@maruru_289 No worries at all! Thank you so much for all your hard work in putting together such a wonderful resource for the community. Take care and best of luck with everything! ๐
@coldniko@great_garbanzo Quick question, sorry if this is obvious: what's the "fast mode" for Qwen on B65? I checked 0.2.8 and the only "Faster mode" I found in the docs is DeepSeek's 5-of-6-experts thing. Which switch am I missing? ๐
@great_garbanzo@coldniko decode: ~37 tok/s on decode-heavy, ~38 short, ~42โ44 small/medium, ~37 at 100K context
prefill: ~850 tok/s at 8K, ~700 at 100K โ long prefill is the weak spot on single B-line.
Haven't tried fast mode yet, will report back if it moves these numbers.
@coldniko@ivanalog_com So, Lord Niko, since youโre holding the secret recipe until next week... if I'm GPU hunting right now (new or used), which card should I prioritize that won't require a second mortgage like a 5090? ๐