docs(model): teach builder-backed model configs in the adding-model-support skill

[kamran-nvidia]

What does this PR do ?

Updates the adding-model-support skill to teach the builder-backed ModelConfig / ModelBuilder path instead of provider_bridge() / to_megatron_provider(). Documentation only; no code changes.

The skill still directs new bridges to override provider_bridge(). For a family that only does that, get_model_config() returns a config built from CONFIG_MAPPING alone, and every attribute set in provider_bridge() is silently dropped. Its YaRN example also described yarn_* as fields missing from the provider; since #6365 they are declared on Bridge TransformerConfig.

Changelog

• SKILL.md:
  • LLM bridges implement hf_config_to_model_config_kwargs(), make hf_config_to_provider_kwargs() return the same kwargs, and set USE_MODEL_CONFIG_FOR_CONVERSION = True.
  • Models that BridgeGPTModelConfig cannot represent (multimodal, hybrid) get a ModelConfig subclass and a ModelBuilder, following Muse Glimmer.
  • The hook table marks provider_bridge() as legacy.
  • Recipes use get_model_config(), and the smoke test builds through get_model().
  • The LLM reference moves from Qwen2 to Qwen3 (Qwen2 is not migrated yet).
• llm-patterns.md:
  • Model config pattern and a full bridge example.
  • Table of the three defaults that differ between the builder config and the legacy provider: position_embedding_type, masked_softmax_fusion and rope_scaling_factor. Missing one of these silently changes the model.
  • Layer-spec guidance, including that MCore's builder default handles dense-first MoE.
  • YaRN guidance that replaces the stale provider-subclass example.
  • A checklist for migrating an existing provider-based bridge.
• recipe-patterns.md: recipes use get_model_config(); the recipe-test pattern uses a strict builder config and rejects the provider API.
• tests-and-examples.md: builder unit-test template covering config mapping, provider/builder parity, and the conversion flag.
  • Also fixes the existing mapping-registry template, which raised TypeError because QKVMapping.hf_param is a dict.
  • Replaces the Mock() HF config, which fails on both paths because unset attributes such as hidden_act come back as Mock objects.
• vlm-patterns.md: adds the builder-backed multimodal pattern (Muse Glimmer) and labels the provider snippets as legacy.
• parity-testing/SKILL.md: config-mapping checks point to hf_config_to_model_config_kwargs() first.

Verification

The snippets were run against main (MCore 8949def81) in the NeMo 26.10 container, with Qwen3 standing in for the placeholder model:

• Snippets: the unit-test template, the recipe-test stub, and the bridge example from llm-patterns.md were extracted from the Markdown and executed. 16 tests passed.
• Defaults table: measured on a bridge with no overrides. Builder: learned_absolute, False, 8.0. Provider: rope, True, 1.0.
• Other claims checked:
  • Unknown kwargs raise ValueError.
  • Undeclared attribute assignment raises AttributeError.
  • Flat assignment routes to the nested transformer config.
  • AutoBridge.from_hf_config(...).get_model_config() works.
  • Qwen3 fills the HF YaRN defaults.
• Smoke test: the SKILL.md smoke snippet (get_model_config() → get_model() → export_hf_weights()) passed on one GB200 GPU with a tiny Qwen3 checkpoint.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

• Follows the provider → builder migration (refactor(model): add Qwen3 builder-backed model config with YaRN support #6365 merged; refactor(model): build Qwen2 and MiMo from builder-backed model configs #6379, refactor(model): build OLMoE and GLM-4.5 from builder-backed model configs #6383, refactor(model): build Hy V3 and MiniMax-M2 from builder-backed model configs #6392 and refactor(model): build Ling MoE V2 and EXAONE 4.0 from builder-backed model configs #6401 open).
• docs/adding-new-models.md still teaches provider_bridge(); left for a separate docs PR.

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[kamran-nvidia]

/ok to test a937f1e

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-6402-57f5bd386329.docs.buildwithfern.com/nemo/megatron-bridge

[kamran-nvidia]

/ok to test 57f5bd3