@kalomaze speculative decoding that invokes the full model under the wrong conditions
Notes/ TwitterApr 14, 2025, 19:32 EDTPrevious in thread · mimi↗@kalomaze cacheing fuckery tokenizer fuckery pseudorandom that isn't distributed properly MoE but reverse a sign in the routing between training vs inference@kalomaze speculative decoding that invokes the full model under the wrong conditions← Previous noteNext note →