parallel-decoding

[EMNLP 2026] 📚 A curated list of Awesome Efficient dLLMs Papers with Codes

[NeurIPS 2026] DMax: Aggressive Parallel Decoding for dLLMs

[ICML 2026] When Drafts Evolve: Speculative Decoding Meets Online Learning

Repository of the paper "Accelerating Transformer Inference for Translation via Parallel Decoding"

[ICML 2026] Jacobi Forcing: Fast and Accurate Diffusion-style Decoding

[ICLR 2026 Oral] Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

[ICLR 2026] Autoregressive Image Generation with Randomized Parallel Decoding

A 4B diffusion grounding model for fast, precise visual evidence extraction. Unifies diverse grounding tasks through parallel decoding, with optional self-speculative decoding delivering 4.51x autoregressive throughput on COCO.

JEV-inspired parallel decisions for CUDA LLMs. One context, many decisions. vLLM API, game-agent examples, and reproducible benchmarks.

[ICLR 2026] dParallel: Learnable Parallel Decoding for dLLMs

[ICLR 2026] ParallelBench: Understanding the Tradeoffs of Parallel Decoding in Diffusion LLMs

Implementation of Flash-DLM (paper: FlashDLM: Accelerating Diffusion Language Models via Efficient KV Caching and Guided Diffusion). Provides training-free methods to accelerate diffusion language model inference.

Parallel Decoding of Language Models: A Survey. A systematic survey of parallel decoding, covering speculative decoding, diffusion LLMs, etc...

This repository contains the source code for COrAL, an Order-Agnostic Language Modeling framework for Efficient Iterative Refinement.

Predictive-State Flows (PSF): a flow prior generates K predictive states, each summarizing only the preceding blocks; a block decoder writes all blocks in parallel. Code and checkpoints for arXiv 2610.12898.

Strata 125B MoE + JEV mode: llama.cpp fork with parallel-decision endpoints (yes/no in ~10ms), an MCP server, and a llama-swap router swarm for Claude Code & DeepSeek Harness.

Educational implementation and visualization of autoregressive, DFlash-inspired, and DFlash2-inspired LLM decoding with a tiny Transformer benchmark.

What parallel commitment costs a masked diffusion LM: a 4-page note, measured on a task with a decidable verifier. Produced by the Rust stack (candle-mi -> canvas).

Diffusion VLM papers indexed with one feature schema, plus decode-behaviour traces measured through a single instrumentation path

Masked diffusion language model (201M params, RTX 3090) — parallel token generation validated by an AR oracle, 1.8-3.9x faster forward pass than autoregressive. Weights on HuggingFace.