nemo-automodel-model-onboarding

Adding Model Support to NeMo AutoModel

Purpose

This skill guides implementation of new model architectures in NeMo AutoModel. Follow the five phases in order.

Instructions

When answering an onboarding question, keep the response in this order:

1. Classify the architecture from config.json.
2. Name the exact implementation files under components/models/<name>/.
3. Identify registry and optional custom-config updates.
4. State the validation tests that must be added before full checkpoint use.

For conceptual onboarding questions, answer from this skill without opening the pattern files unless the user asks you to edit code. Mention pattern filenames as references, then give the direct checklist.

Use direct action verbs: classify the model, name the files, map the weights, register the class, and add tests. Do not discuss distributed strategy, launcher configuration, or general recipe authoring unless the user explicitly connects it to onboarding a new architecture.

Examples

Use these compact answer patterns for common questions:

• Dense causal LM: classify as dense only when architectures contains a ForCausalLM class and expert fields such as num_local_experts, n_routed_experts, or num_experts_per_tok are absent. Create components/models/<name>/model.py and __init__.py; add state_dict_adapter.py only for checkpoint weight conversion and config.py only if needed. Register MODEL_ARCH_MAPPING in _transformers/registry.py, add example YAML, and add tiny-config unit tests plus layer-equivalence tests for rewritten layers.
• MoE state dict: identify expert fields in config.json, reference moe-patterns.md, map router tensors separately, preserve routed-expert index order, map routed experts, shared experts, and gate/up/down projections, add adapter key-map tests and tiny-config numerical equivalence tests, and do not rely only on from_pretrained() or silent tensor reshapes.
• VLM onboarding: classify as VLM only when vision_config, text_config, and a ForConditionalGeneration architecture are present. Reference vlm-patterns.md and existing VLM implementations such as mistral4, kimivl, or kimi_k25_vl; check text backbone, vision tower, projector, processor assumptions, text and vision checkpoint compatibility (adapter mappings when needed), registry registration, and tiny image-text tests before full checkpoints. Do not treat VLM onboarding as a pure causal-LM path or skip processor/image tests.

For MoE state-dict and VLM questions, apply the checklists in Sections 2.4 and 2.5.

Routing Boundary

Use this skill only when the user is adding or modifying model architecture support: model files, custom layers, state-dict adapters, Hugging Face config mapping, registry entries, or model capability flags.

Do not use this skill for standalone training recipe YAML questions about optimizers, datasets, schedulers, validation datasets, or trainer wiring unless they are explicitly part of onboarding a new model architecture. Those recipe questions belong to the nemo-automodel-recipe-development skill.

In-scope examples:

• "Add support for a new Hugging Face causal LM architecture."
• "Map MoE router and expert weights from a Hugging Face checkpoint."
• "Register a new model class in NeMo AutoModel."

Out-of-scope examples:

• "Write a finetuning recipe YAML with optimizer and dataset sections."
• "Choose FSDP2, DDP, tensor parallel, or context parallel settings."
• "Configure Slurm, SkyPilot, containers, mounts, or launch dispatch."

Phase 1: Discovery

Before writing code, gather information about the target model.

1.1 Fetch HuggingFace config.json

Download the model's config.json from the HuggingFace Hub (or use AutoConfig.from_pretrained). Key fields to extract:

• architectures -- determines the class name and registration key (e.g., "LlamaForCausalLM", "Qwen3MoeForCausalLM", "Mistral3ForConditionalGeneration")
• model_type -- used for custom config registration in _CUSTOM_CONFIG_REGISTRATIONS if HF does not have a built-in config class
• hidden_size, intermediate_size, num_hidden_layers, num_attention_heads, num_key_value_heads -- sizing
• vocab_size -- needed for tiny test configs
• tie_word_embeddings -- the saved setting in each supported checkpoint; do not infer it from a bare config constructor
• hidden_act -- activation function (e.g., "silu" for SwiGLU)

