convert-pth-to-ggml.py expects 2 parts for ALPACA-LORA-13B, but it has only one

[stduhpf]

convert-pth-to-ggml.py expects 2 parts for ALPACA-LORA-13B, but it has only one

[stduhpf]

Fixed!
(though you should update README)

[stduhpf]

Also I get this error when I try to run it after quantize:

[Info]: Func('Model') loading model(='ALPACA-LORA-13B') from ./models/ALPACA-LORA-13B/alpaca-lora-q4_0.bin - please wait ...
[Info]: Func('KVCacheBuffer::init') kv self size  =  512.00 MB
[Error]: Func('Model') invalid model file ./models/ALPACA-LORA-13B/alpaca-lora-q4_0.bin (bad magic)
[Error]: Func('FastLlama::Params::build') Unable to load model
Traceback (most recent call last):
  File "examples/python/example-alpaca.py", line 11, in <module>
    model = Model(
  File "examples/python/build/fastllama.py", line 208, in __init__
    raise RuntimeError("Unable to load model")
RuntimeError: Unable to load model
Exception ignored in: <function Model.__del__ at 0x7f4e2d3cb310>
Traceback (most recent call last):
  File "examples/python/build/fastllama.py", line 404, in __del__
    signal.signal(signal.SIGINT, signal.SIG_DFL)
  File "~/miniconda3/envs/fastllama/lib/python3.8/signal.py", line 47, in signal
    handler = _signal.signal(_enum_to_int(signalnum), _enum_to_int(handler))
TypeError: signal handler must be signal.SIG_IGN, signal.SIG_DFL, or a callable object

[PotatoSpudowski]

Ah yes,
Good catch. I was trying to move away from my hacky implementation. In our old method, for large models, creating just one file was not a good idea because of high memory usage during conversion and quantization steps.

I wanted to add a way to split the models into smaller parts during the exporting part.
Please give me some time I will get it done by tonight.

Update: Using the convert.py script should work for alpaca models now as well as the base llama models

[PotatoSpudowski]

Meanwhile you can try the sync_llama branch for now.
All ggml models exported to huggingface work now. Been testing Vicuna and Alpaca 7B and 13B

[PotatoSpudowski]

This should be fixed in the new PR, can you please check?