Features

Diffulex features are grouped by the part of inference they affect: decoding state, parallel topology, attention and cache behavior, runtime capture, adapter loading, and MoE execution.

• Multi-Block Diffusion
• Decoding Strategies
• Model and Data Parallelism
• Optimized Attention
• Prefix Caching
• CUDA Graph
• LoRA Adapters
• Fused MoE

Use the feature pages as tuning references. For a first run, keep defaults and change only the model path, model name, and decoding strategy. Once generation works, tune one feature area at a time.