nemo-mbridge-recipe-recommender

name	nemo-mbridge-recipe-recommender
license	Apache-2.0
description	Recommend and customize Megatron Bridge library and benchmark recipes for a user's model, GPU count, hardware, sequence length, and pretrain/SFT/PEFT goal. Use when selecting a starting recipe, comparing library and benchmark configs, resizing parallelism for a GPU allocation, or distinguishing convergence changes, semantics-preserving execution tuning, and benchmark-only shortcuts.

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. Benchmark recipes under src/megatron/bridge/perf_recipes/ are for upper-bound throughput benchmarks. They own their canonical benchmark data and settings and should not be presented as production training recipes.
3. For a first-time Bridge smoke test, recommend llama3_8b_pretrain_config with mock data via --dataset mock.
4. For normal SFT recommendations, select a finetuning preset such as --dataset squad or --dataset tulu3; for pretrain and mock validation recommendations, use --dataset mock. Do not pair the pretraining-only mock preset with an SFT or PEFT mode.
5. After the recipe and dataset, give the required resizing rules: TP must divide num_key_value_heads, use the lowest TP that still fits the dense state, keep TP within one node unless using NVL72-class interconnect, enable SP when TP > 1, configure CP for long context, and make both dense and expert meshes divide the allocation.
6. State whether each proposed override changes the convergence contract or only the execution/performance mapping. Do not trade convergence semantics for throughput without calling it a new experiment.

Configuration Layers and Change Control

Separate training semantics from their hardware mapping before recommending or tuning a recipe.

Convergence configuration includes the starting checkpoint and trainable parameters; dataset/revision/split/order/seeds; tokenizer, masking, truncation, and packing; sequence length; global batch and token budget; objective and loss coefficients; natural or forced MoE routing and token-dropping policy; optimizer, LR, schedule, warmup, betas, epsilon, weight decay, clipping, and dropout; arithmetic and optimizer-state precision; and PEFT adapter settings. Changing one of these creates a new convergence experiment.

Execution/performance configuration includes hardware count and topology; TP/PP/VP/CP/EP/ETP/DP/SP; recompute and offload; distributed optimizer/FSDP; communication overlap; fusions and attention backends; CUDA graphs and compilation; checkpoint I/O; and MoE transport through all-to-all, DeepEP, or HybridEP when the routing policy is unchanged. These settings should preserve the objective and effective updates, although floating-point reduction order can produce small numerical drift that still needs validation.

Treat micro batch size and gradient accumulation as execution fingerprints. Tune them only with fixed global batch size, global batch membership/order, normalization, optimizer boundaries, and token budget, and validate fresh loss sentinels for each layout. Packing, precision, forced MoE load balancing, token dropping/capacity, and router/auxiliary loss changes are never performance-only knobs.

Treat mock data, forced balancing, disabled correctness checks, and timing-only schedules as benchmark-only shortcuts. They may be appropriate in perf_recipes, but their losses and checkpoints are not convergence evidence.

For comparable model-verification recipes, choose a cohort-wide convergence contract before tuning performance. Keep the same bounded data selection, preprocessing, sequence length, global batch, optimizer/schedule, precision, seeds, routing policy, optimizer-step horizon, and processed-token checkpoints where the architectures permit. Record any necessary model-specific deviation and do not present that result as apples-to-apples convergence evidence. Absolute losses from different architectures or tokenizers are not directly rankable; compare stability and trend at equal token counts.

When a recipe's batch disagrees with the chosen convergence contract, modify and validate the library recipe separately. A declared bounded-verification protocol may explicitly apply the same LR, schedule, sequence, and data overrides across a cohort, but do not make one-off convergence changes merely to improve throughput. Conversely, first try TP/PP/CP/EP, recompute/offload, dispatcher transport, overlap, fusion, and CUDA graphs when optimizing fit or throughput.

