WeDLM

Why we included this project

Diffusion language models promise parallel token generation, but most use bidirectional attention that breaks the KV-cache machinery production inference engines depend on, so the theoretical speedup never shows up in practice. WeDLM, a research release from Tencent's WeChat AI, works around that by performing parallel mask recovery under standard causal attention, which keeps it compatible with FlashAttention, PagedAttention, and CUDA Graphs. The included 7B and 8B models initialize directly from Qwen2.5 and Qwen3 weights and report real wall-clock gains over vLLM-served autoregressive baselines. Along with the inference engine you get a web demo, a fine-tuning framework, and evaluation scripts, and chatllm.cpp support makes local deployment straightforward. If you're evaluating faster serving options, this is a concrete candidate to benchmark against your current stack.