introduce legacy-torch flag for backward compatibility on older Intel…

[csabakecskemeti]

Intel-based Macs cannot upgrade PyTorch beyond version 2.2.2

Users with older Intel-based Macs cannot upgrade PyTorch beyond version 2.2.2, as Apple dropped support for these machines in newer macOS versions. The latest convert_hf_to_gguf.py fails on these systems with:

AttributeError: module 'torch' has no attribute 'uint64'

This occurs because torch.uint64, torch.uint32, and torch.uint16 were introduced in PyTorch 2.3.

As I love my 2013 Mac Pro I try to keep it useful as long as possible. I can fall back to the older types.
There are so many ways to do this but wanted to make it completely user driven with a '-legacy-torch' flag . I wanted to avoid any automatic fallback - user should know what they do and this is a "last resort" on older machines.

Usage

python3 convert_hf_to_gguf.py --legacy-torch /path/to/model --outfile output.gguf --outtype f16

Other option would be maintaining a separate convert_hf_to_gguf_bw_comp script but I do not like that.
Please suggest if you have better suggestion to solve this.

[CISC]

No flag necessary, you can just comment them out, they are unused (mapped from _dtype_str_map where they are commented out):

llama.cpp/convert_hf_to_gguf.py

Lines 11196 to 11197 in 68a7c14

	dtype = cls._dtype_str_map[tensor.dtype]
	numpy_dtype = cls._dtype_byteswap_map[dtype]

Add similar comment to this:

llama.cpp/convert_hf_to_gguf.py

Line 11130 in 68a7c14

# TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734