Production-ready training recipes for state-of-the-art MoE models — DeepSeek-V3, Qwen3, and Mixtral — built on 🚀 Megatron-Core DEV branch.

✅ Performance-tuned configs for H100, B200, and GB200 clusters
✅ Model-specific best practices for training MoE models ✅ One-command launch with sensible defaults
✅ Dry-run mode to validate arguments before submitting jobs

Ready-to-run scripts with optimized configurations:

See best_practice/ for detailed guides.

Install yq for YAML processing (one-time setup):

mkdir -p ~/.local/bin && wget -qO ~/.local/bin/yq https://github.com/mikefarah/yq/releases/download/v4.27.5/yq_linux_amd64 && chmod +x ~/.local/bin/yq
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc && source ~/.bashrc

Dockerfile: dockers/Dockerfile (also available: B200.Dockerfile, GB200.Dockerfile)

Mixtral-8x2B, Mixtral-8x7B, Mixtral-8x22B, DeepSeek-V2, DeepSeek-V2-Lite, DeepSeek-V3, Qwen2-57B-A14B, Qwen3-235B-A22B, Qwen3-30B-A3B, Qwen3-Next-80B-A3B

Basic launch:

MODEL=DeepSeek-V3 bash ./sbatch_benchmarking.sh

With custom/overwritten parameters:

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

💡 Tip: Dry Run Mode — Preview the generated SLURM script and training command without submitting to the cluster:

DRY_RUN=1 MODEL=DeepSeek-V3 bash ./sbatch_benchmarking.sh

This is highly recommended before submitting jobs to verify configurations.

Runtime configs (runtime_configs/benchmarking/runtime.conf):

• Parallelism: TP, PP, EP, CP, VPP, PP_FIRST, PP_LAST, LAYERS_PER_VP
• Batch sizes: MBS, GBS
• Training: NNODES, RUN_TIME, NUM_LAYERS, SEQ_LEN
• MoE: MOE_TOKEN_DISPATCHER, MOE_GROUPED_GEMM

Cluster configs (cluster_configs/benchmarking/template.conf):

• Slurm: ACCOUNT, PARTITION, RUN_NAME, CONTAINER_MOUNTS
• Paths: OUTPUT_PATH, DATA_PATH, TOKENIZER_MODEL, LOAD_PATH

watch -n 1 squeue -u $USER

For HF↔MCore conversion, consider MBridge or Megatron-Bridge.

1. Download and convert to BF16:

git lfs install && git clone https://huggingface.co/deepseek-ai/DeepSeek-V3
python inference/fp8_cast_bf16.py --input-fp8-hf-path /input/fp8/path --output-bf16-hf-path /output/bf16/path

2. Convert to Megatron legacy checkpoint:

MODEL=DeepSeek-V3 bash ./ckpt_convert_scripts/DeepSeek-V3/convert_deepseek_v3.sh

3. Convert to 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/ckpt \
  bash ./sbatch_benchmarking.sh --ckpt-convert-save /path/to/dist/ckpt --ckpt-convert-format torch_dist --no-save-optim

Storage: Legacy ~3.4T, Distributed ~1.4T

• Design Docs - Implementation details for MTP, VPP, EP overlapping, etc.