NeMo MBridge Recipe Recommender
Recommends adjustable Megatron Bridge recipes for your model, GPU budget, and training goal.
The skill distinguishes production training from mock-data performance benchmarking, requests model/GPU/goal/context inputs, and requires no credentials or secrets; however, its commands can trigger costly external training effects and it lacks explicit execution confirmation, permission boundaries, rollback guidance, and configuration-safety validation, so points are deducted.
The documentation provides indexes, entry points, resizing rules, and common-failure warnings with generally consistent paths; however, this review executed nothing, no skill-specific test suite is supplied, and several compatibility and scaling claims require environment validation, so the static cap applies and points are deducted.
Audience, inputs, training modes, GPU ranges, and long-context scenarios are reasonably clear, covering PEFT, VLM, MoE, and diffusion; non-fit boundaries, Chinese-language support, and software-version variance are not sufficiently specified, so points are deducted.
The SKILL.md is readable and includes metadata, entry points, examples, limitations, and versioned recipe indexes; however, the BENCHMARK report identifies missing recommended Instructions and Examples sections and missing author metadata, while maintenance ownership, changelog, and update path remain incomplete, so points are deducted.
The skill can directly provide recipe functions, launch commands, datasets, and parallelism guidance; the BENCHMARK contains only one positive task and was not executed during this review, so recommendations still require hardware, version, and data validation and the score remains below the static ceiling.
A pinned revision, evaluation report, one positive evaluation task, and explicit ground truth provide some auditable support; however, there is no committed skill-specific test suite or broader independent reproduction evidence, and this review did not execute the commands, so coverage is limited.
- Do not treat performance recipes as production training configurations; they use mock data and their throughput targets are not convergence-validated.
- Before execution, verify the Megatron Bridge version, actual model configuration, GPU topology, num_key_value_heads, CP communication type, and dataset; the documented scaling advice is not a runtime guarantee.
- Add explicit execution confirmation, failure diagnostics, a version-compatibility matrix, maintenance contact, and broader scenario evaluation.
What does this skill do, and when should you use it?
This skill helps users select NeMo MBridge recipes for pretraining, supervised fine-tuning, or parameter-efficient fine-tuning. It matches model, GPU count and type, training goal, and sequence length against an indexed recipe library, then provides the recipe function and launcher command. It also gives concrete resizing guidance for tensor, pipeline, context, expert, sequence, and data parallelism, plus batch-size and failure-mode advice. It distinguishes functional library recipes from performance recipes intended only for mock-data throughput benchmarking.
Asks for the model name or size, GPU count and type, training goal, and any non-default sequence length; looks up a matching library or performance recipe; returns a recipe function name and entry-point command; recommends adjustments to TP, PP, CP, EP, SP, implicit DP, and batch sizes; and flags issues involving KV-head divisibility, long context, MoE, CUDA graphs, FSDP, and multimodal data.
- An engineer performing a first Megatron Bridge smoke test with mock data.
- A training engineer selecting a pretraining recipe for a listed Llama, Qwen, DeepSeek, Nemotron, or other supported model family.
- A user fitting SFT or PEFT to a constrained GPU allocation and needing parallelism changes.
- A team planning 16K–128K context training and needing a context-parallel recipe.
- A benchmarking team measuring upper-bound throughput on H100, B-series, or GB-series GPUs with mock data.
What are this skill's strengths and limitations?
- Covers pretraining, SFT, and PEFT across many model families, VLMs, and diffusion models listed in the index.
- Clearly separates functional training recipes from mock-data throughput benchmarks.
- Provides concrete rules for TP, PP, CP, EP, SP, batch sizing, and OOM handling.
- Includes copyable commands for library training and performance benchmarking.
- Recommendations are limited to recipes listed in the supplied SKILL.md; unmatched models require using a nearby size and adjusting configuration.
- Performance recipes use mock data and 50 iterations by default and should not be treated as production convergence configurations.
- Execution depends on a Python, uv, PyTorch, and Megatron Bridge environment; complete environment setup is not documented here.
- Interconnect, model architecture, and parallelism combinations can create compatibility or performance issues that require validation.
How do you install this skill?
Use the CLI flow documented in the repository README: npx skills add nvidia/skills --skill nemo-mbridge-recipe-recommender --yes. The README says you do not need to clone the repository or copy the skill directory manually; it becomes available when the agent next loads relevant skills.
How do you use this skill?
Tell the agent the model name or size, GPU count and type, training goal (pretrain, SFT, or PEFT), and any non-default sequence length. Example: “Recommend an SFT starting recipe for Llama 3 8B on 8 H100 GPUs.” The agent should distinguish library and performance recipes, then provide the recipe name, dataset choice, launcher command, and resizing advice.
How does this skill compare with similar options?
The skill presents functional library recipes and performance recipes as the main alternatives: library recipes use scripts/training/run_recipe.py for training, while performance recipes use scripts/performance/run_script.py for mock-data throughput benchmarks.