NVIDIA-NeMo

Capture a headless Playwright screenshot of a Studio page or a single element by data-testid. Use whenever you need an image of the web UI — attaching to a PR, verifying a visual change, comparing light/dark themes, or showing a bug.

Declares HTTP authorization on NeMo Helix plugin routes with @path_rule, AuthzScope, PermissionSet, and CallerKind. Use when adding or securing a plugin route, minting permission ids, choosing PRINCIPAL vs SERVICE_PRINCIPAL callers, fixing a hard_fail bundle…

Fine-tune or train an open model (Nemotron, Llama, Qwen and others) on your own data with NeMo Helix: SFT, LoRA, DPO preference tuning and GRPO reinforcement learning. Picks the backend, prepares the dataset, submits and monitors the job, and reports results.…

Build, convert, validate and upload a NeMo Gym reward environment for GRPO training: package your own reward code or a NeMo Gym server, convert a Prime Intellect or verifiers environment, and write the rollout-row dataset. Hand off to nemo-customizer to…

End-to-end NeMo Helix recipe for domain embedding and reranking fine-tuning: Data Designer retrieval-generate/prepare, Automodel bi_encoder/cross_encoder, Evaluator retrieve-eval on frozen eval_beir (NIM providers via Deployment Manager). Use when the user…

Creates a new NeMo plugin from scratch. Use when starting plugin development, setting up a plugin package, registering surfaces via entry points, or asking how plugins are discovered by the platform. Trigger keywords: create plugin, new plugin, plugin setup,…

Creates schedulable NemoJob surfaces for NeMo Helix plugins. Use when adding a job, declaring spec_schema / input_spec_schema / to_spec / compile, mounting job routes with add_job_routes, understanding generated job CLI shape, or running jobs in containers.…

Evaluate models, datasets, and agents with the NeMo Helix Evals plugin. Use for metric selection, SDK checks, platform jobs, and result retrieval.

End-to-end reference for inference on the NeMo platform — registering LLM backends as ModelProviders, wiring them to VirtualModels with Switchyard middleware (random routing, translate) and `nemo-guardrails` content-safety rails, and hitting them via the nemo…

Author and validate Platform-managed NeMo Agents agent.yaml files using the nemo-agents-spec-v1 format. Use when the user wants to create, edit, validate, or adapt an agent.yaml file, choose a supported harness, add instructions, skills, MCP servers, tools,…

Use when a customer asks NeMo Platform to build, implement, package, register, or deploy a new agent from an idea or approved Ethos. Builds the supported LangChain Deep Agents shape through Fabric. Do not use for design-only work, focused agent.yaml editing,…

Captures a NeMo Platform agent Ethos as a durable artifact at agents/<name>-ethos/ETHOS.md. Validates the front matter and required markdown sections, writes the file, and uploads it to a NeMo Filesets fileset (the canonical copy). The Ethos location is fully…

NeMo Evaluator SDK-first rubric-to-eval guide for BYOB (Bring Your Own Benchmark): parse domain expert rubrics, choose composable evaluation primitives, generate human-reviewable eval configs/artifacts, and run reproducible exact, numeric, LLM-as-judge,…

End-to-end guide for getting evaluation data into NeMo Platform Intake so it shows up in the Experiments leaderboard. Create an Experiment, create an Evaluation, then log traces and evaluator results via ATIF (Harbor), chat-completions, or OTLP and view the…

Captures what a NeMo Platform agent should do before any code or YAML. Explores the user's codebase and docs first, then asks one intent question at a time for what source cannot supply. Output feeds nemo-ethos. Use over generic brainstorming for any NeMo…

NeMo guardrails CLI reference for creating guardrail configs and applying them to chat completions through an IGW VirtualModel. Covers config CRUD, the standalone `/checks` endpoint for verifying a config, and the `nemo-guardrails` inference middleware plugin…

Guide for adding support for new LLM or VLM models in Megatron-Bridge. Covers bridge, builder-backed model config, recipe, tests, docs, and examples.

