Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing

Motivation

• Open-source dLLMs had not demonstrated faster inference than similarly sized AR LLMs, despite their potential for parallel token generation.Existing approaches pursued KV-cache strategies or parallel sampling, but block-wise sequential generation could eliminate inter-block parallelism.Source: §1 · p. 2

Discrete diffusion forcing

• D2F extends diffusion forcing to discrete text, allowing subsequent blocks to be predicted before preceding blocks are fully denoised.The resulting paradigm combines block-wise sequential generation with inter-block parallelism instead of requiring complete preceding blocks.Source: §1 · p. 2

Method

• D2F dLLMs denoise token blocks with increasing mask ratios, use block-wise causal attention for accurate KV caching, and learn through asymmetric distillation.Inference uses a pipelined parallel decoding algorithm that trades off efficiency and performance.Source: §1 · p. 2, Figure 3 · p. 3

Results

• More than 50× acceleration was achieved over LLaDA-Instruct-8B and Dream-Base-7B without compromising average performance across mathematical and programming benchmarks.The evaluated benchmarks included GSM8K, MATH, HumanEval, and MBPP.Source: §1 · p. 2
• Up to 2.5× faster inference than prevalent AR LLMs such as LLaMA3 and Qwen2.5 establishes D2F as a faster-than-AR open-source dLLM approach.The contribution is presented as the first faster-than-AR dLLMs.Source: §1 · pp. 2–3

Diffusion language models

• dLLMs seek to reduce AR decoding latency by iteratively denoising masked sequences rather than generating tokens sequentially.This parallel denoising paradigm motivates diffusion-based language generation.Source: §2 · p. 3

Acceleration methods

• Existing dLLM acceleration methods use approximate caching schemes or sampling optimizations to reduce redundant computation and decoding steps.Bidirectional attention makes standard KV caching incompatible, while confidence-aware strategies reduce iterations.Source: §2 · p. 3

Pipelined decoding

• D2F dynamically adds blocks when the latest block exceeds a completion threshold, while fully activated blocks decode multiple tokens more aggressively.The illustrated threshold is τadd = 1/6.Source: Figure 4 · p. 4

AR-diffusion hybrids

• AR-diffusion hybrids apply autoregressive dependencies across units while retaining diffusion denoising within units, a design previously explored in video generation.This architecture separates temporal dependency modeling from within-unit denoising.Source: §2 · p. 4

Forward process

• Masked diffusion progressively replaces tokens in an L-token sequence with [MASK] across a continuous time schedule t ∈ [0, 1].At the endpoint, the sequence is fully masked.Source: §3 · p. 4

Reverse process

• dLLMs learn pθ(Y 0|Y t) to denoise a fully masked sequence into language samples while predicting all masked tokens simultaneously at each step.This simultaneous prediction provides the theoretical basis for parallel generation.Source: §3 · p. 4

Inference bottlenecks

• Bidirectional attention conflicts with standard KV caching, causing redundant computation across denoising steps.This is a central practical inference bottleneck for masked dLLMs.Source: §3 · p. 4
• Conditional independence assumptions make interdependent-token generation difficult, requiring more iterative steps for high-quality outputs.Prior methods do not simultaneously provide precise KV caching and efficient parallel decoding.Source: §3 · p. 4

D2F setup

• D2F training divides sequences into blocks, applies a monotonically increasing noise schedule, and updates a D2F model through asymmetric distillation from a pretrained dLLM.The supplied training procedure samples blocks and noise levels before predicting block distributions.Source: §3 · p. 5

Discrete Diffusion Forcing

• D2F provides block-level AR generation for standard KV-cache reuse and inter-block parallel decoding from incomplete predecessors.These capabilities reduce redundant computation while increasing the number of tokens decoded per inference step.Source: §4.1 · p. 5
• D2F partitions sequences into equal-sized blocks and applies monotonically increasing noise levels so earlier blocks are more complete than later blocks.The reverse process can finalize preceding blocks while advancing denoising in subsequent blocks.Source: §4.1 · p. 5
• Block-wise causal attention preserves exact KV reuse across completed blocks while retaining bidirectional attention within each block.This attention structure supports cached computation without requiring fully denoised preceding blocks before later-block decoding advances.Source: §4.1 · p. 5, §4.2 · p. 6

Asymmetric Distillation

• D2F is distilled from a pre-trained bidirectional dLLM using an asymmetric KL-divergence objective under the student's causally restricted view.The teacher uses a global view of noisy blocks, while the student learns to reproduce its mask predictions with block-wise causal attention.Source: §4.2 · pp. 5–6

Pipelined Parallel Decoding

• Algorithm 2 updates the KV cache after blocks are completed during pipelined D2F inference.The procedure appends blocks according to decoding progress and confidence thresholds while operating on cached forward passes.Source: §4.2 · p. 6
• Pipelined parallel decoding maintains a sliding window, dynamically appends masked blocks, and uses dual-state activation to balance efficiency against decoding quality.New blocks begin semi-activated and become fully activated after their predecessors accumulate sufficient decoding progress.Source: §4.3 · p. 6

