Daniele

I'm noticing just now some that the model loaded using hipblas under certain circumstances generates garbage output compared to the no hipblas version even without the f32 memory flag.

Try for example the prompt chat-with-bob. Using hipblas it seems not to understand the context of the conversation.

On the Radeon 5700 XT everything works correctly except the perplexity test (the --memory_f32 perplexity test works). However the card isn't officially supported by ROCm and it doesn't seem to...

> > On the Radeon 5700 XT everything works correctly except the perplexity test (the --memory_f32 perplexity test works). However the card isn't officially supported by ROCm and it doesn't...

Sure. By installing the CUDA toolkit can Windows users build without any problem? I don't have a Nvidia GPU so I can't test it.

Ok, I'll just add that info then.

@SlyEcho Producing the build instructions for Windows in a clear way seems to add quite a lot of space to the readme. If that is not a problem I can...

@SlyEcho I'm finishing a revised version of the build instructions to include make on Windows. Is there any reason why you recommended the fortran version of w64devkit?

So it should be OK to link the fortran version also for the normal build or it may be better to link the vanilla version for that?

I think that it may be quite clear like this.