Create or update concise, agent-readable Megatron Bridge model verification cards. Use when adding a model support card, auditing cross-model convergence comparability or verification coverage, recording conversion, deterministic inference, training,…

Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute. Use for activation memory OOMs or regressions involving recompute_granularity, recompute_num_layers, recompute_modules,…

Practical guidance for training MoE VLMs in Megatron Bridge. Compares FSDP and 3D-parallel approaches, using rounded lessons from Qwen3-VL, Qwen3-Next, and other multimodal experiments.

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…

Structured framework for exact HF↔MCore weight verification, forward-pass logit correlation, and optional strict numerical diagnostics. Use when debugging weight mismatches, verifying HF↔MCore checkpoint round-trips, choosing verification tools, or…

Run Megatron-LM (MLM) and Megatron Bridge training with mock or real data. Covers correlation testing, available recipes, and multi-GPU examples.

Operational guide for choosing and combining parallelism strategies in Megatron Bridge, including sizing rules, hardware topology mapping, and combined parallelism configuration.

Run NOOA Connect headlessly as an agent to discover interfaces, propose model settings, check chat or native decision routes and media, and save registry aliases with the user's permission. Canonical guide for Connect stages, JSON reports, bounded checks,…

Author agents with NVIDIA-labs Object Oriented Agents (NOOA). Use when writing or modifying an Agent subclass, agentic methods (ellipsis bodies), docstring prompts, structured output contracts, strategy selection (CodeAct/Predict), visibility control,…

Attach images, audio, video, and PDFs to NOOA agents. Use typed parameters or CodeAct show(), configure media_inputs, and troubleshoot text-only notices, URLs, MIME types, and size limits.

Understand and troubleshoot NOOA registry settings, aliases, credentials, reply budgets, reasoning levels, caching and reasoning replay. For permission and running Connect headlessly to onboard or edit models, use the nooa-connect skill.

Give a NOOA agent capabilities — methods as tools, built-in tools (ShellTools, TodoManager), MCP servers, agent skills (Skill/TextSkill/SkillRegistry), and multimodal media. Use when adding tools or external integrations to an agent, wiring MCP, or packaging…

Advanced tuning of NOOA strategies — CodeAct prefill, loop guards, text-only recovery callbacks, truncation, code restrictions, execution internals, and PredictStrategy tuning. Use when configuring CodeActConfig or PredictConfig beyond defaults, handling…

Solve an interactive grid game iteratively — observe, hypothesize, experiment, record knowledge, and submit action sequences until every level is won.

Explore and debug agent execution traces. Use when the user asks to analyze a trace, debug an agent run, investigate errors in a trace, or when they paste a trace-explorer prompt from the viewer UI.

Reference and routing skill for NVIDIA Nemotron 3.5 Transcribe. Use for model facts, checkpoint selection, inference constraints, and locating the canonical NeMo Speech main-stage training, fine-tuning, DFlash2, timestamp-serving, and quantization workflows;…

Review Nemotron pull requests for code correctness, instructions that match the implementation, first-user setup, and reproducible notebooks and cookbooks.

Prepare, validate, build, and use Nemotron Customizer airgap image bundles for offline clusters. Use when planning airgapped deployments, editing deploy/nemotron-customizer/airgap/airgap.yaml, selecting workflow targets, grouping step execution images, baking…

Configure and chain repo-native Nemotron steps for curation, tokenizer extension, SDG/BYOB, training, conversion, optimization, and evaluation.

Use for Nemotron embed/rerank recipes, including Nemotron 3.5 multimodal SDG, tuning, evaluation, and deployment. Not for generic retrieval.

