Your comment sent me down a week of runs. Each card repeats itself exactly; the differences come from the order each GPU adds its numbers, which depends on how many processors the card has. The full trace, and the one llama.cpp setting that makes two cards agree on every token, is here: dev.to/gde/one-word-two-gpus-why-g...
We're a place where coders share, stay up-to-date and grow their careers.
Same files, same build, and still 2 BF16 questions graded differently across the two cards is the part I keep coming back to.
Your comment sent me down a week of runs. Each card repeats itself exactly; the differences come from the order each GPU adds its numbers, which depends on how many processors the card has. The full trace, and the one llama.cpp setting that makes two cards agree on every token, is here: dev.to/gde/one-word-two-gpus-why-g...