convert-pth-to-ggml.py error with "Got unsupported ScalarType BFloat16"

[austinchau]

Trying to convert "chavinlo/alpaca-native" alpaca native model's (https://huggingface.co/chavinlo/alpaca-native) weights to ggml but got this error -

Processing part 0

Processing variable: model.embed_tokens.weight with shape: torch.Size([32001, 4096]) and type: torch.float32
Processing variable: model.layers.0.self_attn.q_proj.weight with shape: torch.Size([4096, 4096]) and type: torch.float32
Processing variable: model.layers.0.self_attn.k_proj.weight with shape: torch.Size([4096, 4096]) and type: torch.float32
Processing variable: model.layers.0.self_attn.v_proj.weight with shape: torch.Size([4096, 4096]) and type: torch.float32
Processing variable: model.layers.0.self_attn.o_proj.weight with shape: torch.Size([4096, 4096]) and type: torch.float32
Processing variable: model.layers.0.self_attn.rotary_emb.inv_freq with shape: torch.Size([64]) and type: torch.bfloat16
Traceback (most recent call last):
File "/Users/domeie/projects/llama.cpp/convert-pth-to-ggml.py", line 157, in
main()
File "/Users/domeie/projects/llama.cpp/convert-pth-to-ggml.py", line 151, in main
process_and_write_variables(fout, model, ftype)
File "/Users/domeie/projects/llama.cpp/convert-pth-to-ggml.py", line 109, in process_and_write_variables
data = datao.numpy().squeeze()
TypeError: Got unsupported ScalarType BFloat16

[gjmulder]

Please review and use our issue template to provide more details so we can better understand your problem and answer you.

[prusnak]

The scripts have been rewritten from scratch. Try the new convert.py script from master.