yujiepan / deepseek-v3-tiny-random

This model is for debugging. It is randomly initialized with the config from deepseek-ai/DeepSeek-V3 but is of smaller size.

Codes:

import os

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

model_id = "deepseek-ai/DeepSeek-V3"
repo_id = "yujiepan/deepseek-v3-tiny-random"
save_path = f"/tmp/{repo_id}"
os.system(f"rm -rf {save_path}")

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
config.num_hidden_layers = 2
config.first_k_dense_replace = 1
config.hidden_size = 16
config.intermediate_size = 32
config.moe_intermediate_size = 16
config.q_lora_rank = 16
config.kv_lora_rank = 16
config.qk_rope_head_dim = 16
config.qk_nope_head_dim = 16
config.v_head_dim = 16
config.num_attention_heads = 2
config.num_key_value_heads = 2
# transformers has not supported the customized quantization config
del config.quantization_config

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

enable_full_determinism(seed=42)
model = AutoModelForCausalLM.from_config(
    config, torch_dtype=torch.bfloat16, trust_remote_code=True,
).eval()

try:
    model.generation_config = GenerationConfig.from_pretrained(
        model_id, trust_remote_code=True)
except:
    print("No generation config found")

num_params = 0
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        if 'experts' in name and 'experts.0.' not in name:  # avoid printing too much
            pass
        else:
            print(name, p.shape)
        # torch.nn.init.uniform_(p, -0.2, 0.2)
        num_params += p.numel()
print(f"Number of parameters: {num_params / 1e6:.2f}M")
model.save_pretrained(save_path)

del model
del tokenizer

torch.use_deterministic_algorithms(False)
tokenizer = AutoTokenizer.from_pretrained(save_path)
model = AutoModelForCausalLM.from_pretrained(
    save_path, trust_remote_code=True).eval()
prompt = 'Hello!'
messages = [
    {"role": "system", "content": "You are a helpful assistant."}
]
messages.append({"role": "user", "content": prompt})
tokenized_chat = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")

device = torch.device("cuda")
outputs = model.to(device).generate(
    tokenized_chat.to(device),
    max_new_tokens=16,
    do_sample=False,
    use_cache=True,
)
tokens = tokenizer.convert_ids_to_tokens(outputs[0])
string = tokenizer.decode(outputs[0])
print(tokens)

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