[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M758PH2NDNFZHVZDX8GSE8PH":3,"skill-same:01M758PH2NDNFZHVZDX8GSE8PH":85},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":41,"commitSha":42,"skillDirectory":43,"forksCount":14,"repositoryHtmlUrl":44,"repositoryHomepageUrl":6,"ownerHtmlUrl":45,"ownerBlogUrl":46,"ownerDescription":47,"updatedAt":48,"install":49,"resources":62},"01M758PH2NDNFZHVZDX8GSE8PH","training-llms-megatron",null,"Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies. Use when training models >1B parameters, need maximum GPU efficiency (47% MFU on H100), or require tensor\u002Fpipeline\u002Fsequence\u002Fcontext\u002Fexpert parallelism. Production-ready framework used for Nemotron, LLaMA, DeepSeek.","zengrong233\u002Fdevin-skills","zengrong233","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F55529920?v=4","01M758DNYKQQCQX87XFVAGX5MW","songzengrong","01M758DNZVEJJQNM71DSRVHFEB",0,"MIT","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: training-llms-megatron\ndescription: Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies. Use when training models >1B parameters, need maximum GPU efficiency (47% MFU on H100), or require tensor\u002Fpipeline\u002Fsequence\u002Fcontext\u002Fexpert parallelism. Production-ready framework used for Nemotron, LLaMA, DeepSeek.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ntags: [Megatron-Core, Large-Scale Training, NVIDIA, Tensor Parallelism, Pipeline Parallelism, Model Parallelism, H100, Distributed Training, Production]\ndependencies: [megatron-core, torch, apex, transformer-engine]\n---\n\n# Megatron-Core - Large-Scale LLM Training\n\n## Quick start\n\nMegatron-Core trains LLMs from 2B to 462B parameters with up to 47% Model FLOP Utilization on H100 GPUs through advanced parallelism strategies.\n\n**Installation**:\n```bash\n# Docker (recommended)\ndocker run --gpus all -it --rm nvcr.io\u002Fnvidia\u002Fpytorch:25.04-py3\n\n# Or pip\npip install megatron-core\n```\n\n**Simple distributed training**:\n```bash\n# Train with 2 GPUs using data parallelism\ntorchrun --nproc_per_node=2 examples\u002Frun_simple_mcore_train_loop.py\n\n# Or LLaMA-3 8B training\n.\u002Fexamples\u002Fllama\u002Ftrain_llama3_8b_fp8.sh\n```\n\n## Common workflows\n\n### Workflow 1: Train LLaMA-style model with 3D parallelism\n\nCopy this checklist:\n\n```\nLLaMA Training Setup:\n- [ ] Step 1: Choose parallelism configuration\n- [ ] Step 2: Configure training hyperparameters\n- [ ] Step 3: Launch distributed training\n- [ ] Step 4: Monitor performance metrics\n```\n\n**Step 1: Choose parallelism configuration**\n\nModel size determines parallelism strategy:\n\n| Model Size | GPUs | Tensor Parallel | Pipeline Parallel | Data Parallel | Context Parallel |\n|------------|------|-----------------|-------------------|---------------|------------------|\n| 7B | 8 | 1 | 1 | 8 | 1 |\n| 13B | 8 | 2 | 1 | 4 | 1 |\n| 70B | 64 | 4 | 4 | 4 | 1 |\n| 405B | 128 | 8 | 8 | 2 | 2 |\n\n**Step 2: Configure training hyperparameters**\n\n```bash\n#!\u002Fbin\u002Fbash\n# train_llama_70b.sh\n\nGPUS_PER_NODE=8\nNNODES=8  # 64 GPUs total\nTP=4      # Tensor parallel\nPP=4      # Pipeline parallel\nCP=1      # Context parallel\n\n# LLaMA 70B configuration\nMODEL_SIZE=70  # Billion parameters\nHIDDEN_SIZE=8192\nNUM_LAYERS=80\nNUM_HEADS=64\nSEQ_LENGTH=4096\n\n# Training hyperparameters\nMICRO_BATCH=1\nGLOBAL_BATCH=1024\nLR=3e-4\n\ntorchrun \\\n  --nproc_per_node=$GPUS_PER_NODE \\\n  --nnodes=$NNODES \\\n  pretrain_gpt.py \\\n  --tensor-model-parallel-size $TP \\\n  --pipeline-model-parallel-size $PP \\\n  --context-parallel-size $CP \\\n  --sequence-parallel \\\n  --num-layers $NUM_LAYERS \\\n  --hidden-size $HIDDEN_SIZE \\\n  --num-attention-heads $NUM_HEADS \\\n  --seq-length $SEQ_LENGTH \\\n  --max-position-embeddings $SEQ_LENGTH \\\n  --micro-batch-size $MICRO_BATCH \\\n  --global-batch-size $GLOBAL_BATCH \\\n  --lr $LR \\\n  --train-iters 100000 \\\n  --lr-decay-style cosine \\\n  --lr-warmup-iters 2000 \\\n  --weight-decay 0.1 \\\n  --clip-grad 1.0 \\\n  --bf16 \\\n  --use-mcore-models \\\n  --transformer-impl transformer_engine \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n**Step 3: Launch distributed training**\n\n```bash\n# Single node (8 GPUs)\nbash train_llama_70b.sh\n\n# Multi-node with SLURM\nsbatch --nodes=8 --gpus-per-node=8 train_llama_70b.sh\n```\n\n**Step 4: Monitor performance metrics**\n\nKey metrics to track:\n```\nModel FLOP Utilization (MFU): Target >40% on H100\nThroughput: Tokens\u002Fsec\u002FGPU\nMemory usage: \u003C80GB per GPU for 70B model\nLoss: Should decrease steadily\n```\n\n### Workflow 2: Configure Mixture of Experts (MoE) training\n\nFor sparse MoE models like Mixtral.\n\n```\nMoE Training:\n- [ ] Step 1: Configure expert parallelism\n- [ ] Step 2: Set MoE hyperparameters\n- [ ] Step 3: Launch training with EP\n```\n\n**Step 1: Configure expert parallelism**\n\n```bash\n# Mixtral 8x7B example\nTENSOR_PARALLEL=2\nPIPELINE_PARALLEL=1\nEXPERT_PARALLEL=4  # Split 8 experts across 4 GPUs\nDATA_PARALLEL=4\n\nTOTAL_GPUS=$((TENSOR_PARALLEL * PIPELINE_PARALLEL * EXPERT_PARALLEL * DATA_PARALLEL))\n# = 2 * 1 * 4 * 4 = 32 GPUs\n```\n\n**Step 2: Set MoE hyperparameters**\n\n```bash\ntorchrun \\\n  --nproc_per_node=8 \\\n  pretrain_gpt.py \\\n  --tensor-model-parallel-size 2 \\\n  --pipeline-model-parallel-size 1 \\\n  --expert-model-parallel-size 4 \\\n  --num-experts 8 \\\n  --moe-router-topk 2 \\\n  --moe-router-load-balancing-type aux_loss \\\n  --moe-aux-loss-coeff 0.01 \\\n  --hidden-size 4096 \\\n  --num-layers 32 \\\n  --num-attention-heads 32 \\\n  --seq-length 4096 \\\n  --max-position-embeddings 4096 \\\n  --bf16 \\\n  --use-mcore-models \\\n  --transformer-impl transformer_engine \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n**Step 3: Launch training with EP**\n\nExpert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity.\n\n```\nMemory without EP: 8 experts × 7B = 56GB per GPU\nMemory with EP=4: 2 experts × 7B = 14GB per GPU\nSavings: 75% memory reduction\n```\n\n### Workflow 3: Optimize for maximum throughput\n\nAchieve 47% MFU on H100.\n\n```\nPerformance Optimization:\n- [ ] Step 1: Enable Flash Attention\n- [ ] Step 2: Use FP8 precision (H100)\n- [ ] Step 3: Optimize micro-batch size\n- [ ] Step 4: Tune parallelism degrees\n```\n\n**Step 1: Enable optimizations**\n\n```bash\n--use-mcore-models  # Use Megatron Core models\n--transformer-impl transformer_engine  # Use Transformer Engine\n--sequence-parallel  # Reduce activation memory (use with TP)\n```\n\n**Step 2: Use FP8 precision (H100 only)**\n\n```bash\n--fp8-hybrid  # FP8 mixed precision training\n# Transformer Engine handles FP8 automatically\n```\n\nResult: 1.5-2x speedup on H100 vs BF16.