# Parallel Decoding

BYOD is a LoRA-only recipe for converting autoregressive language models into diffusion language models: every base parameter stays frozen, each layer's causal mask becomes bidirectional, and a masked-diffusion objective trains at most 102.4M tokens in 3.1–4.0 hours on one Google Colab GPU, turning Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B and Ministral 8B into models that generate from fully masked sequences, with adapters at 4.1–5.8% of final parameters. The strongest conversion, BYOD-Gemma, reaches 50.6% average accuracy on nine tasks against 51.9% for LLaDA 8B Instruct in the same harness; open-generation quality at low NFE budgets remains the main limitation.

BYOD is a LoRA-only recipe for converting autoregressive language models into diffusion language models: every base parameter stays frozen, each layer's causal mask becomes bidirectional, and a masked-diffusion objective trains at most 102.4M tokens in 3.1–4.0 hours on one Google Colab GPU, turning Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B and Ministral 8B into models that generate from fully masked sequences, with adapters at 4.1–5.8% of final parameters. The strongest conversion, BYOD-Gemma, reaches 50.6% average accuracy on nine tasks against 51.9% for LLaDA 8B Instruct in the same harness; open-generation quality at low NFE budgets remains the main limitation.

The work introduces WaveFront Decoding (WFD), a training-free self-speculative decoding framework for looped language models that exploits two properties of these architectures, namely that intermediate recurrence outputs provide effective draft predictions and that weight sharing lets token states at different positions and recurrence depths be processed in one batched recurrent-block call, organizing these mixed-depth states into a diagonal wavefront so that drafting and verification proceed concurrently within the same recurrent calls while rejected drafts are corrected using full-depth predictions; across six Spec-Bench task categories it achieves 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B over autoregressive decoding, rising to 4.81x on Huginn-3.

The work introduces WaveFront Decoding (WFD), a training-free self-speculative decoding framework for looped language models that exploits two properties of these architectures, namely that intermediate recurrence outputs provide effective draft predictions and that weight sharing lets token states at different positions and recurrence depths be processed in one batched recurrent-block call, organizing these mixed-depth states into a diagonal wavefront so that drafting and verification proceed concurrently within the same recurrent calls while rejected drafts are corrected using full-depth predictions; across six Spec-Bench task categories it achieves 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B over autoregressive decoding, rising to 4.81x on Huginn-3.

Flash-dLLM is a training-free inference acceleration framework for diffusion large language models: it pairs an I/O-aware fused KV-cache kernel — folding QKV projection, RoPE and cache writes together in SRAM and writing keys and values straight into the cache — with scheduled Flash Attention to ease GPU memory-traffic bottlenecks, keeps only a small set of most-attended tokens through selective cache updates, and lets the dLLM serve as both drafter and verifier (Flash-Verify); evaluated on LLaDA-1.5 across GSM8K, MATH, HumanEval and MBPP, it reports 5.1× and 11.0× speedups over the strongest baseline Elastic-Cache on GSM8K and HumanEval while preserving generation quality.

Flash-dLLM is a training-free inference acceleration framework for diffusion large language models: it pairs an I/O-aware fused KV-cache kernel — folding QKV projection, RoPE and cache writes together in SRAM and writing keys and values straight into the cache — with scheduled Flash Attention to ease GPU memory-traffic bottlenecks, keeps only a small set of most-attended tokens through selective cache updates, and lets the dLLM serve as both drafter and verifier (Flash-Verify); evaluated on LLaDA-1.5 across GSM8K, MATH, HumanEval and MBPP, it reports 5.1× and 11.0× speedups over the strongest baseline Elastic-Cache on GSM8K and HumanEval while preserving generation quality.