---

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 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 squad

# 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 mock \
    'model.tensor_model_parallel_size=2' \
    'train.global_batch_size=64'

Benchmark recipes (throughput benchmarks)

./scripts/training/train.sh \
    --nodes 2 --gpus-per-node 8 \
    --account ACCOUNT --partition PARTITION --container-image IMAGE \
    --recipe qwen3_30b_a3b_pretrain_16gpu_h100_bf16_config \
    --mode pretrain

The total GPU allocation must match the count encoded in the recipe name. The user selects the node shape, and the selected partition must provide the requested hardware. The launcher does not inject benchmark offline defaults or cluster-specific launch policy. Use --env NAME for exported offline or NCCL fabric settings and repeated --srun-arg=ARG options for srun. Configure CPU/NUMA wrappers and Slurm segment sizing through the target cluster integration, or use scripts/performance/setup_experiment.py when its compatibility policies are required. The unified launcher supports exact exported text pretraining, text SFT/PEFT, Qwen-VL pretraining, and Wan pretraining recipes and infers their forward step. Text SFT/PEFT text benchmark recipes retain the flat runner's mock-data default; Qwen-VL and Wan retain their model-specific datasets. Exported benchmark PEFT recipes are fixed LoRA configs; use a configurable library recipe for DoRA. Trailing KEY=VALUE overrides are accepted, but an overridden benchmark recipe no longer represents its canonical benchmark configuration. Use scripts/performance/setup_experiment.py for selector-based invocation, dataset replacement, topology resizing, and specialized benchmark controls.

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

---

Benchmark Recipe Layout

Benchmark recipes use the same Python function format as library recipes, but live in a dedicated namespace for throughput benchmarking:

• Benchmark recipes live in src/megatron/bridge/perf_recipes/<family>/<hardware>/<model>.py
• Each benchmark recipe is a self-contained Python function (e.g. llama3_8b_pretrain_8gpu_h100_bf16_config())
• Recipe names encode model, task, GPU count, hardware, precision, and optional variant
• scripts/performance/utils/utils.py derives compatibility WorkloadBaseConfig views from the flat recipe itself
• 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.

The training launcher discovers library and benchmark recipes from the complete exported function name. Five legacy duplicate names select the benchmark definition; use the corresponding generic alias for those functional workloads. New recipe names should be unique across both packages.

---

Recipe Index (Library & Benchmark)

The full per-family recipe tables — every shipped library recipe (src/megatron/bridge/recipes/) and benchmark recipe (src/megatron/bridge/perf_recipes/), with parallelism degrees, minimum GPU counts, and hardware coverage — are kept in a dedicated reference file so this skill stays concise:

→ See references/recipe-index.md — Library Recipe Index (Llama, Qwen2/2.5/3, Qwen3-MoE, Qwen3-Next, DeepSeek, GLM-4.5, Gemma, Nemotron, VLM, Diffusion) and Benchmark Recipe Index (per-hardware throughput configs).

Load that file to pull an exact recipe function name or its default parallelism; the guidance below tells you which entry to look up.

---

Recommendation Decision Tree

User wants to train a model
│
├─ Know the model name?
│   ├─ Yes → Look up in references/recipe-index.md
│   │   ├─ 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 benchmark recipes (src/megatron/bridge/perf_recipes/)
│   │   ├─ Exact exported recipe → scripts/training/train.sh --recipe <exact function name>
│   │   └─ Selector/specialized workflow → scripts/performance/setup_experiment.py
│   └─ 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. Dense and expert meshes overlap. Do not multiply TP and EP together. The minimum MoE world size is PP × max(TP × CP, EP × ETP). Dense DP is world_size / (TP × PP × CP) and expert EDP is world_size / (PP × EP × ETP); both quotients must be integral, and the expert count must be divisible by EP.

Fit-First, Then Fill Tuning Loop

For an MoE recipe on a fixed allocation, use this practical order:

1. Start with the lowest legal TP that can hold the dense parameters, optimizer state, and workspaces. Do not raise TP only to reduce activation memory; TP communication is expensive, so try recompute first.
2. Use the largest legal EP that matches the expert count and topology. EP=8 is a common first candidate on one 8-GPU node, and EP=32 is a common first candidate on a 32-GPU allocation. These are starting points, not universal defaults: EP shards expert weights only, not dense layers or activations.
3. Prefer architecture-specific selective recompute. For Qwen3.5 GDN layers, verify gdn_norm_out on the exact Megatron Core revision instead of assuming core_attn covers the GDN peak. Use full-layer recompute only when the targeted boundaries still do not make the complete step fit.
4. After a stable baseline completes optimizer initialization and several steady steps, inspect W&B memory and confirm every rank's peak allocated and reserved memory in the runtime logs; rank 0 can miss the hot EP or PP rank. If there is safe headroom, test one change at a time: lower TP by one legal step or increase MBS while keeping GBS, data order, and optimizer boundaries fixed.
5. Keep the candidate only when it improves end-to-end tokens/s/GPU or step time with healthy loss and no skipped/NaN iterations. Leave headroom for checkpointing, evaluation, routing imbalance, and optional CUDA graphs.

