refactor(model): build Qwen2 and MiMo from builder-backed model configs

[kamran-nvidia]

What does this PR do ?

Moves Qwen2/Qwen2.5 and MiMo onto the builder-backed ModelConfig / ModelBuilder path, for recipes and checkpoint conversion, following Muse Glimmer and #6365 (Qwen3).

Builds on #6365 (merged). It reuses #6365's TransformerConfig YaRN fields, without which Qwen2.5 YaRN checkpoints could not convert through the builder. Rebased onto main after #6365's squash-merge, so the PR now contains only the Qwen2 and MiMo commits.

Changelog

• Qwen2Bridge: provider_bridge() is replaced by hf_config_to_model_config_kwargs(), and the provider path reuses the same kwargs (hf_config_to_provider_kwargs). Provider-implied defaults are explicit (masked_softmax_fusion, rope_scaling=False, rope_scaling_factor=1.0, position_embedding_type via setdefault("rope") so YaRN is preserved). USE_MODEL_CONFIG_FOR_CONVERSION = True.
• MimoBridge (subclasses Qwen2Bridge): extends the Qwen2 kwargs with qk_layernorm=False, add_qkv_bias=True, and MTP depth (mtp_num_layers, mtp_loss_scaling_factor=0.1) instead of overriding provider_bridge(). It inherits the conversion flag.
• Recipes: all 30 Qwen2/2.5 H100 recipes use get_model_config().
• Tests:
  • New test_qwen2_bridge.py; Qwen2 had no bridge unit tests. It covers the direct builder mapping, provider/builder parity (tied, fp16, YaRN), and the YaRN transformer fields plus HF export.
  • A strict recipe test runs every Qwen2/2.5 recipe against a real BridgeGPTModelConfig.
  • MiMo tests cover the builder MTP mapping and provider parity.

Verification

Each family was validated separately on tiny random-weight checkpoints (single GB200 node, NeMo 26.10 container).

Conversion paths:

• scripts/conversion GPU import at TP2 and export.
• AutoBridge.import_ckpt / export_ckpt.
• Load + forward as in compare.py.
• roundtrip_checkpoint.

Each was run on this branch and on main.

Check	Qwen2 (untied, tied, YaRN)	MiMo (MTP=1, MTP=0)
Provider fields, main vs this branch	identical	identical
get_model_config() vs provider	0 diffs (9 on main)	0 diffs (9 on main)
Imported run_config.yaml	BridgeGPTModelConfig + GPTModelBuilder	same, mtp_num_layers 1 / 0
Export vs source HF weights (GPU TP2, CPU, and from a main-imported checkpoint)	bit-exact, 9/9	bit-exact incl. MTP, 6/6
Logits: this branch vs main, and this branch on a main-imported checkpoint	bitwise-identical, 6/6	bitwise-identical, 4/4
Logits vs HF	cosine 0.99999	n/a (remote-code HF model)
roundtrip_checkpoint TP2	pass	pass

Unit tests:

• 268 (Qwen2 run) and 130 (MiMo run).
• Regression: 40 Qwen2/2.5 recipe factories and 517 tests across models/qwen, models/mimo, test_auto_bridge.py, recipes/qwen, test_qwen_recipes.py and doc_consistency.
• After rebasing onto main (Megatron Core 8949def81): 528 passed, 1 skipped across models/qwen, models/mimo, recipes/qwen, test_qwen_recipes.py, test_auto_bridge.py, test_gpt_model_config.py and test_model_bridge.py.

ruff check / ruff format --check are clean (ruff 0.16.7; pre-commit itself not run).

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

• Part of the incremental provider → builder migration (refactor(model): add Qwen3 builder-backed model config with YaRN support #6365 Qwen3, merged; then two families per PR: 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, refactor(model): build Ling MoE V2 and EXAONE 4.0 from builder-backed model configs #6401).
• Qwen2.5 YaRN configs that omit beta_fast / beta_slow still fail at model construction, on both paths, exactly as on main; fix(model): apply Hugging Face YaRN defaults for omitted rope_scaling keys #6367 fixes this for all GPT families.

🤖 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 e4e13cd

[kamran-nvidia]

/ok to test 4903e5c