This model is for debugging. It is randomly initialized with the config from Qwen/QwQ-32B-Preview but is of smaller size.
Codes:
from transformers import AutoModelForCausalLM, AutoTokenizer
import transformers
import torch
import os
from huggingface_hub import create_repo, upload_folder
import accelerate
model_id = 'Qwen/QwQ-32B-Preview'
save_path = '/tmp/yujiepan/QwQ-preview-tiny-random'
repo_id = 'yujiepan/QwQ-preview-tiny-random'
os.system(f'rm -rf {save_path}')
config = transformers.AutoConfig.from_pretrained(
model_id,
trust_remote_code=True,
)
config._name_or_path = model_id
config.hidden_size = 8
config.intermediate_size = 16
config.num_key_value_heads = 1
config.num_attention_heads = 2
config.num_hidden_layers = 2
config.max_window_layers = 1
model = transformers.AutoModelForCausalLM.from_config(
config,
trust_remote_code=True,
)
model.generation_config = transformers.GenerationConfig.from_pretrained(
model_id)
model = model.to(torch.bfloat16)
transformers.set_seed(42)
num_params = 0
with torch.no_grad():
for name, p in sorted(model.named_parameters()):
print(name, p.shape)
torch.nn.init.uniform_(p, -0.5, 0.5)
num_params += p.numel()
print("Total number of parameters:", num_params)
model.save_pretrained(save_path)
tokenizer = transformers.AutoTokenizer.from_pretrained(
model_id,
trust_remote_code=True,
)
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)
def try_example(model, tokenizer):
prompt = "How many r in strawberry."
messages = [
{"role": "system", "content": "You are a helpful and harmless assistant. You should think step-by-step."},
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
**model_inputs,
max_new_tokens=32
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
try_example(model, tokenizer)