1.2 Determine model type

1.3 Check for existing similar architectures

Look in components/models/ for architectures with similar attention or MLP patterns:

components/models/
  llama/           # Standard GQA + SwiGLU with separate HF-compatible projections
  qwen2/           # Same as Llama but with attention bias + QKV bias
  baichuan/        # ALiBi attention variant
  deepseek_v3/     # MLA attention + MoE (DeepSeek-style grouped experts)
  mistral4/        # MLA + MoE + VLM (Pixtral vision)
  kimivl/          # DeepSeek-V3 backbone + MoonVit vision
  kimi_k25_vl/     # Updated KimiVL with different projector
  qwen3_moe/       # Qwen3 with MoE layers
  nemotron_v3/     # Hybrid mamba-attention

1.4 Identify custom components

Check whether the model needs:

• Custom attention: GQA (standard), MLA (DeepSeek/Mistral4), sliding window, bidirectional
• Custom RoPE: Standard (Llama), YaRN scaling, NTK-aware, complex-number (DeepSeek)
• Custom normalization: RMSNorm (standard), LayerNorm, different eps values
• Custom MLP: SwiGLU (standard), GeGLU, ReLU-squared, MoE routing
• Custom config class: Needed only if HF AutoConfig cannot parse the model's config.json (check auto_map field)

1.5 Note dimensions for test config

For unit tests, create a tiny config. Target: ~1M parameters or less.

Phase 2: Implementation

2.1 Create directory structure

2.2 Implementation order

Implement files in dependency order:

1. config.py (if needed) -- Custom PretrainedConfig subclass
2. rope_utils.py (if needed) -- RoPE implementation
3. layers.py (if needed) -- Attention, MLP, decoder block classes
4. model.py -- The main ForCausalLM (or ForConditionalGeneration) class
5. state_dict_adapter.py (if needed) -- HF weight conversion. Treat checkpoint I/O performance as part of the implementation, and evaluate the low-memory DCP capability as described in Section 2.6.
6. init.py -- Re-export the main model class

See the pattern files for detailed implementation guidance:

• Dense LLM: llm-patterns.md
• MoE: moe-patterns.md
• VLM: vlm-patterns.md
• Capabilities and fp32 precision: capabilities-and-precision.md

Most custom models need state_dict_adapter.py for HF weight conversion. Omit the file and attribute only when HF names and tensor layouts already match across supported backend/config variants, as in Llama, Qwen2, and Qwen3. Weight tying remains the model's responsibility (Section 2.3).

2.3 Causal LM weight tying

Every registered model class with a causal lm_head must:

• Declare tie_word_embeddings_support: TieSupport as BOTH, TIED_ONLY, or UNTIED_ONLY.
• Call reject_unsupported_tie_word_embeddings(type(self), config) at the top of __init__, using the original config before unwrapping text_config or thinker_config.

Only classes with no causal LM head may be explicitly exempted from the registry test.

Choose the policy from the implementation and the actual supported checkpoint configs, not from a bare config constructor:

• BOTH: tied and untied configurations are both supported.
• TIED_ONLY: only a tied configuration is supported.
• UNTIED_ONLY: only an untied configuration is supported.

Runtime helpers must treat TIED_ONLY and UNTIED_ONLY as authoritative and only resolve a per-checkpoint config flag for BOTH. All current BOTH VLMs honor the outer tie_word_embeddings flag, so do not add a model-specific resolver until a supported BOTH model actually requires another config path.

For BOTH and TIED_ONLY, always declare _tied_weights_keys and implement tie_weights() with the actual lm_head and input-embedding FQNs. Do not rely on inherited Hugging Face tying, and re-tie after any language-model swap.

Add policy-specific tests:

• BOTH: tied aliases; untied does not alias.
• TIED_ONLY: tied aliases; untied is rejected.
• UNTIED_ONLY: weights stay separate; tied is rejected.

Do not tie architectures with intentionally separate heads, asymmetric vocab sizes, or stages that do not own both tensors.

For from_pretrained, the checkpoint's saved tie_word_embeddings value is authoritative, even for BOTH. The NeMoAuto* bridge rejects flips in either direction. A model-owned from_pretrained that bypasses that bridge must call reject_tie_word_embeddings_flip(checkpoint_config, requested_config, model_class_name).

2.4 MoE state-dict adapter checklist

For MoE models, verify all weights below. When their HF and native layouts differ, the adapter must explicitly map:

• Router weights, including gate bias or correction-bias tensors when the Hugging Face model has them.
• Expert weights, preserving expert index order across local and routed experts.
• Gate/up/down projections, including combined or split projection layouts.
• Shared experts separately from routed experts when the architecture has both.

