QuantaViolet

Turn a checkpoint into a GGUF you can load. Krea 2 included.

A 13 GB fp16 model does not fit a 12 GB card; the quantized copy is what you run. For most architectures someone has published one. For Krea 2 nobody has, and no published converter handles it, so QuantaViolet carries the architecture template itself. Point it at a .safetensors checkpoint and it reads the header first, before the long write: which architecture it is, how big each quant level will come out, whether the disk has room, and which tensors will refuse to quantize. Then Convert writes a loadable full-size GGUF, and Convert and quantize adds the Q4_K_M, Q5_K_M, Q6_K and Q8_0 levels.

• Plan first — architecture, size per level, disk space and the tensors that fall back, from the file header alone
• Convert to GGUF — a loadable full-size file, with Krea 2 detected and written as krea2
• Quantize — Q4_K_M, Q5_K_M, Q6_K and Q8_0 from the converted file
• Command line — the same two steps as a tool that prints one line of JSON per event
• Where it lives — the QuantaViolet panel in the Trainer screen of VIOLINET OS

Python · gguf · llama.cpp · inside VIOLINET OS · Apache-2.0 converter base

Built. Tested end to end on a stand-in Krea 2 checkpoint; a full-size Krea 2 file and the quantize step still need their first real run. Release is coming soon.