Decode · tok/s
122.5
Prefill
2100
tok/100W
17.5
and 122.5 peak... Prefill is 2100 as per logs, not sure I can squeze more.
First-person 2× 3090, same thread as the IQ3_S GSQ-RCO Strata setup. Rank is the later quoted 122.5 peak. Prefill rank is the log figure 2,100, not the earlier up-to 2,944. The 12K up-to 117.3 and the 87–96 long-context range stay off the rank. Abliteration is a control vector, not a new weight file.
Uncensored