upload sft script

Author: Kunlun-Zhu

scripts/train_sft.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+set -x
+
+# Set default values for required parameters
+default_nproc=2
+default_save_path="./output"
+default_use_lora=true
+default_data_prefix="sft/data/"
+
+# Parse command line arguments
+nproc_per_node=${1:-$default_nproc}
+save_path=${2:-$default_save_path}
+use_lora=${3:-$default_use_lora}
+data_prefix=${4:-$default_data_prefix}
+
+# Shift the processed arguments
+shift $(($# > 3 ? 4 : $#))
+
+# Display usage information if needed
+if [ "$1" = "help" ] || [ "$1" = "--help" ] || [ "$1" = "-h" ]; then
+    echo "Usage: $(basename $0) [nproc_per_node] [save_path] [use_lora] [data_prefix] [other_configs...]"
+    echo "  nproc_per_node: Number of processes per node (default: $default_nproc)"
+    echo "  save_path: Directory to save model and logs (default: $default_save_path)"
+    echo "  use_lora: Whether to use LoRA for fine-tuning (true/false, default: $default_use_lora)"
+    echo "  data_prefix: Path prefix for training data (default: $default_data_prefix)"
+    exit 0
+fi
+
+# Create save directory if it doesn't exist
+if [ ! -d "$save_path" ]; then
+    mkdir -p "$save_path"
+    echo "Created directory: $save_path"
+fi
+
+# Setup LoRA parameters if enabled
+lora_config=""
+if [ "$use_lora" = true ]; then
+    lora_config="model.lora_rank=64 model.lora_alpha=32 model.target_modules=all-linear"
+fi
+
+# Extract environment type from data path for naming
+env_type=$(basename $data_prefix)
+
+# Generate a unique experiment name with timestamp
+timestamp=$(date +"%Y%m%d_%H%M%S")
+experiment_name="finetune_${env_type}_${timestamp}"
+
+# Run the training process
+echo "Starting fine-tuning"
+echo "Processes per node: $nproc_per_node"
+echo "Save path: $save_path"
+echo "Data path: $data_prefix"
+echo "LoRA enabled: $use_lora"
+
+torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
+     -m verl.trainer.fsdp_sft_trainer \
+    data.train_files=${data_prefix}/train.parquet \
+    data.val_files=${data_prefix}/test.parquet \
+    data.prompt_key=prompt \
+    data.response_key=response \
+    data.max_length=2048 \
+    optim.lr=1e-4 \
+    data.train_batch_size=128 \
+    data.micro_batch_size=4 \
+    model.partial_pretrain=Qwen/Qwen2.5-0.5B \
+    trainer.default_local_dir=$save_path \
+    trainer.experiment_name=$experiment_name \
+    trainer.logger=['console','wandb'] \
+    trainer.total_epochs=5 \
+    trainer.default_hdfs_dir=null \
+    trainer.validate_before_training=True \
+    model.enable_gradient_checkpointing=False \
+    $lora_config \
+    "$@" \
+    2>&1 | tee $save_path/train.log
+
+echo "Training completed. Logs saved to $save_path/train.log"