training-llms-megatron（训练 LLMS 威震天）

Megatron-Core - Large-Scale LLM Training

Quick start

Megatron-Core trains LLMs from 2B to 462B parameters with up to 47% Model FLOP Utilization on H100 GPUs through advanced parallelism strategies.

Installation:

# Docker (recommended)
docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:25.04-py3

# Or pip
pip install megatron-core

Simple distributed training:

# Train with 2 GPUs using data parallelism
torchrun --nproc_per_node=2 examples/run_simple_mcore_train_loop.py

# Or LLaMA-3 8B training
./examples/llama/train_llama3_8b_fp8.sh

Common workflows

Workflow 1: Train LLaMA-style model with 3D parallelism

Copy this checklist:

LLaMA Training Setup:
- [ ] Step 1: Choose parallelism configuration
- [ ] Step 2: Configure training hyperparameters
- [ ] Step 3: Launch distributed training
- [ ] Step 4: Monitor performance metrics

Step 1: Choose parallelism configuration

Model size determines parallelism strategy:

Step 2: Configure training hyperparameters

#!/bin/bash
# train_llama_70b.sh

GPUS_PER_NODE=8
NNODES=8  # 64 GPUs total
TP=4      # Tensor parallel
PP=4      # Pipeline parallel
CP=1      # Context parallel

# LLaMA 70B configuration
MODEL_SIZE=70  # Billion parameters
HIDDEN_SIZE=8192
NUM_LAYERS=80
NUM_HEADS=64
SEQ_LENGTH=4096

# Training hyperparameters
MICRO_BATCH=1
GLOBAL_BATCH=1024
LR=3e-4

torchrun \
  --nproc_per_node=$GPUS_PER_NODE \
  --nnodes=$NNODES \
  pretrain_gpt.py \
  --tensor-model-parallel-size $TP \
  --pipeline-model-parallel-size $PP \
  --context-parallel-size $CP \
  --sequence-parallel \
  --num-layers $NUM_LAYERS \
  --hidden-size $HIDDEN_SIZE \
  --num-attention-heads $NUM_HEADS \
  --seq-length $SEQ_LENGTH \
  --max-position-embeddings $SEQ_LENGTH \
  --micro-batch-size $MICRO_BATCH \
  --global-batch-size $GLOBAL_BATCH \
  --lr $LR \
  --train-iters 100000 \
  --lr-decay-style cosine \
  --lr-warmup-iters 2000 \
  --weight-decay 0.1 \
  --clip-grad 1.0 \
  --bf16 \
  --use-mcore-models \
  --transformer-impl transformer_engine \
  --data-path /path/to/data \
  --vocab-file /path/to/vocab.json \
  --merge-file /path/to/merges.txt

Step 3: Launch distributed training

# Single node (8 GPUs)
bash train_llama_70b.sh

# Multi-node with SLURM
sbatch --nodes=8 --gpus-per-node=8 train_llama_70b.sh

Step 4: Monitor performance metrics

Key metrics to track:

Model FLOP Utilization (MFU): Target >40% on H100
Throughput: Tokens/sec/GPU
Memory usage: <80GB per GPU for 70B model
Loss: Should decrease steadily

Workflow 2: Configure Mixture of Experts (MoE) training

For sparse MoE models like Mixtral.

MoE Training:
- [ ] Step 1: Configure expert parallelism
- [ ] Step 2: Set MoE hyperparameters
- [ ] Step 3: Launch training with EP

Step 1: Configure expert parallelism

# Mixtral 8x7B example
TENSOR_PARALLEL=2
PIPELINE_PARALLEL=1
EXPERT_PARALLEL=4  # Split 8 experts across 4 GPUs
DATA_PARALLEL=4

TOTAL_GPUS=$((TENSOR_PARALLEL * PIPELINE_PARALLEL * EXPERT_PARALLEL * DATA_PARALLEL))
# = 2 * 1 * 4 * 4 = 32 GPUs

Step 2: Set MoE hyperparameters

torchrun \
  --nproc_per_node=8 \
  pretrain_gpt.py \
  --tensor-model-parallel-size 2 \
  --pipeline-model-parallel-size 1 \
  --expert-model-parallel-size 4 \
  --num-experts 8 \
  --moe-router-topk 2 \
  --moe-router-load-balancing-type aux_loss \
  --moe-aux-loss-coeff 0.01 \
  --hidden-size 4096 \
  --num-layers 32 \
  --num-attention-heads 32 \
  --seq-length 4096 \
  --max-position-embeddings 4096 \
  --bf16 \
  --use-mcore-models \
  --transformer-impl transformer_engine \
  --data-path /path/to/data \
  --vocab-file /path/to/vocab.json \
  --merge-file /path/to/merges.txt

Step 3: Launch training with EP

Expert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity.

