This guide provides detailed instructions, best practices, and optimized configurations for testing Mixtral, DeepSeek, and Qwen series models using the Megatron-Core framework to achieve optimal performance and reliability.

• DeepSeek-V3 best practices in a single command
  • Current includes H100, B200, and Long Context. GB200 config coming soon.

• Container Setup
• Design Docs
• Environment Setup
• Performance Benchmarking
• DeepSeek Checkpoint Conversion

• Dockerfile: dockers/Dockerfile

Please refer to design_docs.

Before entering the container, you need to install yq to process .yaml configuration files.

Before running any scripts, you need to set up the following environment variables:

export WANDB_API_KEY="your_wandb_api_key_here"
export MEGATRON_PATH="/path/to/your/megatron/directory"
export MCORE_RELEASE_VERSION="0.13"
export CONTAINER_IMAGE="/path/to/container/image.sqsh"
export CLUSTER="your_cluster_name"

• WANDB_API_KEY: Your Weights & Biases API key for experiment tracking.
  • Get your key from wandb.ai/authorize.
• MEGATRON_PATH: Absolute path to your Megatron-MoE installation directory.
  • Example: path/to/Megatron-LM
• MCORE_RELEASE_VERSION: Version of Megatron-Core to use.
  • Currently recommended: 0.13
• CONTAINER_IMAGE: Path to the container image file (.sqsh).
  • Example: path/to/container/image.sqsh
• CLUSTER: Name of your cluster environment (e.g., EOS, CW).

For performance benchmarking, you can launch scripts either with sbatch via sbatch_benchmarking.sh or on an interactive node via interactive_benchmarking.sh.

• MODEL

  • This is a required environment variable that must be set in your script or command.
  • Predefined models include: Mixtral-8x2B, Mixtral-8x7B, Mixtral-8x22B, DeepSeek-V2, DeepSeek-V2-Lite, DeepSeek-V3, DeepSeek-V3-Lite, and Qwen2-57B-A14B.

• CLUSTER, MCORE_RELEASE_VERSION, and MEGATRON_PATH

  • These required variables must be defined in your script or command for proper execution.

• CONTAINER_IMAGE

• Using WandB for Experiment Tracking

  Click here to view WandB setup instructions.

  • To use WandB for experiment tracking, set WANDB_API_KEY with your key from wandb.ai/authorize. It is highly recommended to add export WANDB_API_KEY="your_own_wandb_api_key" to your ~/.bashrc.
  • If you do not wish to use WandB, comment out the following lines in your model's .yaml configuration file:

    # --wandb-project: wandb_project_name
    # --wandb-exp-name: wandb_experiment_name

All model-specific runner configurations can be adjusted through runtime_configs/benchmarking/runtime.conf or via the benchmarking command.

• Available Model-Specific Runner Configurations

  Click here to view available model-specific benchmarking configurations.

  • Parallel Mappings: TP, PP, EP, CP, VPP, PP_FIRST, PP_LAST, and LAYERS_PER_VP
  • Batch Sizes: MBS and GBS
  • Model Architecture: NUM_LAYERS
  • MoE Configurations: MOE_TOKEN_DISPATCHER, MOE_GROUPED_GEMM, and --moe-extended-ep
  • Training Configurations: NNODES, RUN_TIME, and PRETRAIN. Note that specifying a shorter run time may improve your job's priority in the Slurm queue.
  • Data Configurations: SEQ_LEN and DATASET

• All available optimial configurations are listed in runtime_configs/benchmarking/runtime.conf.

All cluster configurations can be customized either through cluster_configs/benchmarking/your_own_cluster.conf or via the benchmarking command. For guidance on creating your own cluster configurations, refer to the template provided in cluster_configs/benchmarking/template.conf.

• Required Cluster-Specific Slurm Settings: ACCOUNT, PARTITION, RUN_NAME, and CONTAINER_MOUNTS
• Required Cluster-Specific Paths: OUTPUT_PATH, DATA_PATH, TOKENIZER_MODEL, and LOAD_PATH

• To benchmark a model from scratch with preconfigured parameters:

  # Example for DeepSeek-V3
  MODEL=DeepSeek-V3 bash ./sbatch_benchmarking.sh

• To train a model with custom parameters:

  # Example for DeepSeek-V3
  MODEL=DeepSeek-V3 TP=2 PP=8 EP=64 VPP=1 PP_FIRST=8 PP_LAST=5 RUN_TIME=00:60:00 NNODES=64 bash sbatch_benchmarking.sh --recompute-granularity selective --recompute-modules mla_up_proj layernorm

• To monitor your jobs, use squeue -u $USER for a one-time status check or watch -n 1 squeue -u $USER for continuous monitoring. For detailed logging, refer to the WandB dashboard.

| Please try MBridge and Megatron-Bridge for better HF<->MCore conversion support.

Download the DeepSeek-V3 checkpoint from HuggingFace:

# Make sure git-lfs is installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/deepseek-ai/DeepSeek-V3

The downloaded checkpoint is in FP8 format. Run the following command to convert it to BF16 format, using this script:

python inference/fp8_cast_bf16.py --input-fp8-hf-path /your/input/fp8/hf/path --output-bf16-hf-path /your/output/bf16/hf/path

To convert the BF16 HuggingFace checkpoint to a Megatron legacy checkpoint, execute the following command:

# Example for DeepSeek-V3
MODEL=DeepSeek-V3 bash ./ckpt_convert_scripts/DeepSeek-V3/convert_deepseek_v3.sh

Finally, run this command to convert the legacy checkpoint into a distributed checkpoint:

MODEL=DeepSeek-V3 TP=1 PP=4 EP=64 VPP=1 PP_FIRST=16 PP_LAST=13 NNODES=32 LOAD_PATH=/path/to/legacy/checkpoint bash ./sbatch_benchmarking.sh --ckpt-convert-save /path/to/save/distributed/checkpoint --ckpt-convert-format torch_dist --no-save-optim

For reference, after conversion, the legacy checkpoint is approximately 3.4T, and the distributed checkpoint is about 1.4T.