Reference desk for NVIDIA Nemotron 3.5 Lightning (30B-A3B), a hybrid Mamba-Transformer MoE with MTP built from Nano 3 Base by distillation from Super 3 Base, single-stage text SFT, RLVR, MTP boosting, and W4A16 NVFP4 quantization. Use when the user asks facts…

Reference desk for NVIDIA Nemotron 3.5 Super (120B-A12B), covering vision-language architecture, MTP, distillation pretraining from Ultra 3 Base, Indic CPT, SFT, RLVR, specialist teachers, two-phase MOPD, MTP boosting, NVFP4 quantization, and inference. Use…

Run the Nemotron-3.5 Super VL finance LoRA cookbook (NeMo Megatron-Bridge) end-to-end for the user, on one 8-GPU node or a Slurm cluster: data prep, checkpoint conversion, text+vision LoRA fine-tuning of the 120B-A12B model, merge to a Hugging Face…

Guide for adding or updating an agent harness in NeMo-Gym: a CLI or SDK agent runtime (for example Hermes, Pi, Codex, OpenCode, or an agent framework) wrapped as an Agent Server adapter that runs through Gym's Environment Server and agent sessions. Use when…

Guide for adding a new benchmark or training environment to NeMo-Gym. Use when the user asks to add, create, or integrate a benchmark, evaluation, training environment, or resources server into NeMo-Gym. Also use when wrapping an existing 3rd-party benchmark…

Onboard a model's serving config for NeMo Gym evaluations: write the vLLM policy service (model path, served name, tensor/pipeline/data/expert parallelism, max model length, max sequences, tool-call and reasoning parsers, chat template, generation config,…

Take a NeMo Gym evaluation from a ready config to a trusted result: submit it, let the resumable Slurm chain continue through walltime limits, preemption and node loss, inspect and recover the run from its saved evidence with `gym eval resubmit --recover`,…

Use when creating, validating, or documenting Nemo Gym pivot datasets from rollout, trajectory, chat-completion, Responses API, or tool-call artifacts. Covers Gym Responses-style row conversion, reconstructing model calls from flattened rollout output,…

Use when debugging a Nemo Gym run or reward profiling job. Covers rollout collection failures, empty or partial JSONL outputs, stale materialized inputs, verifier/schema errors, Ray or Slurm issues, vLLM readiness, judge failures, tool/sandbox failures, cache…

Use to help users get started with Nemo Gym reward profiling. Covers the basic gym env start, gym eval run, and gym eval profile workflow, repeated rollouts, materialized inputs, rollout JSONL artifacts, task and rollout identity, output inspection, partial…

NeMo Gym pull-request review rubric used by the formal /review service, including automatic reviews.

Testing conventions for NeMo-RL. Covers what not to unit-test, Ray actor coverage pragmas, nightly test requirements, and recipe naming rules.

Contribution conventions for NeMo-RL. Covers the rules every code change follows, PR title format, commit sign-off, and CI triggering.

CI/CD reference for NeMo-RL. Covers GitHub Actions pipeline structure, CI triggering via /ok to test, and CI failure investigation.

Agent-team-based parallel code review for NVIDIA-NeMo/RL pull requests. Spawns specialized agents (RL expert, submodule experts, bug finder, design reviewer, test agent, devil's advocate, comment reviewer) that coordinate via shared task list and direct…

After a published PR review, check the author's fixes, scan the new commits for major issues, close fixed threads, stage follow-ups

Build and dependency management for NeMo-RL. Covers Docker image building and running, uv usage, venv setup, and adding dependencies.

Manually cherry-pick a merged NVIDIA-NeMo/RL PR from a release branch (r0.X.0) back onto main when the auto cherry-pick bot failed. Stops at conflicts for the user to resolve, then verifies, pushes, opens a cp PR, copies CI labels and posts /ok to test.

Manually cherry-pick a merged NVIDIA-NeMo/RL PR onto the current release branch (r0.X.0) and open a cp PR with matching CI labels and /ok to test.