Add tests that assert expected key mappings and run numerical equivalence with tiny configs before trying full checkpoints.

Do not use these shortcuts:

• Do not validate the adapter only by calling from_pretrained().
• Do not accept missing or extra expert keys without an explicit mapping reason.
• Do not change dtype, transpose dimensions, or reshape tensors unless the HF and NeMo layouts require it and a test proves the conversion is reversible.
• Do not skip router or shared-expert tests because dense-layer tests pass.

2.5 VLM onboarding checklist

For VLMs, confirm the Hugging Face config has vision_config and text_config and that architectures points to a conditional-generation class. Start from the closest VLM pattern file, usually vlm-patterns.md, and compare existing implementations such as mistral4, kimivl, or kimi_k25_vl.

The implementation should explicitly cover:

• Text backbone, vision tower, projector, and processor or image preprocessing assumptions.
• Checkpoint compatibility for both text and vision modules, with adapter mappings where needed.
• Registration of the ForConditionalGeneration class in _transformers/registry.py.
• Tiny tests that exercise image-text inputs and verify checkpoint load/export, plus adapter round-trip when present.

2.6 Checkpoint I/O performance

Treat checkpoint performance as an implementation requirement, not a later optimization. For every new or materially changed state-dict adapter, evaluate both latency and peak host/device memory for loading and for any save or export path the change affects. In particular:

• Avoid full-checkpoint or model-sized temporary copies when tensors can load directly into final model storage or be transformed in bounded parts.
• Keep distributed reads and conversions rank-local when a rank needs only its shard; do not materialize a global tensor on every rank unnecessarily.
• Avoid repeated tensor merges, copies, full-heap garbage collections, shard scans, or file opens inside model-sized loops.
• Record representative before/after latency and peak-memory evidence for an optimized path, including the model, dtype, backend, and topology.

Every adapter must evaluate supports_low_memory_dcp_load. Set _supports_low_memory_dcp_load = True only when most checkpoint tensors write directly into final model storage and every remaining allocating conversion has a small, bounded temporary footprint for every runtime variant that reports support. Keep it false when a backend, topology, dtype, quantization mode, or model option requires model-sized rebuilding. A false value selects the safe fallback; it does not mean checkpoint loading is unsupported.

An opt-in needs focused tests that write sentinel values through direct destinations and prove the final model storage changes, bound any allocating conversions, and verify unsafe runtime variants report the capability as false. This storage test is also a correctness requirement: a false positive can cause the adapter to treat a temporary tensor as loaded in place and skip rebuilding the real parameter.

2.7 Register in registry

Add the model to MODEL_ARCH_MAPPING in _transformers/registry.py:

If the model has a custom config class with auto_map in its config.json, also register in _CUSTOM_CONFIG_REGISTRATIONS:

2.8 Declare capabilities and precision-sensitive params

Every class registered in MODEL_ARCH_MAPPING must declare parallelism capabilities, either with a static nested ModelCapabilities dataclass or a variant-aware get_capabilities(cls, config) method. Pick exactly one pattern. Capabilities should reflect recipe YAMLs that have been validated end to end.

If the model has precision-sensitive parameters such as Mamba A_log / dt_bias, MoE sigmoid gate bias, attention-sink bias, or per-head scale, declare _keep_in_fp32_modules_strict so sharding keeps those params in fp32 compute. See capabilities-and-precision.md for examples, variant dispatch rules, and frozen-submodule dtype guidance.

Phase 3: Onboarding Example Config

This phase is only for adding a minimal example config that proves the newly onboarded architecture can load and run. Use nemo-automodel-recipe-development for general recipe authoring or existing recipe modifications.

3.1 Create example YAML config

Create an example config under examples/llm_finetune/<name>/ (or examples/vlm_finetune/<name>/):

For new full-parameter Adam/AdamW examples, set model.dtype: float32. See training precision for compute precision and other training modes.

3.2 Verify model loads

Test that the model loads from a HuggingFace checkpoint:

3.3 Test with tiny config first

Before using full-size models, verify with a tiny config (1-2 layers, small hidden dim) to catch shape mismatches early.

