Quantization made by Richard Erkhov.
This model is randomly initialized, using the config from
microsoft/Phi-3-mini-128k-instruct but with smaller size.
Note the model is in float16.
1import transformers
2import torch
3import os
4from huggingface_hub import create_repo, upload_folder
5
6source_model_id = 'microsoft/Phi-3-mini-128k-instruct'
7save_path = '/tmp/yujiepan/phi-3-tiny-random'
8repo_id = 'yujiepan/phi-3-tiny-random'
9
10config = transformers.AutoConfig.from_pretrained(
11 source_model_id, trust_remote_code=True)
12config.hidden_size = 16
13config.intermediate_size = 32
14config.num_attention_heads = 4
15config.num_hidden_layers = 2
16config.num_key_value_heads = 4
17config.rope_scaling['long_factor'] = [1.0299, 1.0499]
18config.rope_scaling['short_factor'] = [1.05, 1.05]
19
20model = transformers.AutoModelForCausalLM.from_config(
21 config, trust_remote_code=True)
22model = model.to(torch.float16)
23model.save_pretrained(save_path)
24
25tokenizer = transformers.AutoTokenizer.from_pretrained(
26 source_model_id, trust_remote_code=True)
27tokenizer.save_pretrained(save_path)
28
29result = transformers.pipelines.pipeline(
30 'text-generation',
31 model=model.float(), tokenizer=tokenizer)('Hello')
32print(result)
33
34os.system(f'ls -alh {save_path}')
35create_repo(repo_id, exist_ok=True)
36upload_folder(repo_id=repo_id, folder_path=save_path)
37
38from transformers import AutoProcessor
39AutoProcessor.from_pretrained(source_model_id, trust_remote_code=True).push_to_hub(repo_id)