\n\n**Step 3: Optimize micro-batch size**\n\nFind largest micro-batch that fits in memory:\n\n```bash\n# Start with 1, increase until OOM\nfor MBS in 1 2 4 8; do\n  echo \"Testing micro-batch-size=$MBS\"\n  torchrun ... --micro-batch-size $MBS\ndone\n```\n\nTypical values:\n- 7B model: 4-8\n- 70B model: 1-2\n- 405B model: 1\n\n**Step 4: Tune parallelism degrees**\n\nRules of thumb:\n```\nTensor Parallel: Use ≤8 (limited by NVLink within node)\nPipeline Parallel: Use for >70B models\nContext Parallel: Use for sequences >8K tokens\nData Parallel: Fill remaining GPUs\n```\n\nExample 405B on 128 H100s:\n```\nTP=8 (1 node)\nPP=8 (across nodes)\nCP=2 (long sequences)\nDP=1\nTotal = 8 × 8 × 2 × 1 = 128 GPUs\n```\n\n## When to use vs alternatives\n\n**Use Megatron-Core when:**\n- Training models >10B parameters\n- Need maximum efficiency (target >40% MFU)\n- Using NVIDIA GPUs (A100, H100)\n- Production training at scale\n- Want fine-grained parallelism control\n\n**Use alternatives instead:**\n- **PyTorch FSDP**: Models \u003C70B, simpler API, PyTorch native\n- **DeepSpeed**: Easier setup, good for \u003C100B models\n- **HuggingFace Accelerate**: Prototyping, simpler workflows\n- **LitGPT**: Educational, single-file implementations\n\n## Common issues\n\n**Issue: Low GPU utilization (\u003C30% MFU)**\n\nCauses:\n1. Micro-batch too small\n2. Too much parallelism overhead\n3. Not using Flash Attention\n\nFixes:\n```bash\n# Increase micro-batch\n--micro-batch-size 4  # Was 1\n\n# Enable optimizations\n--use-flash-attn\n--sequence-parallel\n\n# Reduce TP if >8\n--tensor-model-parallel-size 4  # Was 16\n```\n\n**Issue: Out of memory**\n\nReduce memory with:\n```bash\n--tensor-model-parallel-size 2  # Split model across GPUs\n--recompute-granularity full  # Gradient checkpointing\n--recompute-method block  # Checkpoint transformer blocks\n--recompute-num-layers 1  # Checkpoint every layer\n```\n\nOr use CPU\u002FNVMe offloading:\n```bash\n--cpu-optimizer  # Offload optimizer to CPU\n--cpu-optimizer-type ADAM  # CPU Adam variant\n```\n\n**Issue: Training slower than expected**\n\nCheck:\n1. **Network bottleneck**: Ensure InfiniBand\u002FNVLink enabled\n2. **Pipeline bubbles**: Use interleaved pipeline schedule\n   ```bash\n   --num-layers-per-virtual-pipeline-stage 2\n   ```\n3. **Data loading**: Use fast data loader\n   ```bash\n   --dataloader-type cyclic\n   ```\n\n**Issue: Diverging loss**\n\nStabilize training:\n```bash\n--lr-warmup-iters 2000  # Longer warmup\n--clip-grad 1.0  # Gradient clipping\n--init-method-std 0.006  # Smaller init\n--attention-dropout 0.0  # No dropout in attention\n--hidden-dropout 0.0  # No dropout in FFN\n```\n\n## Advanced topics\n\n**Parallelism strategies**: See [references\u002Fparallelism-guide.md](references\u002Fparallelism-guide.md) for detailed comparison of TP\u002FPP\u002FDP\u002FCP\u002FEP with performance analysis and when to use each.\n\n**Performance benchmarks**: See [references\u002Fbenchmarks.md](references\u002Fbenchmarks.md) for MFU numbers across different model sizes and GPU configurations.\n\n**Production configurations**: See [references\u002Fproduction-examples.md](references\u002Fproduction-examples.md) for real-world setups from LLaMA 3 405B, Nemotron-4 340B, and DeepSeek-V3 671B.\n\n**Training recipes**: See [references\u002Ftraining-recipes.md](references\u002Ftraining-recipes.md) for complete hyperparameter configurations for GPT\u002FLLaMA\u002FMixtral architectures.\n\n## Hardware requirements\n\n- **GPU**: NVIDIA Ampere+ (A100, H100, B200)\n  - Turing works but slower\n  - FP8 requires Hopper\u002FAda\u002FBlackwell\n- **Network**: InfiniBand or 400Gb+ Ethernet for multi-node\n- **Memory per GPU**:\n  - 7B model: 40GB+\n  - 70B model: 80GB (with TP=4)\n  - 405B model: 80GB (with TP=8, PP=8)\n- **Storage**: Fast NVMe for checkpoints (1TB+ for 70B+ models)\n\n## Resources\n\n- Docs: https:\u002F\u002Fdocs.nvidia.com\u002Fmegatron-core\u002F\n- GitHub: https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM\n- Papers:\n  - \"Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism\" (2019)\n  - \"Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM\" (2021)\n- NeMo Framework: https:\u002F\u002Fdocs.nvidia.com\u002Fnemo-framework\u002F (built on Megatron-Core)\n\n\n","# Megatron-Core - Large-Scale LLM Training\n\n## Quick start\n\nMegatron-Core trains LLMs from 2B to 462B parameters with up to 47% Model FLOP Utilization on H100 GPUs through advanced parallelism strategies.