All learning paths

Read inference-efficiency claims critically

Distinguish fewer network evaluations from lower wall-clock latency and identify which comparisons apply to your workload.

Source-linked starting points, not an exhaustive survey or a current leaderboard. Reading exercises below are suggestions, not experiments run by this site.

Before you start

Complete the foundations route, or know masked-token denoising. Bring a target sequence length and serving workload, not a benchmark result.

Network function evaluations (NFE) · Latency versus throughput · Remasking versus carry-over · Block diffusion · Evaluation protocol

1. Compare quality at a compute budget

Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution — Lou, Meng & Ermon · ICML 2024

Sampling and compute–quality comparisons; check the definition of the evaluator.

SEDD studies trading network evaluations for sample quality. Such a trade-off is not itself a wall-clock speedup measurement.

Check your understanding

What must match before comparing two reported perplexities?

Reasoning to check against

Check dataset/split, tokenization, whether the number is a model likelihood bound or an external evaluator score, and decoding settings. Missing conditions mean not comparable.

2. Read the cache assumptions

Simple and Effective Masked Diffusion Language Models — Sahoo et al. · NeurIPS 2024

Sampling discussion and Appendix E.2, including its hardware and batch settings.

MDLM studies caching unchanged denoiser inputs. Inspect when reuse is valid rather than assuming every time step needs a fresh call.

Check your understanding

Is a cache win at batch size one necessarily the same win at a large batch size?

Reasoning to check against

No. Reuse opportunities and accelerator utilization change. Record NFE, batch size, token length, precision, hardware, and elapsed time separately.

3. Locate the sequential dependency

Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models — Arriola et al. · ICLR 2025

Introduction and blockwise factorization; inspect the cache and sampling protocol.

Block Diffusion combines autoregression across blocks with diffusion inside each block, allowing KV caching and within-block parallelism.

Check your understanding

If tokens inside a block are parallel, is the entire sequence generated in one pass?

Reasoning to check against

No. Count both dependencies between blocks and denoising work inside blocks. Block size changes this balance; a latency claim still needs a measured protocol.

Apply the reading

Build a comparison worksheet: quality definition, dataset, length, NFE, batch, hardware, precision, cache policy, wall-clock latency, and throughput. Leave unreported cells unknown. This site has not run these experiments.

Inference acceleration

Long context

These links search the published corpus; a topic can be empty.