On EP=8, a large expert footprint can remain capacity-limited enough to need full recompute. EP=32 usually reduces expert-weight pressure substantially, making selective recompute a better candidate, but Qwen3.5 GDN coverage and the hot rank's measured peak still decide.

Batch Size Tuning

• Start with the recipe's micro_batch_size. If OOM, reduce to 1; after the run fits, use measured headroom to sweep MBS upward again.
• 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

Pitfall	Symptom	Fix
TP > num_kv_heads	Crash: "TP must divide num_query_groups"	Reduce TP to a divisor of num_kv_heads
PP without VP	Poor throughput (large bubble)	Set virtual_pipeline_model_parallel_size
EP too low for large MoE	OOM on expert params	Increase EP; each expert lives on EP/num_experts ranks
CUDA graphs + packed sequences	Assert: "CUDA graph accepts only Tensor inputs"	Disable packing or use local full-iteration graphs
CUDA graphs + full recompute	Assert: "full recompute only with full iteration CUDA graph"	Disable recompute or switch to local impl
use_te_rng_tracker not set	Assert on provider init when CUDA graphs enabled	Set cfg.model.use_te_rng_tracker = True and cfg.rng.te_rng_tracker = True
FSDP + TP > 1 on H100	Possible comm bottleneck	Prefer FSDP with TP=1 or TP=2 on H100; FSDP shines on GB/B-series
Long context without CP	OOM on activations	Add CP=2/4/8; use *_16k, *_64k, or *_128k recipe variants
MoE overlap_grad_reduce on H100	May hurt throughput (False in many H100 presets)	Set overlap_grad_reduce=False for MoE on H100
VLM SFT missing image data	Runs but produces garbage	Provide actual multimodal dataset or use mock VLM data
Qwen35-VL MoE FSDP	Tested on Blackwell only	May not work on H100; validate first

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 mock

# Run the native 4-GPU Qwen3-MoE 30B PEFT topology
uv run python -m torch.distributed.run --nproc_per_node=4 scripts/training/run_recipe.py \
    --recipe qwen3_30b_a3b_peft_config \
    --dataset tulu3

# 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 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 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?

I want to...	Start with	GPUs needed
Try Bridge for the first time	llama3_8b_pretrain_config + mock data	2
Fine-tune a 7-8B model	llama3_8b_sft_config or qwen3_8b_sft_config	2–4
LoRA on 1 GPU	llama3_8b_peft_config or qwen3_8b_peft_config	1
Pretrain a dense 70B	llama3_70b_pretrain_config	32–64
Train a small MoE	qwen3_30b_a3b_pretrain_config	16
Train a large MoE (235B+)	qwen3_235b_a22b_pretrain_config	256–512
Benchmark text-pretrain throughput	Benchmark recipe via train.sh --recipe <exact name>	Exact encoded count
Long-context training	llama3_8b_128k_pretrain_config or add CP override	16+
VLM fine-tuning	qwen3_vl_8b_sft_config or gemma3_vl_*_sft_config	4–8
Diffusion training	wan_1_3B_pretrain_config or flux_12b_pretrain_config	8

---

Code Anchors

What	Path
Library recipes root	src/megatron/bridge/recipes/
Recipe __init__.py (all exports)	src/megatron/bridge/recipes/__init__.py
Common recipe helpers	src/megatron/bridge/recipes/common.py
Training entry point	scripts/training/run_recipe.py
Training Slurm launcher	scripts/training/train.sh
Benchmark recipes root	src/megatron/bridge/perf_recipes/
Benchmark compatibility launcher	scripts/performance/setup_experiment.py
Benchmark recipe helpers	scripts/performance/utils/utils.py
Benchmark overrides	scripts/performance/utils/overrides.py

Last signature refresh: 2026-08-03.