\n\n**Installation**:\n```bash\n# Docker (recommended)\ndocker run --gpus all -it --rm nvcr.io\u002Fnvidia\u002Fpytorch:25.04-py3\n\n# Or pip\npip install megatron-core\n```\n\n**Simple distributed training**:\n```bash\n# Train with 2 GPUs using data parallelism\ntorchrun --nproc_per_node=2 examples\u002Frun_simple_mcore_train_loop.py\n\n# Or LLaMA-3 8B training\n.\u002Fexamples\u002Fllama\u002Ftrain_llama3_8b_fp8.sh\n```\n\n## Common workflows\n\n### Workflow 1: Train LLaMA-style model with 3D parallelism\n\nCopy this checklist:\n\n```\nLLaMA Training Setup:\n- [ ] Step 1: Choose parallelism configuration\n- [ ] Step 2: Configure training hyperparameters\n- [ ] Step 3: Launch distributed training\n- [ ] Step 4: Monitor performance metrics\n```\n\n**Step 1: Choose parallelism configuration**\n\nModel size determines parallelism strategy:\n\n| Model Size | GPUs | Tensor Parallel | Pipeline Parallel | Data Parallel | Context Parallel |\n|------------|------|-----------------|-------------------|---------------|------------------|\n| 7B | 8 | 1 | 1 | 8 | 1 |\n| 13B | 8 | 2 | 1 | 4 | 1 |\n| 70B | 64 | 4 | 4 | 4 | 1 |\n| 405B | 128 | 8 | 8 | 2 | 2 |\n\n**Step 2: Configure training hyperparameters**\n\n```bash\n#!\u002Fbin\u002Fbash\n# train_llama_70b.sh\n\nGPUS_PER_NODE=8\nNNODES=8  # 64 GPUs total\nTP=4      # Tensor parallel\nPP=4      # Pipeline parallel\nCP=1      # Context parallel\n\n# LLaMA 70B configuration\nMODEL_SIZE=70  # Billion parameters\nHIDDEN_SIZE=8192\nNUM_LAYERS=80\nNUM_HEADS=64\nSEQ_LENGTH=4096\n\n# Training hyperparameters\nMICRO_BATCH=1\nGLOBAL_BATCH=1024\nLR=3e-4\n\ntorchrun \\\n  --nproc_per_node=$GPUS_PER_NODE \\\n  --nnodes=$NNODES \\\n  pretrain_gpt.py \\\n  --tensor-model-parallel-size $TP \\\n  --pipeline-model-parallel-size $PP \\\n  --context-parallel-size $CP \\\n  --sequence-parallel \\\n  --num-layers $NUM_LAYERS \\\n  --hidden-size $HIDDEN_SIZE \\\n  --num-attention-heads $NUM_HEADS \\\n  --seq-length $SEQ_LENGTH \\\n  --max-position-embeddings $SEQ_LENGTH \\\n  --micro-batch-size $MICRO_BATCH \\\n  --global-batch-size $GLOBAL_BATCH \\\n  --lr $LR \\\n  --train-iters 100000 \\\n  --lr-decay-style cosine \\\n  --lr-warmup-iters 2000 \\\n  --weight-decay 0.1 \\\n  --clip-grad 1.0 \\\n  --bf16 \\\n  --use-mcore-models \\\n  --transformer-impl transformer_engine \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n**Step 3: Launch distributed training**\n\n```bash\n# Single node (8 GPUs)\nbash train_llama_70b.sh\n\n# Multi-node with SLURM\nsbatch --nodes=8 --gpus-per-node=8 train_llama_70b.sh\n```\n\n**Step 4: Monitor performance metrics**\n\nKey metrics to track:\n```\nModel FLOP Utilization (MFU): Target >40% on H100\nThroughput: Tokens\u002Fsec\u002FGPU\nMemory usage: \u003C80GB per GPU for 70B model\nLoss: Should decrease steadily\n```\n\n### Workflow 2: Configure Mixture of Experts (MoE) training\n\nFor sparse MoE models like Mixtral.