Model Training

Model Training Skill

Launch and monitor model training on Ascend NPU environment.

When to Invoke

• User wants to start model training
• User wants to run performance training
• User wants to run accuracy training
• User asks about training options

Information to Collect

Ask user for the following:

1. Training mode preference:
   - Performance training (FP32/FP16)
   - Accuracy training (full epochs)
   - Custom training
   
2. Number of GPUs/NPUs to use

3. Batch size (if custom)

4. Any specific training parameters

Training Options

Provide user with training options:

Training Options:
1. Performance Training (FP32, 8 GPUs) - Quick performance test
2. Performance Training (FP16, 8 GPUs) - Quick performance test with mixed precision
3. Accuracy Training (24 epochs) - Full accuracy training
4. Custom training configuration

Workflow

Step 1: Select Training Mode

Ask user to select training mode or provide custom configuration.

Step 2: Prepare Environment

Ensure environment variables are set:

export ASCEND_SLOG_PRINT_TO_STDOUT=0
export ASCEND_GLOBAL_LOG_LEVEL=3
export TASK_QUEUE_ENABLE=2
export COMBINED_ENABLE=1
export HCCL_WHITELIST_DISABLE=1
export HCCL_IF_IP=$(hostname -I | awk '{print $1}')
export HCCL_CONNECT_TIMEOUT=1200

Step 3: Launch Training

Execute training script based on selected mode.

Step 4: Monitor Progress

Check training log and verify training started successfully.

BEVFormer Training Commands

Performance Training (FP32, 8 GPUs)

cd <working_directory>
bash test/train_performance_8p_base_fp32.sh --batch-size=1 --num-npu=8

Performance Training (FP16, 8 GPUs)

cd <working_directory>
bash test/train_performance_8p_base_fp16.sh --batch-size=1 --num-npu=8

Accuracy Training (24 epochs)

cd <working_directory>
bash test/train_full_8p.sh --batch-size=1

Custom Training

cd <model_directory>
bash ./tools/dist_train.sh ./projects/configs/bevformer/bevformer_base.py <num_gpus>

Training Verification

Check Training Started

# Check training process
ps aux | grep torchrun

# Check training log
tail -20 <working_directory>/test/output/train_performance_8p_base_fp32.log

Success Indicators

Training is considered successfully started when:

1. Log file shows training iterations
2. Loss values are being printed
3. No error messages in recent log
4. Process is running (torchrun)

Example successful log:

2026-03-12 09:47:40,838 - mmdet - INFO - Epoch [1][26/41]       lr: 7.333e-05, eta: 0:36:35, time: 1.883, data_time: 0.023, memory: 25332, loss_cls: 1.2614, loss_bbox: 1.7827

Monitoring Training

Check Training Log

tail -f <working_directory>/test/output/train_performance_8p_base_fp32.log

Key Metrics

• Loss values: loss_cls, loss_bbox, total loss
• Learning rate: lr
• Time per iteration: time
• Memory usage: memory
• ETA: estimated time remaining

Validation Metrics (per epoch)

• NDS: NuScenes Detection Score
• mAP: mean Average Precision
• mATE: mean Translation Error
• mASE: mean Scale Error
• mAOE: mean Orientation Error

Training Completion

Training is complete when:

1. All epochs finished
2. Final validation metrics printed
3. Checkpoint saved

Check for completion:

grep "Saving checkpoint" <log_file> | tail -1

Troubleshooting

Training Fails to Start

1. Check NPU devices: npu-smi info
2. Check environment variables
3. Check conda environment is activated
4. Check dataset and weights are linked

Out of Memory

1. Reduce batch size
2. Reduce number of GPUs
3. Enable gradient checkpointing

Training Hangs

1. Check NPU status: npu-smi info
2. Check for dead processes: ps aux | grep python
3. Check log for last activity

Port Already in Use

# Kill existing training processes
pkill -f torchrun
pkill -f dist_train

Log File Locations

Reference

• Training scripts: <model_directory>/test/
• Model config: <model_directory>/projects/configs/
• DrivingSDK README: DrivingSDK/model_examples/<model>/README.md