deepseek-v2-tiny-random

This model is randomly initialized, using the config from deepseek-ai/DeepSeek-V2-Chat but with smaller size. Note the model is in bfloat16.

Codes:

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

source_model_id = 'deepseek-ai/DeepSeek-V2-Chat'
save_path = '/tmp/yujiepan/deepseek-v2-tiny-random'
repo_id = 'yujiepan/deepseek-v2-tiny-random'

os.system(f'rm -rf {save_path}')

config = transformers.AutoConfig.from_pretrained(
    source_model_id,
    trust_remote_code=True,
)
config._name_or_path = source_model_id
config.hidden_size = 8
config.intermediate_size = 16
config.moe_intermediate_size = 4
config.num_attention_heads = 2
config.num_key_value_heads = 2
config.num_hidden_layers = 2
config.kv_lora_rank = 2
config.q_lora_rank = 2
config.v_head_dim = 2
config.qk_nope_head_dim = 2
config.qk_rope_head_dim = 2
config.torch_dtype = "bfloat16"

model = transformers.AutoModelForCausalLM.from_config(
    config,
    trust_remote_code=True,
)
model.generation_config = transformers.GenerationConfig.from_pretrained(source_model_id)
model = model.to(torch.bfloat16)
model.save_pretrained(save_path)

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

print(model.float().generate(torch.tensor([[1, 2, 3]]).long(), max_length=16))

os.system(f'ls -alh {save_path}')
# os.system(f'rm -rf {save_path}/model.safetensors')
create_repo(repo_id, exist_ok=True)
upload_folder(repo_id=repo_id, folder_path=save_path)