\n\n```\nMoE Training:\n- [ ] Step 1: Configure expert parallelism\n- [ ] Step 2: Set MoE hyperparameters\n- [ ] Step 3: Launch training with EP\n```\n\n**Step 1: Configure expert parallelism**\n\n```bash\n# Mixtral 8x7B example\nTENSOR_PARALLEL=2\nPIPELINE_PARALLEL=1\nEXPERT_PARALLEL=4  # Split 8 experts across 4 GPUs\nDATA_PARALLEL=4\n\nTOTAL_GPUS=$((TENSOR_PARALLEL * PIPELINE_PARALLEL * EXPERT_PARALLEL * DATA_PARALLEL))\n# = 2 * 1 * 4 * 4 = 32 GPUs\n```\n\n**Step 2: Set MoE hyperparameters**\n\n```bash\ntorchrun \\\n  --nproc_per_node=8 \\\n  pretrain_gpt.py \\\n  --tensor-model-parallel-size 2 \\\n  --pipeline-model-parallel-size 1 \\\n  --expert-model-parallel-size 4 \\\n  --num-experts 8 \\\n  --moe-router-topk 2 \\\n  --moe-router-load-balancing-type aux_loss \\\n  --moe-aux-loss-coeff 0.01 \\\n  --hidden-size 4096 \\\n  --num-layers 32 \\\n  --num-attention-heads 32 \\\n  --seq-length 4096 \\\n  --max-position-embeddings 4096 \\\n  --bf16 \\\n  --use-mcore-models \\\n  --transformer-impl transformer_engine \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n**Step 3: Launch training with EP**\n\nExpert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity.\n\n```\nMemory without EP: 8 experts × 7B = 56GB per GPU\nMemory with EP=4: 2 experts × 7B = 14GB per GPU\nSavings: 75% memory reduction\n```\n\n### Workflow 3: Optimize for maximum throughput\n\nAchieve 47% MFU on H100.\n\n```\nPerformance Optimization:\n- [ ] Step 1: Enable Flash Attention\n- [ ] Step 2: Use FP8 precision (H100)\n- [ ] Step 3: Optimize micro-batch size\n- [ ] Step 4: Tune parallelism degrees\n```\n\n**Step 1: Enable optimizations**\n\n```bash\n--use-mcore-models  # Use Megatron Core models\n--transformer-impl transformer_engine  # Use Transformer Engine\n--sequence-parallel  # Reduce activation memory (use with TP)\n```\n\n**Step 2: Use FP8 precision (H100 only)**\n\n```bash\n--fp8-hybrid  # FP8 mixed precision training\n# Transformer Engine handles FP8 automatically\n```\n\nResult: 1.5-2x speedup on H100 vs BF16.\n\n**Step 3: Optimize micro-batch size**\n\nFind largest micro-batch that fits in memory:\n\n```bash\n# Start with 1, increase until OOM\nfor MBS in 1 2 4 8; do\n  echo \"Testing micro-batch-size=$MBS\"\n  torchrun ... --micro-batch-size $MBS\ndone\n```\n\nTypical values:\n- 7B model: 4-8\n- 70B model: 1-2\n- 405B model: 1\n\n**Step 4: Tune parallelism degrees**\n\nRules of thumb:\n```\nTensor Parallel: Use ≤8 (limited by NVLink within node)\nPipeline Parallel: Use for >70B models\nContext Parallel: Use for sequences >8K tokens\nData Parallel: Fill remaining GPUs\n```\n\nExample 405B on 128 H100s:\n```\nTP=8 (1 node)\nPP=8 (across nodes)\nCP=2 (long sequences)\nDP=1\nTotal = 8 × 8 × 2 × 1 = 128 GPUs\n```\n\n## When to use vs alternatives\n\n**Use Megatron-Core when:**\n- Training models >10B parameters\n- Need maximum efficiency (target >40% MFU)\n- Using NVIDIA GPUs (A100, H100)\n- Production training at scale\n- Want fine-grained parallelism control\n\n**Use alternatives instead:**\n- **PyTorch FSDP**: Models \u003C70B, simpler API, PyTorch native\n- **DeepSpeed**: Easier setup, good for \u003C100B models\n- **HuggingFace Accelerate**: Prototyping, simpler workflows\n- **LitGPT**: Educational, single-file implementations\n\n## Common issues\n\n**Issue: Low GPU utilization (\u003C30% MFU)**\n\nCauses:\n1. Micro-batch too small\n2. Too much parallelism overhead\n3. Not using Flash Attention\n\nFixes:\n```bash\n# Increase micro-batch\n--micro-batch-size 4  # Was 1\n\n# Enable optimizations\n--use-flash-attn\n--sequence-parallel\n\n# Reduce TP if >8\n--tensor-model-parallel-size 4  # Was 16\n```\n\n**Issue: Out of memory**\n\nReduce memory with:\n```bash\n--tensor-model-parallel-size 2  # Split model across