Audit evaluation coverage, find gaps, and recommend improvements against ETHOS.md. Use for "audit my existing evals", "generate audit.md from ETHOS.md", schema validation, measuring an ATIF trace, or aggregating coverage reports. Full audits first read…

Use for "help me with my evals", "help me onboard to expand my evals", "what's the state of the eval suite here?", "what happened in this trace?", or requests to create, discover, audit, or improve evaluations. For fresh full workflows, read this skill, show…

For a fresh start, read the eval-author core skill at ../eval-author/SKILL.md, use its pure conversation renderer for orientation and end the turn before repository inspection or prerequisite work. Confirm ETHOS.md, plan evaluation tasks, and build first…

Required by Eval Author's first-eval and task-create flows before they write a Harbor or NeMo Gym task's tests, and available to trace-environment: builds the environment the task runs in, meaning the agent's databases, files, services, and tools, seeded…

Experimental compatibility workflow for an explicit trace-environment request or continuation of an existing trace-environment workspace. New requests to derive tasks from recorded interactions go to task-create ATIF source mode or first-eval. Supports NeMo…

Captures what an AI agent is supposed to do — its purpose, boundaries, and success criteria — as a durable snapshot for anyone building, evaluating, reviewing, or changing it. Explores the user's codebase and docs first, then asks one intent question at a…

Writes an AI agent's ETHOS.md — a durable snapshot of its purpose, intended behavior, boundaries, and success criteria — from the answers gathered by ethos-explore. Renders the schema-v1 template, validates the front matter and required markdown sections,…

Entry point for an AI agent's ETHOS.md: a durable snapshot of what the agent is for, what it must not do, and what success looks like. Routes a request to explore and capture intent (ethos-explore), or write or edit the file (ethos-write). Owns the standard…

Build and verify an inference-aligned training forward for a NeMo AutoModel model so the trainer recomputes the rollout engine's per-token logprobs bitwise (true on-policy RL). Use for "bitwise", "batch-invariant", "aligned kernels", "train/inference…

Testing reference for NeMo AutoModel — unit and functional test layout, tier semantics (L0/L1/L2), running tests locally, adding or disabling tests, and pytest conventions. Use when adding, running, or disabling tests; debugging a test failure; choosing the…

Maintain the NeMo AutoModel Fern docs site under docs/ (MDX content) + docs/fern/ (infra) — add, update, move, or remove pages; manage redirects, slugs, navigation, and version aliases; run validation and previews. Use when editing or adding documentation…

Guide for selecting and configuring distributed training strategies in NeMo AutoModel, including FSDP2, Megatron FSDP, DDP, and parallelism settings. Use when adding or modifying distributed training strategies (FSDP2, HSDP, DDP), debugging multi-GPU or…

Dev environment setup for NeMo AutoModel — container-based development, uv package management, installation options, environment variables, and common build pitfalls. Use when setting up a dev environment, adding or removing dependencies, switching container…

CI/CD reference for NeMo AutoModel — pipeline structure, commit and PR workflow, CI failure investigation, and common failure patterns. Use when investigating a CI failure, understanding the pipeline structure, writing a commit or PR, triggering CI, 'CI is…

Code style and quality rules for NeMo AutoModel — ruff configuration, naming conventions, type hints, docstrings, copyright headers, and the code review checklist. Use when writing or reviewing code for style compliance, fixing ruff errors, understanding type…

Verify numerical parity between NeMo AutoModel implementations and reference HuggingFace models, including state dict and forward-pass checks. Use when verifying numerical correctness of a new or modified model against its HuggingFace reference, debugging…

Repository review rubric and review principles for the formal /review command.

Verify a NeMo Speech change or pull request before commit, push, or review by checking the complete diff, formatting, tests, documentation, and DCO sign-offs. Use when asked to verify changes, prepare a commit or PR, or review a PR's validation and coverage.

