Testing DiffusionGemma 26B-A4B-it Q4_K_M GGUF on RTX 3090 24GB.
Prompt: "Explain DeepMind."
The highest raw TPS was 222.8 tok/s at max_steps=4, but the output was corrupted.
By English standards, the usable point is measured around 49.5 tok/s at max_steps=24.
Conservative quality setting max_steps=48 was 42.9 tok/s.
I thought English would produce better results, but it seems similar.
Sample summary:
max_steps=4:
DeepMind is a British artificial intelligence (AI) research laboratory based in London
subsidiary of subsidiary of subsidiary subsidiary parent parent...
max_steps=12:
DeepMind is a British artificial intelligence research laboratory based in London. It is a
subsidiary of Alphabet Inc...
...
AlphaZero ... playing playing against itself ... surpassing surpassing...
max_steps=24 had quite usable English answers. However, there is a phenomenon of repeating the same paragraph.
max_steps=48 is also the most stable but not completely clean.
