yujiepan/qwen1.5-moe-tiny-random

This model is randomly initialized, using the config from Qwen/Qwen1.5-MoE-A2.7B-Chat but with smaller size. Note the model is in float16.

Codes:

import transformers
import torch
import os
from huggingface_hub import create_repo, upload_folder

source_model_id = 'Qwen/Qwen1.5-MoE-A2.7B-Chat'
save_path = '/tmp/yujiepan/qwen1.5-moe-tiny-random'
repo_id = 'yujiepan/qwen1.5-moe-tiny-random'

config = transformers.AutoConfig.from_pretrained(
    source_model_id, trust_remote_code=True)
config.hidden_size = 4
config.intermediate_size = 2
config.num_attention_heads = 4
config.num_hidden_layers = 2
config.num_key_value_heads = 2
config.moe_intermediate_size = 2
config.shared_expert_intermediate_size = 2
config.max_window_layers = 1
config.use_sliding_window = True
config.torch_dtype = torch.float16

model = transformers.AutoModelForCausalLM.from_config(
    config, trust_remote_code=True, torch_dtype=torch.float16)
model = model.half()

tokenizer = transformers.AutoTokenizer.from_pretrained(
    source_model_id, trust_remote_code=True)

result = transformers.pipelines.pipeline(
    'text-generation',
    model=model, tokenizer=tokenizer,
    device=0,
    max_new_tokens=16,
)('Hello World!')
print(result)

model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)

os.system(f'ls -alh {save_path}')
create_repo(repo_id, exist_ok=True)
upload_folder(repo_id=repo_id, folder_path=save_path)