Kirill R.
Kirill R.
imho seems way too hardcoded for halfs In weights loading code, instead of casts/args scattered all over, would be nicer to find a single place that would call a passed...
needs a test `test/extra/test_utils.py` also `load_single_weight` is called in multiple places, and it won't realize in `fake_torch_load` it seems
ok, nice and clean now 😅 @geohot
what do you mean by it exactly?
afaik ggml files are just weights, the model is C code calling the library. So making weights loader is likely not hard, but the actual model has to be reimplemented
hmm, ggml can export computational graphs now (not for all models it seems), so might be actually possible: https://github.com/ggerganov/ggml/pull/108
Also would be good to add a test for operations on `inf`s
which weights you use? If default ones, the shasum should be this: ``` $ shasum weights/sd-v1-4.ckpt 210783247af4f65a3d23d026490cc37a670964dd weights/sd-v1-4.ckpt ``` you can try to remove `weights/sd-v1-4.ckpt` and redownload if hash is...
can you add `print(t[0:64].numpy())` here above `if`: https://github.com/geohot/tinygrad/blob/079ea217a38bc90ce8e64137cf7d04d664ad9db6/tinygrad/state.py#L94-L95 and show what it prints?
I have this hacky solution for `inf`s, but idk. torch also works correctly with `nan`s. Also not sure about non-floats and cases of different dtypes. Maybe there should be an...