想着五一前后挑战llama4 400B发现并没什么挑战性，q4模型用7970X (150GB/s)纯CPU prefll 108 t/s decode 13.8 t/s，用8G显存offload dense层27 t/s，塞满双卡96G显存能30.8 t/s
不过llama.cpp的override tensors的prefill看起来是用纯GPU走PCIe访问内存里的模型，还有优化空间。至少不应该比纯CPU差
BY David's random thoughts
Telegram today rolling out an update which brings with it several new features.The update also adds interactive emoji. When you send one of the select animated emoji in chat, you can now tap on it to initiate a full screen animation. The update also adds interactive emoji. When you send one of the select animated emoji in chat, you can now tap on it to initiate a full screen animation. This is then visible to you or anyone else who's also present in chat at the moment. The animations are also accompanied by vibrations. This is then visible to you or anyone else who's also present in chat at the moment. The animations are also accompanied by vibrations.
Importantly, that investor viewpoint is not new. It cycles in when conditions are right (and vice versa). It also brings the ineffective warnings of an overpriced market with it.Looking toward a good 2022 stock market, there is no apparent reason to expect these issues to change.
David& 39;s random thoughts from kr