Add train loss

README.md

@@ -89,7 +89,7 @@ python infer_vllm.py --model_name mistralai/Ministral-8B-Instruct-2410
 
 1. 启动server
 ```bash
-vllm serve mistralai/Ministral-8B-Instruct-2410 --tokenizer_mode mistral --config_format mistral --load_format mistral
+vllm serve mistralai/Ministral-8B-Instruct-2410 --tokenizer-mode mistral --config-format mistral --load-format mistral
 ```
 
 2. client测试:
@@ -110,12 +110,20 @@ curl --location 'http://<your-node-url>:8000/v1/chat/completions' \
 ```
 
 ## result
+Prompt: "Do we need to think for 10 seconds to find the answer of 1 + 1?"
 <div align=center>
     <img src="./doc/results.png"/>
 </div>
 
 ### 精度
-DCU与GPU精度一致，推理框架：pytorch。
+训练框架：[Llama-Factory](https://developer.sourcefind.cn/codes/OpenDAS/llama-factory)
+
+训练脚本：[ministral_lora_sft.yaml](llama-factory/train_lora/ministral_lora_sft.yaml)
+
+|   device   |   train_loss   |
+|:----------:|:-------:|
+| DCU K100AI | 0.9068 |
+|  GPU A800  |  |
 
 ## 应用场景
 ### 算法类别

llama-factory/train_full/ministral_full_sft.yaml

@@ -17,12 +17,13 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4
 
 ### output
-output_dir: saves/ministral/full/sft
+output_dir: saves/ministral-8B/full/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
 
 ### train
 per_device_train_batch_size: 1

llama-factory/train_lora/ministral_lora_sft.yaml

@@ -20,12 +20,13 @@ preprocessing_num_workers: 16
 dataloader_num_workers: 4
 
 ### output
-output_dir: saves/ministral/lora/sft
+output_dir: saves/ministral-8B/lora/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 save_only_model: false
+report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]
 
 ### train
 per_device_train_batch_size: 1