training-llms-megatron
Trains large language models (2B-462B parameters) with NVIDIA Megatron-Core using tensor, pipeline, sequence, context, and expert parallelism, plus FP8 on H100 and MoE configuration for Mixtral-style models. Covers choosing TP/PP/DP/CP sizes, launching distributed training, tuning micro-batch size, and fixing low MFU, out-of-memory errors, and diverging loss. Use when training models above 10B parameters on NVIDIA A100/H100 GPUs, when setting up 3D parallelism for a LLaMA-style model, when configuring expert parallelism for MoE training, or when trying to raise MFU toward 40-47%. Use PyTorch FSDP, DeepSpeed, or HuggingFace Accelerate instead for models under 70B or simpler setups.
Install this skill
Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.
npx skills add KalarisLabs/research-agent-skills --skill training-llms-megatronuvx research-agent-skills install training-llms-megatronDocumentation & Instructions
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:
| Model Size | GPUs | Tensor Parallel | Pipeline Parallel | Data Parallel | Context Parallel |
|---|---|---|---|---|---|
| 7B | 8 | 1 | 1 | 8 | 1 |
| 13B | 8 | 2 | 1 | 4 | 1 |
| 70B | 64 | 4 | 4 | 4 | 1 |
| 405B | 128 | 8 | 8 | 2 | 2 |
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:
- Micro-batch too small
- Too much parallelism overhead
- 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:
- Network bottleneck: Ensure InfiniBand/NVLink enabled
- Pipeline bubbles: Use interleaved pipeline schedule
--num-layers-per-virtual-pipeline-stage 2 - 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)
Agent operating procedure
- Check the environment. Check GPU type, memory and driver/CUDA versions (
nvidia-smi), framework versions, and dataset location and size. - Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
- Run a small version first. Do a smoke run: tiny model or subset, few steps, and confirm loss decreases and checkpoints save.
- Execute the full task using the instructions and references above.
- Validate the result. Track metrics on held-out data, compare against a baseline, and record seeds, configs and hardware.
- Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.
| If this happens | Do this |
|---|---|
| CUDA out-of-memory | Reduce batch size, enable gradient accumulation/checkpointing or mixed precision, or shard the model. |
| Loss is NaN or diverges | Lower the learning rate, check data for invalid values, and enable gradient clipping. |
| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API. |
| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |
Integrity rules
- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
- Never claim training results without logs; estimate compute cost before launching large jobs.
- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.
Related skills
tensorrt-llm: Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency.moe-training: Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace.axolotl: Provides guidance for fine-tuning large language models with Axolotl, covering YAML training configs, LoRA and QLoRA, preference training w…
Trigger Prompts & Evaluation Corpus
These evaluation prompts test that the agent correctly selects training-llms-megatron for relevant tasks and leaves it out for near misses.
Should trigger (3)
- should trigger
“How should I split TP/PP/DP for a 70B LLaMA on 64 H100s with Megatron?”
expects Recommends a concrete layout such as TP=4, PP=4, DP=4 (CP=1), explains that TP stays within an NVLink node and PP spans nodes, and gives the launch arguments. Says to run a small smoke test and check memory and MFU before the full run.
- should trigger
“I'm pretraining a 13B GPT on 2 nodes of 8xA100-80GB with Megatron-LM, using pretrain_gpt.py and torchrun. At micro-batch 4 with TP=2, PP=1, I get 'CUDA out of memory' in the first forward pass. MFU was around 28% on a smaller test. What should I change, and in what order?”
expects Gives an ordered set of memory fixes: smaller micro-batch, activation recomputation, distributed optimizer, sequence parallel, and a higher TP or PP if needed. Also covers low-MFU causes such as Flash Attention, micro-batch size and interleaved pipeline schedule. Asks for or checks the GPU, CUDA and Megatron versions, and warns that settings depend on the installed version.
- should trigger
“We want to train a Mixtral-style 8x7B sparse model from scratch on a 64-GPU H100 cluster and keep the experts spread across devices so it fits. Our loss also spikes after a few thousand steps. How do we set it up?”
expects Explains how to configure expert parallelism (expert-model-parallel-size, number of experts, top-k, load-balancing/aux loss) together with TP/PP/DP, and how to launch it. For the loss spikes it suggests a lower learning rate, gradient clipping, a data check and a warmup review. Notes the option of FP8 on H100, and that settings should be validated against the installed version.
Near misses (2)
- near miss
“I have a 7B Llama checkpoint and want to fine-tune it with LoRA on 2 GPUs using a YAML config and a DPO dataset. What's the setup?”
expects Recognizes this as small-scale LoRA/DPO fine-tuning, not large-scale Megatron pretraining. Points to axolotl (or another PEFT tool) and does not give a Megatron 3D-parallelism setup.
- near miss
“Our trained model is ready. How do I serve it with FP8 quantization and in-flight batching on H100s for lowest latency?”
expects Recognizes this as inference deployment, not training. Points to TensorRT-LLM for FP8 quantization, in-flight batching and multi-GPU serving, and does not give Megatron training advice.