Strix Halo 1.95 BPW
Really REALLY like the 1.95 BPW quant on the 128GB Strix Halo. It's replaced Bartowski's IQ2_S quant for me. Similar results (And better than the IQ2_XS) and at a much lower memory usage. Can fit a HUGE cache on there without anything accidentally offloading onto pagefile.
I would love love love to see a 2.25 BPW quant. That'd make the absolute most of the Strix Halo I believe. The 2.5 BPW quant is too big to fit, but I feel like there's still room for improvement over the 1.95 BPW quant.
Sure, I can add this to my list. Unfortunately my tooling is all setup for making Minimax 2.7 quants right now so it might be a few days.
Thanks for that!! I'm starting to play around with your MiniMax M2.7 quants right now actually.
Really REALLY like the 1.95 BPW quant on the 128GB Strix Halo. It's replaced Bartowski's IQ2_S quant for me. Similar results (And better than the IQ2_XS) and at a much lower memory usage. Can fit a HUGE cache on there without anything accidentally offloading onto pagefile.
I would love love love to see a 2.25 BPW quant. That'd make the absolute most of the Strix Halo I believe. The 2.5 BPW quant is too big to fit, but I feel like there's still room for improvement over the 1.95 BPW quant.
2.25bpw is uploading, stats posted for it on main page. Might be up to 5-6 hours from this message until it's up.
I saw it! Nice, the PPL looks promising.
I have high hopes for it, no doubt it'll be the best 397B quant for 128GB RAM.
Thank you!!