versus 5.482 s/image, and 48.737 versus 43.859 GPU-seconds/image. TARGET uses one active rank and 12.535 GPU-seconds/image. The 2.286× speedup over TARGET includes drafting and parallel verification.
Scope of the additional experiments:
We added shorter schedules and higher resolution with DiT and fixed-posterior DDPM. At 256 resolution, 250 steps, and B=1, 256 paired cases give 2.753/2.599 s/image for eight-GPU strict FREE/RELOAD: 5.6% lower latency, or 1.059× [1.050, 1.068].
At 512 resolution, we evaluated DiT-XL/2 with DDPM-1000, B=1, L=8, CFG 1.5, and 128 paired cases. All three configurations use eight active GPU ranks:
Frozen re-encoding gives 1.457× [1.446, 1.468] over one-pass drafting. Role-LoRA adds 1.054× [1.046, 1.061] over frozen re-encoding, or 5.11% lower latency. All intervals are paired 95% CIs.
Time constraints prevented strict FREE-512 and non-DiT evaluation. If accepted, we will add FREE-512 under the same target sampler and hardware, and evaluate a non-DiT model satisfying the verifier conditions.
The new 256 experiment uses another A100 80GB server, case set, and evaluator supporting short schedules/batching. Baselines were remeasured; speedups are paired within this experiment. Checkpoints and FP32/TF32 are unchanged. Timing includes communication and excludes loading, warmup, and VAE decoding.