GPUs\n--recompute-granularity full  # Gradient checkpointing\n--recompute-method block  # Checkpoint transformer blocks\n--recompute-num-layers 1  # Checkpoint every layer\n```\n\nOr use CPU\u002FNVMe offloading:\n```bash\n--cpu-optimizer  # Offload optimizer to CPU\n--cpu-optimizer-type ADAM  # CPU Adam variant\n```\n\n**Issue: Training slower than expected**\n\nCheck:\n1. **Network bottleneck**: Ensure InfiniBand\u002FNVLink enabled\n2. **Pipeline bubbles**: Use interleaved pipeline schedule\n   ```bash\n   --num-layers-per-virtual-pipeline-stage 2\n   ```\n3. **Data loading**: Use fast data loader\n   ```bash\n   --dataloader-type cyclic\n   ```\n\n**Issue: Diverging loss**\n\nStabilize training:\n```bash\n--lr-warmup-iters 2000  # Longer warmup\n--clip-grad 1.0  # Gradient clipping\n--init-method-std 0.006  # Smaller init\n--attention-dropout 0.0  # No dropout in attention\n--hidden-dropout 0.0  # No dropout in FFN\n```\n\n## Advanced topics\n\n**Parallelism strategies**: See [references\u002Fparallelism-guide.md](references\u002Fparallelism-guide.md) for detailed comparison of TP\u002FPP\u002FDP\u002FCP\u002FEP with performance analysis and when to use each.\n\n**Performance benchmarks**: See [references\u002Fbenchmarks.md](references\u002Fbenchmarks.md) for MFU numbers across different model sizes and GPU configurations.\n\n**Production configurations**: See [references\u002Fproduction-examples.md](references\u002Fproduction-examples.md) for real-world setups from LLaMA 3 405B, Nemotron-4 340B, and DeepSeek-V3 671B.\n\n**Training recipes**: See [references\u002Ftraining-recipes.md](references\u002Ftraining-recipes.md) for complete hyperparameter configurations for GPT\u002FLLaMA\u002FMixtral architectures.\n\n## Hardware requirements\n\n- **GPU**: NVIDIA Ampere+ (A100, H100, B200)\n  - Turing works but slower\n  - FP8 requires Hopper\u002FAda\u002FBlackwell\n- **Network**: InfiniBand or 400Gb+ Ethernet for multi-node\n- **Memory per GPU**:\n  - 7B model: 40GB+\n  - 70B model: 80GB (with TP=4)\n  - 405B model: 80GB (with TP=8, PP=8)\n- **Storage**: Fast NVMe for checkpoints (1TB+ for 70B+ models)\n\n## Resources\n\n- Docs: https:\u002F\u002Fdocs.nvidia.com\u002Fmegatron-core\u002F\n- GitHub: https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM\n- Papers:\n  - \"Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism\" (2019)\n  - \"Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM\" (2021)\n- NeMo Framework: https:\u002F\u002Fdocs.nvidia.com\u002Fnemo-framework\u002F (built on Megatron-Core)\n\n\n",{"name":5,"tags":24,"author":34,"license":15,"version":35,"description":7,"dependencies":36},[25,26,27,28,29,30,31,32,33],"Megatron-Core","Large-Scale Training","NVIDIA","Tensor Parallelism","Pipeline Parallelism","Model Parallelism","H100","Distributed Training","Production","Orchestra Research","1.0.0",[37,38,39,40],"megatron-core","torch","apex","transformer-engine","skills","4fb202b97f2676132ac8713dbd2370cefacacb81",".agents\u002Fskills\u002Fmegatron-core","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills","https:\u002F\u002Fgithub.com\u002Fzengrong233","https:\u002F\u002Fwtu.edu.cn\u002F","my name songzengrong\r\n  research interest's:NLP","2026-06-04 10:48:22",{"source":8,"skillName":5,"commands":50,"prompt":60,"zipAvailable":61},[51,54,57],{"tool":52,"command":53},"npx","npx skills add zengrong233\u002Fdevin-skills --skill training-llms-megatron",{"tool":55,"command":56},"bunx","bunx skills add zengrong233\u002Fdevin-skills --skill training-llms-megatron",{"tool":58,"command":59},"pnpm","pnpm