Nemo Mbridge Recipe Recommender

Auto Recipe — Recipe Index & Recommendation

This skill indexes every shipped recipe and helps users pick the right starting config, adjust parallelism, and avoid common pitfalls.

How to Use This Skill

1. Ask the user for: model name/size, GPU count & type, training goal (pretrain / SFT / PEFT), and sequence length (if non-default).
2. Look up the best-match recipe in the index below.
3. Recommend the recipe function name + entry-point command.
4. Provide adjustment advice (parallelism resizing, batch tuning, pitfalls).

First Answer Checklist

When recommending recipes, always include these distinctions before the long index details:

1. Library recipes under src/megatron/bridge/recipes/ are for functional training and use scripts/training/run_recipe.py.
2. Performance recipes under scripts/performance/ are for upper-bound throughput benchmarks. They use mock data and should not be presented as production training recipes.
3. For a first-time Bridge smoke test, recommend llama3_8b_sft_config with mock data via --dataset llm-pretrain-mock. Do not use llm-finetune for the setup-only tryout unless the user specifically asks for an SFT data path.
4. For normal SFT recommendations, use --dataset llm-finetune; for pretrain and mock validation recommendations, use --dataset llm-pretrain-mock.
5. After the recipe and dataset, give the required resizing rules: TP must divide num_key_value_heads, keep TP within one node unless using NVL72-class interconnect, enable SP when TP > 1, configure CP for long context, DP is implicit, and reduce micro_batch_size first on OOM.

Entry Points

Library recipes (functional training)

# Pretrain with mock data
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe <recipe_function_name> \
    --dataset llm-pretrain-mock

# SFT with SQuAD
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe <recipe_function_name> \
    --dataset llm-finetune

# Override any field via CLI
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe llama3_8b_pretrain_config \
    --dataset llm-pretrain-mock \
    'model.tensor_model_parallel_size=2' \
    'training.global_batch_size=64'

Performance recipes (throughput benchmarks)

python scripts/performance/run_script.py \
    --recipe <model_family> \
    --gpu_type h100 \
    --num_gpus 64 \
    --data mock

See the Performance Recipe Index for important caveats before using these for anything beyond throughput benchmarking.

Recipe Unification (Coming Soon — PR #2803)

PR #2803 is unifying performance recipes into the same Python function format used by library recipes. Key changes:

• Perf recipes move from scripts/performance/configs/ → src/megatron/bridge/recipes/<family>/<model>_perf.py
• Each perf recipe becomes a self-contained Python function (e.g. llama3_8b_h100_bf16_pretrain_config())
• The old WorkloadBaseConfig → set_workload_base_configs → get_perf_optimized_recipe pipeline is removed
• Shared helpers: _benchmark_common() (50 iters, timing, TE RNG), _perf_precision() (bf16 / fp8_cs / fp8_mx / nvfp4)

Why Python, not YAML? Previous YAML-based approaches had problems: recipe logic was split across multiple indirection layers, configs were not self-contained, and the two-level pipeline made maintenance and debugging difficult. Python functions are explicit, greppable, and composable.

After #2803 lands, both library and perf recipes will be invocable through the same run_recipe.py entry point.

Library Recipe Index

All recipes live under src/megatron/bridge/recipes/. Each function returns a ConfigContainer with model, training, optimizer, and data settings.

Llama

Qwen2 / Qwen2.5

Qwen3 (Dense)

Qwen3 MoE

Qwen3-Next

DeepSeek

GLM-4.5

Gemma

NemotronH / Nemotron

Other Models

VLM Recipes

Diffusion Recipes

Performance Recipe Index

All perf recipes live under scripts/performance/. They are invoked via run_script.py and use WorkloadBaseConfig presets per GPU type.

Important: Perf recipes are designed for upper-bound throughput benchmarks, not production training. They run 50 iterations on mock data by default. Throughput numbers are aspirational targets, not validated convergence configs.

Llama 3 / 3.1

SFT/LoRA variants also exist (e.g. 8B SFT with packed sequences, 70B SFT on 32 GPUs).

