NVIDIA-NeMo/Automodel

nemo-automodel-recipe-development

Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow. Use when creating or modifying training, SFT, or eval recipes, adding new YAML config fields, debugging recipe construction or trainer…

Langue du texte source : anglais

build-and-dependency

NeMo AutoModel dev environment: containers, uv dependency management, env vars, and build pitfalls. Use when uv sync fails, imports break, TransformerEngine mismatches, or .venv is stale.

Langue du texte source : anglais

cicd

NeMo AutoModel CI: pipeline structure, commits and PRs, triggering CI with /ok to test, and investigating failures. Use when CI is red or didn't run, or when writing a commit or PR.

Langue du texte source : anglais

fern-docs

Edit or validate the NeMo AutoModel Fern docs (docs/*.mdx, docs/fern/): page edits, new, moved or removed pages, slugs, redirects, version cuts, fern check or preview failures.

Langue du texte source : anglais

linting-and-formatting

NeMo AutoModel code style: ruff config, naming, type hints, docstrings, copyright headers, and the review checklist. Use for style questions, ruff errors, or pre-commit failures.

Langue du texte source : anglais

parity-testing

Verify NeMo AutoModel models against HuggingFace references (state dicts, forward pass). Use for loss divergence, output mismatches, or checking state-dict mappings.

Langue du texte source : anglais

testing

NeMo AutoModel tests: unit and functional layout, L0/L1/L2 tiers, running, adding, or disabling tests, and flaky tests. Use when adding, running, or debugging tests or choosing a tier.

Langue du texte source : anglais

train-inference-consistency

Make a model's training forward match the rollout engine's logprobs bitwise (batch-invariant, aligned kernels) for on-policy RL, e.g. when vllm_kl or bitwise_frac is nonzero. Not HF parity.

Langue du texte source : anglais

nemo-automodel-distributed-training

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…

Langue du texte source : anglais

nemo-automodel-launcher-config

Configure NeMo AutoModel job launches for interactive runs, Slurm clusters, and SkyPilot cloud execution. Use when configuring Slurm or SkyPilot job submission, setting up multi-node launch scripts, debugging job submission failures, or switching between…

Langue du texte source : anglais

nemo-automodel-model-onboarding

Guide for onboarding new model architectures into NeMo AutoModel, including architecture discovery, implementation patterns, registration, and validation. Use when adding or modifying model architecture support in NeMo AutoModel, such as LLM/VLM/MoE model…

Langue du texte source : anglais

pr-review

Repository review rubric for the formal /review command.

Langue du texte source : anglais