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.
- License 许可
- MIT
- 最近更新
- 2026-11-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- training-llms-megatron
- license
- MIT
- metadata
- { "tags": "Megatron-Core, Large-Scale Training, NVIDIA, Tensor Parallelism, Pipeline Parallelism, Model Parallelism, H100, Distributed Training, Production", "version": "1.0.0", "category": "ml-training", "maintainer": "Kalaris Labs", "dependencies": "megatron-core, torch, apex, transformer-engine" }
- description
- 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.
02 资源文件 4
- 参考文档references/benchmarks.md7.2 KB
- 参考文档references/parallelism-guide.md9.4 KB
- 参考文档references/production-examples.md10.6 KB
- 参考文档references/training-recipes.md11.3 KB
03 SKILL.md 正文
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…
04 同仓库更多 Skills
youtube-full
Use when the user needs YouTube transcripts, video search, channel browsing, playlist extraction, or content monitoring. Trigger phrases: 'get the transcript for', 'search YouTube for', 'what are the latest videos on', 'list this playlist', 'monitor this channel', or any request involving a YouTube URL, video ID, or @handle. Do NOT use for downloading video or audio files, YouTube engagement data (likes, comments), or private/age-restricted videos.
x-twitter-growth
X/Twitter growth engine for building audience, crafting viral content, and analyzing engagement. Use when the user wants to grow on X/Twitter, write tweets or threads, analyze their X profile, research competitors on X, plan a posting strategy, or optimize engagement. Complements social-content (generic multi-platform) with X-specific depth: algorithm mechanics, thread engineering, reply strategy, profile optimization, and competitive intelligence via web search.
write-a-skill
Create new agent skills with proper structure, progressive disclosure, and bundled resources. Use when user wants to create, write, build, or author a new skill.
wrap-up
Close out a launched Claude Managed Agent — recap every primitive the founder now owns, regenerate the single-file overview page, and suggest the next 1-2 upgrades. Use when the user says "wrap up", "close this out", "what do I own now", "give me the summary", "recap the agent", or when the orchestrator routes phase=wrap-up. primitives_inventory.py tables everything owned (agent, environment, session, memory, outcome, deployment); overview_page.py regenerates a self-contained ./my-agent/agent-overview.html; upgrade_suggester.py ranks the next moves from recorded deferrals plus standing hardening steps. Companion to run-without-you; the last stop before phase=done.
workflow-builder
Design and write deterministic multi-agent workflow scripts (.js files in .claude/workflows/) for Claude Code's Workflow tool. Use when a user wants to build, create, author, scaffold, or run a custom Claude Code workflow, orchestrate sub-agents (fan-out, pipeline, loop, judge-panel), or automate a repeatable multi-step task across fresh-context agents.
weekly-review
Use when someone wants to run a weekly review, close open loops, audit stalled projects and commitments, get their system back to trusted, restart a lapsed review habit, or says "/cs:weekly-review". Walks David Allen's three-phase loop — GET CLEAR, GET CURRENT, GET CREATIVE — with deterministic scripts that inventory open loops, gate the checklist with named gaps, and score commitment health 0-100.