refactor ray integration, support save ckpt

Former-commit-id: 2f50b27e608b2092bfceab6c6e84e6631e973ee2

examples/README.md

@@ -95,6 +95,12 @@ FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500
 FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
+#### Supervised Fine-Tuning with Ray on 4 GPUs
+
+```bash
+USE_RAY=1 llamafactory-cli train examples/train_full/llama3_lora_sft_ray.yaml
+```
+
 ### QLoRA Fine-Tuning
 
 #### Supervised Fine-Tuning with 4/8-bit Bitsandbytes/HQQ/EETQ Quantization (Recommended)

examples/README_zh.md

@@ -95,6 +95,12 @@ FORCE_TORCHRUN=1 NNODES=2 NODE_RANK=1 MASTER_ADDR=192.168.0.1 MASTER_PORT=29500
 FORCE_TORCHRUN=1 llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml
 ```
 
+#### 使用 Ray 在 4 张 GPU 上微调
+
+```bash
+USE_RAY=1 llamafactory-cli train examples/train_full/llama3_lora_sft_ray.yaml
+```
+
 ### QLoRA 微调
 
 #### 基于 4/8 比特 Bitsandbytes/HQQ/EETQ 量化进行指令监督微调（推荐）

examples/train_lora/llama3_lora_sft.yaml

@@ -9,7 +9,6 @@ finetuning_type: lora
 lora_target: all
 
 ### dataset
-dataset_dir: /home/ray/default/LLaMA-Factory/data/
 dataset: identity,alpaca_en_demo
 template: llama3
 cutoff_len: 2048
@@ -39,10 +38,3 @@ val_size: 0.1
 per_device_eval_batch_size: 1
 eval_strategy: steps
 eval_steps: 500
-
-
-### ray setup
-resources_per_worker:
-  GPU: 1
-num_workers: 4
-# placement_strategy: ...

examples/train_lora/llama3_lora_sft_ray.yaml

@@ -0,0 +1,48 @@
+### model
+model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct  # or use local absolute path
+trust_remote_code: true
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: identity,alpaca_en_demo
+dataset_dir: REMOTE:llamafactory/demo_data  # or use local absolute path
+template: llama3
+cutoff_len: 2048
+max_samples: 1000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: tmp_dir
+logging_steps: 10
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 3.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 500
+
+### ray
+ray_run_name: llama3_8b_sft_lora
+ray_num_workers: 4  # number of GPUs to use
+resources_per_worker:
+  GPU: 1
+placement_strategy: PACK