GGUF conversions of inclusionAI/LLaDA2.2-flash, a ~103B-parameter (13B active) Mixture-of-Experts diffusion language model (dLLM). > ⚠️ These files do not work with llama.cpp. LLaDA2 is a block diffusion > model, not an autoregressive one — it denoises a block of masked tokens over > several steps instead of predicting one token at a time. Use > diffuse-cpp. Hugging Face caps individual files at 50 GB, so each model is uploaded as raw byte-range parts. Rejoin them with cat before use (they are not llama.cpp gguf-split shards): The zero-padded names sort correctly under a shell glob. Verify the result is 59086787040 bytes (Q4KS) or 205852119520 bytes (F16), then delete the parts. Source weights are bf16 safetensors (32 shards). Both passes stream tensors one at a time, so peak RAM stays near the largest tensor rather than the model: 1. Convert — tools/convert-llada2-fast.py, bf16 → F16 GGUF. Reads the safetensors shards by raw byte parsing (no PyTorch), converts bf16 → f16 in 4M-element numpy chunks, and writes metadata first then streams tensor data. The 256 per-expert matrices in each MoE layer are stacked into single 3-D tensors ([nexpert, out, in]) so the graph can use…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: Akicou
Теги: gguf, dllm, diffusion, diffusion-language-model, llada, moe, diffuse-cpp
Лайков: 3 | Загрузок: 670
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
