Not sam. I am running it with ollama on a server on my lan with two 7900XT. I get about 50-60 tokens per second on phi4-mini with full precision, it only loads

Not sam. I am running it with ollama on a server on my lan with two 7900XT. I get about 50-60 tokens per second on phi4-mini with full precision, it only loads on a single card. The few requests I tried were correct, I think that phi4 the 14 b parameters model produced better code though. I don't recall what it was, it was rather simple stuff though. QwQ seems to produce okay code as well, but with only 40GB of vram I can only use about 8k context with 8bit quantization.