feat: support bloom-1b4-zh

[xingchensong]

• model link: https://huggingface.co/Langboat/bloom-1b4-zh this is a 1.4B model based on bigscience/bloom-1b7.
  The vocabulary is pruned from 250880 to 46145 with Chinese corpus to reduce GPU memory usage. So the total parameter is 1.4b now.
• results look good to me

fp16：

q4_0

[xingchensong]

PPL test on single sentence:

torch fp16	ggml fp16	ggml q8_0	ggml q5_0	ggml q4_0
1.9083	1.90837552	1.91213162	1.95716960	2.21781030

the code I used to test ppl is from this PR: #521

[ggerganov]

Thanks for this contribution!

Are you interested in trying to add support for this model straight into llama.cpp?
Recently, we have started to support many different model architectures and people will help out with testing and keeping up-to-date.

[xingchensong]

Thanks for this contribution!

Are you interested in trying to add support for this model straight into llama.cpp?
Recently, we have started to support many different model architectures and people will help out with testing and keeping up-to-date.

sure, will do it ASAP

[ggerganov]

Great! Take a look at this PR to get an idea of what is needed:

ggerganov/llama.cpp#3187