4 ms·
We don't current compile in CLBlast or ROCm support but if there's a lot of demand for this, we'll definitely add it in the future. One concern is not wanting t
by Patrick_Devine 3y ago
We don't current compile in CLBlast or ROCm support but if there's a lot of demand for this, we'll definitely add it in the future. One concern is not wanting to bloat out the binary size too much (CUDA is already huge!) but given how big the LLM models are anyway, maybe it's not a huge concern.
- globuous 3y agoAMD support would be amazing <3 I get the boot concern, and the maintenance concern (!!!), but as you say, these models are already quite huge anyway :)
- capableweb 3y agoOffer two builds :) One AMD and one NVIDIA.
- Patrick_Devine 3y agoPossibly, but a core guiding principle for us is to keep everything as simple as possible. We're a small project, so if we add too many features/permutations it really makes it hard to keep everything working!
- nebster 3y agoAlso, some of us weird folk have both an AMD and NVIDIA GPU in one machine... Actually, I guess that is more common now that AMD CPUs have a GPU built in?
- zamalek 3y agoYou could download both binaries for that scenario.
- zamalek 3y ago> permutations I don't believe that AMD/NVIDIA is a low entropy bit so far as configuration permutations go. Although NVIDIA is far more widespread, AMD has significant market share. The Darwin bit you already facilitate for is probably lower entropy.