dlx skills add zengrong233\u002Fdevin-skills --skill training-llms-megatron","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002FSKILL.md",true,[63,69,73,77,81],{"path":64,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":67,"sourceHtmlUrl":68},".openskills.json","REFERENCE","参考文档",243,"https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002F.openskills.json",{"path":70,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":71,"sourceHtmlUrl":72},"references\u002Fbenchmarks.md",7363,"https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002Freferences\u002Fbenchmarks.md",{"path":74,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":75,"sourceHtmlUrl":76},"references\u002Fparallelism-guide.md",9591,"https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002Freferences\u002Fparallelism-guide.md",{"path":78,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fproduction-examples.md",10903,"https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002Freferences\u002Fproduction-examples.md",{"path":82,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Ftraining-recipes.md",11588,"https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fmegatron-core\u002Freferences\u002Ftraining-recipes.md",[86,93,100,109,117,124],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M759188DV0S9ARY1ETHCEY45","whisper","OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual ASR.","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fwhisper\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M75913ZB4EDVWYGYRTRWTKD1","weights-and-biases","Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fweights-and-biases\u002FSKILL.md",[],[],{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":104,"licenseFileUrl":105,"sourceHtmlUrl":106,"categories":107,"keywords":108,"publishedAt":19},"01M75910YNW7H0KZQDR96TGWZ9","webapp-testing","Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fwebapp-testing\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fwebapp-testing\u002FSKILL.md",[],[],{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":104,"licenseFileUrl":113,"sourceHtmlUrl":114,"categories":115,"keywords":116,"publishedAt":19},"01M7590X6XJR78KVSHCVFYB59A","web-artifacts-builder","Suite of tools for creating elaborate, multi-component claude.ai HTML artifacts using modern frontend web technologies (React, Tailwind CSS, shadcn\u002Fui). Use for complex artifacts requiring state management, routing, or shadcn\u002Fui components - not for simple single-file HTML\u002FJSX artifacts.","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fweb-artifacts-builder\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fweb-artifacts-builder\u002FSKILL.md",[],[],{"skillNo":118,"name":119,"nameZh":6,"description":120,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":121,"categories":122,"keywords":123,"publishedAt":19},"01M7590RENMJJTXV44P8G553ZQ","serving-llms-vllm","Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency\u002Fthroughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ\u002FAWQ\u002FFP8), and tensor parallelism.","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fvllm\u002FSKILL.md",[],[],{"skillNo":125,"name":126,"nameZh":6,"description":127,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":128,"categories":129,"keywords":130,"publishedAt":19},"01M7590GA6EZKJA8GAAMMPAXBB","verl-rl-training","Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.","https:\u002F\u002Fgithub.com\u002Fzengrong233\u002Fdevin-skills\u002Fblob\u002Fskills\u002F.agents\u002Fskills\u002Fverl\u002FSKILL.md",[],[]]