Adnan Masood, PhD.·4d agoNext-Concept Modeling (NCM) and the Non-Autoregressive Frontier: Architecture, Taxonomy, and…How Stanford’s Next-Concept Modeling enables parallel text generation: architecture, inference efficiency, and benchmarks against…
Junbum Lee (Beomi)·Apr 17dLLM into TPU: An End-to-End Diffusion LM Stack in Pure JAXFrom an autoregressive HuggingFace checkpoint to a KV-cached Soft Parallel Decoder on TPU v4/v5/v6— without PyTorch. #TPUSprint #TRC
Dr. Shouke Wei·Dec 25, 2025Fast‑dLLM: Training‑Free Acceleration for Diffusion LLMsReimagining Large Language Model Inference with KV Caching and Parallel Decoding
Sirsh Amarteifio·Apr 15, 2025Quickly trying out Inception MercuryInception Labs recently introduced their commercial scale diffusion language model.
InTowards DevbyArXiv In-depth Analysis·Mar 8, 2025Exploring Mercury, the First Commercial-Scale Diffusion Large Language ModelMercury, is making waves as the first commercial-scale dLLM, promising to revolutionize text generation with its speed and efficiency.
InThe Low End DisruptorbyZheng "Bruce" Li·Feb 28, 2025What is Diffusion LLM and why it mattersIntroductionA response icon3A response icon3
![HF autoregressive ckpt (Qwen3–8B, safetensors) │ │ torch-free loader: safetensors + numpy + huggingface_hub ▼ Flax NNX model graph │ │ 2D FSDP + TP sharding, Pallas flash attention, │ CPU-first init + remat for 8B+ ▼ DMax OPUT training loop │ · two-stream [noised | clean] │ · block-diffusion attention mask │ · per-step on-policy rollout (stop_gradient forward → argmax → re-noise) │ · gradient accumulation, cosine LR │ │ Orbax DCP sharded checkpoints → GCS (us-central2, us-east1,](https://miro.medium.com/v2/resize:fill:160:106/1*5JeMPZgr8dhaAbyO9jWN2A.png)
![HF autoregressive ckpt (Qwen3–8B, safetensors) │ │ torch-free loader: safetensors + numpy + huggingface_hub ▼ Flax NNX model graph │ │ 2D FSDP + TP sharding, Pallas flash attention, │ CPU-first init + remat for 8B+ ▼ DMax OPUT training loop │ · two-stream [noised | clean] │ · block-diffusion attention mask │ · per-step on-policy rollout (stop_gradient forward → argmax → re-noise) │ · gradient accumulation, cosine LR │ │ Orbax DCP sharded checkpoints → GCS (us-central2, us-east1,](https://miro.medium.com/v2/resize:fill:320:214/1*5JeMPZgr8dhaAbyO9jWN2A.png)