DeepSeek V3

Qwen3 MoE

Qwen3-VL

Kimi K2

NemotronH

GPT-OSS

Recommendation Decision Tree

User wants to train a model
│
├─ Know the model name?
│   ├─ Yes → Look up in Library Recipe Index above
│   │   ├─ Has a recipe for their size + mode? → Use it directly
│   │   └─ No exact match? → Use closest size, adjust parallelism
│   └─ No → Ask for model name, size, and HF model ID
│
├─ What's the training goal?
│   ├─ Pretrain → Use *_pretrain_config
│   ├─ SFT (full fine-tune) → Use *_sft_config
│   └─ PEFT (LoRA/DoRA) → Use *_peft_config (lowest GPU requirement)
│
├─ How many GPUs?
│   ├─ 1 GPU → Only PEFT recipes work (TP=1, PP=1)
│   ├─ 8 GPUs (1 node) → Most 8B–16B models, small MoE (EP=8)
│   ├─ 16–64 GPUs → 70B dense, medium MoE
│   └─ 128+ GPUs → 405B+, large MoE (DeepSeek V3, Kimi K2)
│
├─ Want throughput benchmarks?
│   ├─ Yes → Use perf recipes (scripts/performance/)
│   │   └─ ⚠️ These run on mock data for upper-bound perf only
│   └─ No → Use library recipes (scripts/training/run_recipe.py)
│
└─ Long context?
    ├─ > 8K → Need CP (context parallelism), check *_16k / *_64k / *_128k variants
    └─ ≤ 8K → Default recipes work

Adjustment Advice (When Recommending)

Parallelism Resizing Rules

When the user's GPU count differs from the recipe default:

1. TP must divide num_key_value_heads (GQA constraint). E.g. if num_key_value_heads=8, valid TP = {1, 2, 4, 8}.
2. TP should stay within a single node (NVLink). TP > 8 requires inter-node NVLink (e.g., GB200 NVL72).
3. PP adds pipeline bubbles. Minimize PP; only increase when TP alone can't fit the model. Use VP (virtual pipeline) to mitigate bubble overhead.
4. EP doesn't reduce dense-layer memory. Only expert parameters shard with EP. Shared attention/embeddings are replicated. For "OOM with MoE", increase EP first, not TP.
5. SP should be True whenever TP > 1. It eliminates redundant activation copies and is essentially free.
6. CP requires all-to-all or ring attention. Check cp_comm_type. For GQA models, a2a+p2p hierarchical CP allows CP > num_kv_heads.
7. world_size = DP × TP × PP × CP × EP. DP is implicit. Make sure the product of explicit parallelisms divides your total GPU count.

Batch Size Tuning

• Start with the recipe's micro_batch_size. If OOM, reduce to 1.
• global_batch_size determines learning dynamics. Scale with DP: GBS = micro_batch_size × DP × gradient_accumulation_steps.
• For MoE, micro_batch_size=1 is typical at scale.

Common Pitfalls to Warn About

Recipe Override Examples

# Scale Llama3 8B from 2 GPUs to 8 GPUs (increase DP)
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe llama3_8b_pretrain_config \
    --dataset llm-pretrain-mock

# Reduce parallelism for Qwen3-MoE 30B to fit on 4 GPUs
uv run python -m torch.distributed.run --nproc_per_node=4 scripts/training/run_recipe.py \
    --recipe qwen3_30b_a3b_sft_config \
    --dataset llm-finetune \
    'model.expert_model_parallel_size=4'

# Add long context to an existing recipe
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe llama3_8b_pretrain_config \
    --dataset llm-pretrain-mock \
    'model.seq_length=32768' \
    'model.context_parallel_size=4'

# Enable CUDA graphs on any recipe
uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
    --recipe qwen3_30b_a3b_pretrain_config \
    --dataset llm-pretrain-mock \
    'model.cuda_graph_impl=transformer_engine' \
    'model.cuda_graph_scope=[attn,moe_router,moe_preprocess]' \
    'model.use_te_rng_tracker=True' \
    'rng.te_rng_tracker=True'

Quick Reference: Which Recipe for My Situation?

Code Anchors