yujiepan / deepseek-v2-0628-tiny-random

This model is for debugging. It is randomly initialized using the config from deepseek-ai/DeepSeek-V2-Chat-0628 but with smaller size.

Codes:

from huggingface_hub import create_repo, upload_folder
from transformers import (
    pipeline,
    set_seed,
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
)
import torch
import transformers
import os

model_id = "deepseek-ai/DeepSeek-V2-Chat-0628"
repo_id = "yujiepan/deepseek-v2-0628-tiny-random"
save_path = f"/tmp/{repo_id}"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
config._name_or_path = 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"
print(config)

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

model = AutoModelForCausalLM.from_config(
    config, torch_dtype=torch.bfloat16, attn_implementation="eager", trust_remote_code=True
)
model.generation_config = GenerationConfig.from_pretrained(model_id, trust_remote_code=True)

set_seed(42)
with torch.no_grad():
    for _, p in sorted(model.named_parameters()):
        torch.nn.init.uniform_(p, -0.1, 0.1)

model.save_pretrained(save_path)

# pipe = pipeline("text-generation", model=save_path, device="cuda", trust_remote_code=True)
# print(pipe("Hello World!"))

# messages = [
#     {"role": "system", "content": "You are a robot."},
#     {"role": "user", "content": "Hi!"},
# ]
# chatbot = pipeline("text-generation", model=save_path, max_length=1000, max_new_tokens=16, trust_remote_code=True)
# print(chatbot(messages))

messages = [{"role": "user", "content": "Write a piece of quicksort code in C++"}]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0][input_tensor.shape[1] :], skip_special_tokens=True)
print(result)