Adapt Nemotron Transcribe to customer audio with NeMo-RL GSPO. Use for reward-based speech fine-tuning, assessing rollout signal, selecting informative training recordings, choosing a bounded training run, or diagnosing RL that improves reward but not task…

Train a SALM speech LLM from a pretrained LLM and pretrained speech encoders with NeMo speechlm2 and the Automodel backend on a multi-node cluster. Use to set up, launch, monitor, continue, fork, or debug a large multitask training run (ASR, contextual ASR,…

Train or adapt a DFlash or DFlash2 speculative draft against a frozen SALM speech-LLM target with NeMo speechlm2. Use for online argmax or offline model-generated labels, pretrained draft adaptation, projection-only ablations, checkpoint resume/export, and…

Fine-tune a SALM speech-LLM checkpoint (speech encoder + projection + LLM decoder) with NeMo speechlm2 on a new domain or task, and evaluate it with vLLM. Use to adapt a speech-to-text LLM to new audio, improve WER on a domain corpus, train speech…

Quantize a SALM speech-LLM checkpoint with a NemotronH decoder to FP8 or NVFP4 for vLLM serving with the NeMo SpeechLM plugin, including the audio encoder and the MTP draft head, and verify the result. Use to produce a quantized checkpoint, check one before…

Fix a GitHub issue in NeMo Speech (NVIDIA-NeMo/Speech). Read the issue, reproduce the bug with a failing test, implement the fix, and verify tests pass. Only opens a PR if the user explicitly asks for it.

Commit current changes with a clear, descriptive message. The argument may contain special instructions for the commit message.

Create a GitHub PR with a well-formatted description matching the repository PR template (flat Changes by default; optional Added/Changed/Removed/Fixed grouping). The argument may contain special instructions for the pull-request summary.

Perform a thorough code review of the current branch or a GitHub PR by number. The argument may start with a pull-request number to review that PR, optionally followed by focus instructions such as "focus on the API changes"; with no number, the current…

Search local Fern documentation for content related to a topic. The argument is the topic to search for.

Search GitHub issues, discussions, and PRs for content related to a topic. The argument is the topic to search for.

Update an existing GitHub PR description to reflect current changes after incorporating feedback. The argument may include the "--title" flag to also update the pull-request title, plus special instructions for the summary.

Maintain the NeMo Data Designer Fern docs site under fern/. Use for any documentation change. Triggered by: "edit docs", "add doc page", "update docs", "rename page", "fix broken link", "add redirect", "preview docs", "publish docs", "regenerate notebooks",…

Review Data Designer changes for correctness, architecture, public contracts, and meaningful test gaps.

Triage or resume Switchyard's repository-wide community-maintenance queue, prepare a gardening handoff, or report a gardener's activity and impact. Do not use for one isolated issue or pull request, implementation work, or an ordinary development-status…

Deep dive on Harbor trial results for tasks that use SWE-Bench-style F2P (FAIL_TO_PASS) and P2P (PASS_TO_PASS) reference tests. Diagnoses why an agent failed and audits whether a failing task is genuinely hard or unfair (instruction-vs-verifier mismatch). Use…

Review CRAFT Search gold answer quality for a specific task. Use when the user asks to review, check, investigate, or look closely at a craft-* task ID. Evaluates gold files, functions, assertions, and explanation against actual source code and agent tier…

Review Switchyard Rust changes for correctness and maintainability. Use for pull requests or diffs touching crates, PyO3 bindings, async runtime behavior, streaming, protocol types, translation, algorithms, or LLM clients.

Build or publish nemo-switchyard Python distributions through the repository's GitHub Actions workflow. Use when asked to build a development wheel, cut a Python release, publish to PyPI, create a release tag, or debug .github/workflows/publish.yml.

Edit or debug the published Switchyard MkDocs site. Use for docs pages, mkdocs.yml navigation, mkdocs_hooks.py source links, strict build failures, local previews, or .github/workflows/docs.yml.

Select or debug Switchyard validation and GitHub CI. Use when asked which tests to run, whether a change is ready, why a CI job failed, how to reproduce CI, or whether live provider tests are appropriate.