import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda"

model = AutoModelForCausalLM.from_pretrained(
    'Kendamarron/Qwen2.5-7B-o1-ja-v0.1',
    torch_dtype=torch.bfloat16,
    device_map=device,
)
tokenizer = AutoTokenizer.from_pretrained('Kendamarron/Qwen2.5-7B-o1-ja-v0.1')

messages = [
  {"role": "system", "content": "あなたは優秀で論理的なアシスタントです。まずは<Thought></Thought>タグの中であなたの思考の過程を記載し、<Output></Output>タグの中に最終的にユーザーに提供する出力を記載します。"},
  {"role": "user", "content": "1から10までの整数を足すと？"}
]
text = tokenizer.apply_chat_template(
  messages,
  tokenize=False,
  add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(
  model_inputs.input_ids,
  max_new_tokens=256,
  do_sample=True,
  top_p=0.95,
  top_k=40,
  temperature=0.7,
  repetition_penalty=1.1,
  pad_token_id=tokenizer.eos_token_id,
  eos_token_id=tokenizer.eos_token_id,
  no_repeat_ngram_size=2
  )
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)

### model
model_name_or_path: Qwen/Qwen2.5-7B-Instruct

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: cot_normal, cot_math
template: qwen_ja
cutoff_len: 8192
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/qwen2.5/full/sft
logging_steps: 1
save_steps: 500
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 16
learning_rate: 1.0e-5
num_train_epochs: 2.0
lr_scheduler_type: cosine
optim: adamw_bnb_8bit
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

### eval
val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

### logging
report_to: wandb