While going out to play on the weekend, I was looking for the optimal values of ComfyUI's Minimax H3,
I used DeepSpeech V4 Flash with OpenRouter, and it sold out for $10 --;
I used more tokens than I thought. It's a bit of an error to spend money on a machine that you bought to save money.
I'm doing pretty well setting up Comfy with Meta Muse Grimmer 30B Q4, which came out yesterday.
Build according to pytorch and rocm versions, and if there are any problems, look for them on the internet. It builds itself by looking it up.
It's a shame that the context is 128K, but even if you fill 128K, the prefill is about 300 per second, so it seems like it would be too much to increase it further.
Try this combination with pi-agent! Try that combination! Download int8convrot and try it! etc.
I've been doing well for two days ㄷ ㄷ ㄷ
When the video generation runs, the LLM can't speak, but even though it sometimes runs together and there are problems, pi agent usually runs it as a script.
By the way, creating videos and photos on Strixhalo is really slow -_-;