Memory without EP: 8 experts × 7B = 56GB per GPU
Memory with EP=4: 2 experts × 7B = 14GB per GPU
Savings: 75% memory reduction

Workflow 3: Optimize for maximum throughput

Achieve 47% MFU on H100.

Performance Optimization:
- [ ] Step 1: Enable Flash Attention
- [ ] Step 2: Use FP8 precision (H100)
- [ ] Step 3: Optimize micro-batch size
- [ ] Step 4: Tune parallelism degrees

Step 1: Enable optimizations

--use-mcore-models  # Use Megatron Core models
--transformer-impl transformer_engine  # Use Transformer Engine
--sequence-parallel  # Reduce activation memory (use with TP)

Step 2: Use FP8 precision (H100 only)

--fp8-hybrid  # FP8 mixed precision training
# Transformer Engine handles FP8 automatically

Result: 1.5-2x speedup on H100 vs BF16.

Step 3: Optimize micro-batch size

Find largest micro-batch that fits in memory:

# Start with 1, increase until OOM
for MBS in 1 2 4 8; do
  echo "Testing micro-batch-size=$MBS"
  torchrun ... --micro-batch-size $MBS
done

Typical values:

• 7B model: 4-8
• 70B model: 1-2
• 405B model: 1

Step 4: Tune parallelism degrees

Rules of thumb:

Tensor Parallel: Use ≤8 (limited by NVLink within node)
Pipeline Parallel: Use for >70B models
Context Parallel: Use for sequences >8K tokens
Data Parallel: Fill remaining GPUs

Example 405B on 128 H100s:

TP=8 (1 node)
PP=8 (across nodes)
CP=2 (long sequences)
DP=1
Total = 8 × 8 × 2 × 1 = 128 GPUs

When to use vs alternatives

Use Megatron-Core when:

• Training models >10B parameters
• Need maximum efficiency (target >40% MFU)
• Using NVIDIA GPUs (A100, H100)
• Production training at scale
• Want fine-grained parallelism control

Use alternatives instead:

• PyTorch FSDP: Models <70B, simpler API, PyTorch native
• DeepSpeed: Easier setup, good for <100B models
• HuggingFace Accelerate: Prototyping, simpler workflows
• LitGPT: Educational, single-file implementations

Common issues

Issue: Low GPU utilization (<30% MFU)

Causes:

1. Micro-batch too small
2. Too much parallelism overhead
3. Not using Flash Attention

Fixes:

# Increase micro-batch
--micro-batch-size 4  # Was 1

# Enable optimizations
--use-flash-attn
--sequence-parallel

# Reduce TP if >8
--tensor-model-parallel-size 4  # Was 16

Issue: Out of memory

Reduce memory with:

--tensor-model-parallel-size 2  # Split model across GPUs
--recompute-granularity full  # Gradient checkpointing
--recompute-method block  # Checkpoint transformer blocks
--recompute-num-layers 1  # Checkpoint every layer

Or use CPU/NVMe offloading:

--cpu-optimizer  # Offload optimizer to CPU
--cpu-optimizer-type ADAM  # CPU Adam variant

Issue: Training slower than expected

Check:

1. Network bottleneck: Ensure InfiniBand/NVLink enabled
2. Pipeline bubbles: Use interleaved pipeline schedule --num-layers-per-virtual-pipeline-stage 2
3. Data loading: Use fast data loader --dataloader-type cyclic

Issue: Diverging loss

Stabilize training:

--lr-warmup-iters 2000  # Longer warmup
--clip-grad 1.0  # Gradient clipping
--init-method-std 0.006  # Smaller init
--attention-dropout 0.0  # No dropout in attention
--hidden-dropout 0.0  # No dropout in FFN

Advanced topics

Parallelism strategies: See references/parallelism-guide.md for detailed comparison of TP/PP/DP/CP/EP with performance analysis and when to use each.

Performance benchmarks: See references/benchmarks.md for MFU numbers across different model sizes and GPU configurations.

Production configurations: See references/production-examples.md for real-world setups from LLaMA 3 405B, Nemotron-4 340B, and DeepSeek-V3 671B.

Training recipes: See references/training-recipes.md for complete hyperparameter configurations for GPT/LLaMA/Mixtral architectures.

Hardware requirements

• GPU: NVIDIA Ampere+ (A100, H100, B200)

- Turing works but slower - FP8 requires Hopper/Ada/Blackwell

• Network: InfiniBand or 400Gb+ Ethernet for multi-node
• Memory per GPU:

- 7B model: 40GB+ - 70B model: 80GB (with TP=4) - 405B model: 80GB (with TP=8, PP=8)

• Storage: Fast NVMe for checkpoints (1TB+ for 70B+ models)

Resources

• Docs: https://docs.nvidia.com/megatron-core/
• GitHub: https://github.com/NVIDIA/Megatron-LM
• Papers:

- "Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism" (2019) - "Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM" (2021)

• NeMo Framework: https://docs.nvidia.com/nemo-framework/ (built on Megatron-Core)