Phase 4: Tests

Create tests/unit_tests/models/<name>/ and cover the checks below before loading full checkpoints:

• Forward-shape smoke test with a tiny config.
• State-dict adapter round-trip, when present: from_hf -> to_hf preserves mapped names, shapes, dtypes, and values.
• HF load/export and native save/reload preserve weights and ties, with or without an adapter; see tests/unit_tests/checkpoint/test_native_hf_state_dict.py.
• Layer equivalence tests for every rewritten attention, MLP, normalization, RoPE, or MoE layer. Use the model dtype from config, identical seeded weights, identical inputs, and dtype-appropriate torch.allclose tolerances.
• Short functional test that verifies loss decreases over a few training steps.

Phase 5: Documentation

5.1 Update model coverage page

Edit the appropriate file in docs/model-coverage/:

• LLM/MoE: docs/model-coverage/llm/index.md
• VLM: docs/model-coverage/vlm/index.md

Add a row with the model name, supported features (TP, PP, FSDP, LoRA, QLoRA), and any limitations.

Phase 6: Parity Testing

After implementation and unit tests are complete, run the full parity-testing workflow to verify that the new model produces numerically equivalent results to the reference HuggingFace implementation.

Run three levels of comparison:

1. State-dict round-trip: load a reference HuggingFace checkpoint, convert it into the NeMo AutoModel layout, export it back, and verify that all mapped tensors match the reference names, shapes, dtypes, and values within the expected tolerance.
2. Component-level parity: compare rewritten attention, MLP, normalization, RoPE, and MoE components against the HuggingFace implementation with fixed seeds and identical dtype.
3. End-to-end forward pass: run the full NeMo AutoModel and HuggingFace model on the same tokenized input and compare logits, hidden states, and loss.

Do not skip this phase. A model that passes unit tests can still diverge from HF due to subtle weight-conversion bugs, backend differences, or RoPE mismatches that only surface in a full parity comparison.

Key Files Reference

Checklist

• Fetched and analyzed config.json from HuggingFace
• Determined model type (dense LLM / MoE / VLM)
• Identified custom components (attention, RoPE, normalization, MLP)
• Created components/models/<name>/ directory
• Implemented config.py (if custom config needed)
• Implemented layers.py (if custom layers needed)
• Implemented rope_utils.py (if custom RoPE needed)
• Implemented model.py with HFCheckpointingMixin
• Implemented state_dict_adapter.py if needed; evaluated checkpoint latency and peak memory per Section 2.6
• For adapters, evaluated supports_low_memory_dcp_load; any opt-in proves direct destinations reach model storage, bounds allocating conversions, and reports False for unsafe variants
• Implemented init.py with re-export
• Registered in MODEL_ARCH_MAPPING in _transformers/registry.py
• Registered custom config in _CUSTOM_CONFIG_REGISTRATIONS (if applicable)
• Declared ModelCapabilities nested dataclass (static) OR get_capabilities(cls, config) classmethod (variant dispatch, e.g. ERNIE-4.5 MoE vs dense) — never both, never neither
• Declared TieSupport and called the constructor guard for every class with a causal lm_head (or added an explicit no-head exemption) -- see §2.3
• Added explicit _tied_weights_keys and tie_weights() for BOTH / TIED_ONLY, plus policy-specific alias and rejection tests -- see §2.3
• Guarded any model-owned from_pretrained that bypasses the NeMoAuto* bridge against checkpoint flips -- see §2.3
• Created example YAML config
• Verified model loads via NeMoAutoModelForCausalLM.from_pretrained()
• Created unit tests (forward shape, state_dict round-trip)
• Declared _keep_in_fp32_modules_strict for every intrinsically-fp32 param (SSM A_log/dt_bias, Mamba D when reference-fp32, MoE gate bias, attention-sink bias, scale, …) — see §2.8
• Created layer equivalence tests for every rewritten layer (matching model dtype)
• Created functional tests (training loss decreases)
• Updated docs/model-coverage page
• Ran state-dict round-trip, component parity, and E2E forward-pass parity checks
• Set ModelClass = <Name>ForCausalLM at module bottom