Experimental Setup

• D2F is evaluated on GSM8K, GSM8K-CoT, MATH, HumanEval, and MBPP using LLaDA-Instruct-8B and Dream-Base-7B.Comparisons include Fast-dLLM, dLLM-Cache, LLaMA3-Instruct-8B, and Qwen2.5-Base-7B.Source: §5.2 · pp. 7–8, §5.1 · p. 7

Main Results

• 119.9 tokens/s on GSM8K gives D2F-Dream-Base-7B a 2.5× speedup over LLaMA3-Instruct-8B and 2.3× over Qwen2.5-Base-7B.The maximum generation length is 512 for all methods.Source: §5.2 · p. 8
• 52.9× speedup on LLaDA-Instruct-8B preserves comparable performance, while D2F reaches 90.2 tokens/s on MATH, 2.1× Fast-dLLM’s Dual-Cache.The LLaDA comparison reports 47.6 versus 0.9 tokens/s and scores of 38.0 versus 39.0.Source: §5.2 · p. 8
• 91.2 tokens/s on GSM8K-CoT yields a 9.6× speedup over Dream-Base-7B and a 1.8× speedup over Fast-dLLM, with slight performance improvement.A unified maximum length of 256 is used for this comparison because the base model struggles with the stop token.Source: §5.2 · p. 8

Throughput-Performance Trade-off

• 150.9 tokens/sec with a score of 71.2 on GSM8K gives D2F 3.1× LLaMA3’s throughput while exceeding its score of 70.1.Dream-Base-7B instead drops from 71.4 to 42.8 when sampling steps decrease from 512 to 128.Source: §5.3 · p. 9

Ablations and Analysis

• Increasing inference block size reduces throughput and first improves then worsens performance; block size 48 peaks at 77.5 versus 75.9 for size 16.The ablation tests integer multiples of the training block size, which is 16.Source: §5.3 · p. 9, Figure 5 · p. 9
• With τact = 0.85, the dual-state setting τadd = 0.7 raises score from 72.6 to 74.2 and throughput from 136.8 to 139.0 TPS.Further reducing τadd improves score while marginally reducing throughput.Source: §5.3 · p. 10
• On MBPP, structured D2F noise scheduling improves score by 5.0 points and throughput by 24.0 TPS over independent random schedules.The comparison uses τact = 0.95, maximum length 512, block size 32, τconf = 0.9, and τadd = 0.1.Source: §5.3 · p. 10, Table 4 · p. 10

6 CONCLUSION

• D2F conditions next-block prediction on partially predicted tokens from previous blocks, supporting KV cache and parallel generation.The paper reports D2F as the first dLLM to support faster-than-AR inference.Source: §6 · p. 10

D2F Components

• Cache-only enables standard KV caching but generates blocks serially, whereas Cache + Para combines caching with pipelined parallel decoding.The two configurations isolate cache use from asynchronous block generation.Source: §A.1 · p. 14
• Adding the parallel pipeline increases GSM8K speedup for LLaDA from 2.4x to 7.3x beyond the cache-only configuration.The result highlights the contribution of asynchronous generation.Source: §A.1 · p. 14

• Dream-Base* was created by directly fine-tuning Dream-Base on the same distillation dataset used for D2F.The control model used LoRA with exactly the same parameters as D2F training.Source: §A.2 · p. 14
• Dream-Base* shows minor score improvements on some tasks but suffers a significant inference-speed drop.The speed degradation is attributed to computational overhead from the added LoRA layers.Source: §A.2 · p. 14
• The control experiment identifies D2F methodology, rather than the training data, as the key driver of the exceptional inference acceleration.Source: §A.2 · p. 14

Sampling hyperparameters

• Table 8 specifies maximum generation length, inference block size, and pipelined-decoding thresholds for D2F models.The listed thresholds include token addition, block activation, and token confirmation thresholds.Source: §A.3 · p. 14, Table 8 · p. 16
• Baseline hyperparameters are adopted from prior work to ensure fair comparison with D2F configurations.Source: §A.3 · p. 14, Table 8 · p. 16

Ablation and control tables

• Tables 5, 6, and 7 report ablations and baseline-relative performance ratios for D2F-LLaDA, Dream-Base*, and D2F-Dream.Table 6 compares Dream-Base with its directly fine-tuned version, Dream-Base*; Tables 5 and 7 study proposed-method variants.Source: Table 5 · p. 14, Table 6 · p. 15, Table 7 · p. 15

Sampling hyperparameters

• For D2F, the length parameter sets only a maximum generation limit and does not affect the sampling distribution because generation is block-wise.For Dream-based models, D2F’s maximum length matches the baseline because the base model often fails to generate proper termination tokens.Source: Table 8 · p. 16