Discrete Diffusion Parallel Sampling Delivers Lossless LLM Speedups Without Draft Models
IFM · hf · 2026-09-08
IFM released work on diffusion-augmented autoregressive language models on Hugging Face: distilled diffusion weights plus a specialized sampler enable parallel token sampling on top of an autoregressive backbone.
- Multiple tokens are generated in parallel, breaking the serial decoding bottleneck.
- Claimed to be lossless (no quality degradation) and requires no draft model, unlike speculative decoding.
- Relevant to inference cost and latency optimization for LLM serving.
More from Infra
- CPU-Only LLM Tests: 35B MoE at Q2 Beats a 2B Model Despite Half the Speed — ML-Future · 2026-09-08
- HydraDB: a Rust graph database that lives entirely in S3, with nothing on disk — thisdudelikesAI · 2026-09-08
- Running Qwen 27B and Gemma 31B locally on one RTX 4090: quantization and context tradeoffs — MooseEfficient2151 · 2026-09-08
- Could distributed iPhones form an inference network? A 2AM open question — gajesh · 2026-09-08
- Candidate depth 10→500 lifts score by just 0.01: Qdrant on diagnosing before tuning vector search — qdrant_engine · 2026-09-08
- Custom llama.cpp build pushes 7900XTX to 1600tk/s prefill on Qwen 27B Q8 — nasone32 · 2026-09-08