User avatar
plautumnpus @autumn@cafe.autumn.town
4mo
ROCm is so awful 😭 i just tried vulkan backend for ollama and it's so much faster
❤️1
2
0
0
1
User avatar
plautumnpus @autumn@cafe.autumn.town
4mo
okay i may have been to hasty, it requires further testing but: ROCm ran the MoE (gemma4:26b-a4b & qwen3.6 35b-a3b) models a lot better than Vulkan, which won out in the smaller models that fit into VRAM (e2b & e2b QAT)
0
0
0
0