refactor(model): build Hy V3 and MiniMax-M2 from builder-backed model configs

[kamran-nvidia]

What does this PR do ?

Moves Hy V3 and MiniMax-M2 onto the builder-backed ModelConfig / ModelBuilder path for checkpoint conversion, following Muse Glimmer. Neither family has library recipes.

Changelog

• Opt-in after fix(model): default MODEL_CONFIG_CLASS to None and opt in per family #6305: MODEL_CONFIG_CLASS now defaults to None, so each migrated bridge sets MODEL_CONFIG_CLASS = BridgeGPTModelConfig next to USE_MODEL_CONFIG_FOR_CONVERSION = True (as Qwen3 and Llama do), and its tests assert it.
• HYV3Bridge: provider_bridge() is replaced by hf_config_to_model_config_kwargs(), and the provider path reuses the same kwargs. The builder config leaves transformer_layer_spec unset because MCore's builder default for MoE models is already get_gpt_decoder_block_spec, which honors the dense-first moe_layer_freq. The deprecated provider path keeps its explicit decoder block spec in hf_config_to_provider_kwargs(). The forced bf16 dtype fields carry over. USE_MODEL_CONFIG_FOR_CONVERSION = True.
• MiniMaxM2Bridge: same kwargs migration. The builder config carries transformer_layer_spec=minimax_m2_layer_spec; that spec reads only fields the builder config resolves (num_moe_experts, moe_grouped_gemm), so both paths build the full-dimension QK norm. rotary_dim → rotary_percent handling is unchanged. USE_MODEL_CONFIG_FOR_CONVERSION = True.
• Tests:
  • New tests/unit_tests/models/hy_v3/; Hy V3 had no bridge unit tests. It covers the builder mapping, provider/builder parity for bf16 and fp16 sources, and megatron_to_hf_config from a builder config.
  • MiniMax-M2 covers the builder mapping, parity with and without rotary_dim, and layer-spec equivalence between provider and builder config.

Verification

Each family was validated separately on tiny random-weight checkpoints (single GB200 node, NeMo 26.10 container, transformers 5.15 for both HF models).

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	Hy V3 (dense-first MoE, shared expert)	MiniMax-M2 (sigmoid MoE, full-dim QK norm)
Provider fields, main vs this branch	identical (bf16 and fp16 sources)	identical (incl. rotary_dim)
get_model_config() vs provider	0 diffs (19 / 22 on main)	0 diffs (14 / 15 on main)
Export vs source HF weights (GPU TP2, CPU, and from a main-imported checkpoint)	bit-exact except expert-bias dtype (see below)	bit-exact except expert-bias dtype (see below)
Logits: this branch vs main, and this branch on a main-imported checkpoint	bitwise-identical	bitwise-identical
Logits vs HF	cosine 0.999987	cosine 0.999914
roundtrip_checkpoint TP2	pass	pass
Unit tests (family run)	7 passed	36 passed

Regression: 334 tests (1 skipped) across models/hy_v3, models/minimax_m2, test_auto_bridge.py, test_gpt_model_config.py, the AutoBridge registration matrix, conversion/ and doc_consistency.

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

Known, pre-existing: a bf16 source checkpoint exports the router expert bias (mlp.expert_bias / e_score_correction_bias) as fp32, because Megatron keeps it in fp32. The values are identical after a cast, and main's own code exports these checkpoints the same way (verified for both families).

After rebasing onto main with #6305 (Megatron Core a98aed553): 215 passed, 1 skipped, including #6305's all-bridges opt-in check in test_gpt_model_config.py; ruff clean.

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, refactor(model): build Qwen2 and MiMo from builder-backed model configs #6379 Qwen2 + MiMo, refactor(model): build OLMoE and GLM-4.5 from builder-backed model configs #6383 OLMoE + GLM-4.5, two families per PR).
• Independent of the other PRs; targets main directly.

🤖 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 d44b78b

[kamran-nvidia]

/ok to test 3d8a518

[github-actions]

Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing:

/review

Add model=claude to use a Claude reviewer (the default is model=codex). mode=light|strict selects the review depth; for example, /review model=claude mode=strict. Comment /review help for all options.