refactor(modelopt): delegate real-quant export to ModelOpt

[mxinO]

Summary

Delegates real-quant packing and deployment metadata to ModelOpt while keeping distributed topology and Hugging Face naming in Megatron-Bridge.

• Reuses existing TP, PP, EP, and ETP conversion tasks for quantized weights.
• Handles fused QKV, gate/up, and numbered grouped-expert weights.
• Stores only topology and format metadata in the reusable plan.
• Recaptures current amax and scale tensors for each export stream.
• Uses object collectives only for small metadata and typed collectives for tensor state.
• Forwards all ModelOpt-emitted packed weights and sidecars without format-specific scale tables.

Unmerged adapter export and quantized mappings requiring dimension permutation are rejected explicitly. Ordinary non-ModelOpt export paths are unchanged.

Ownership

• ModelOpt: format interpretation, state transforms, packing, sidecars, and quantization_config.
• Megatron-Bridge: topology conversion, expert identity, and canonical Hugging Face names.

Validation

• 13 focused tests passed, including distributed state transport and expert gathering.
• ModelOpt and Megatron-Bridge pre-commit hooks passed.
• Downstream two-step W4A16/NVFP4 QA-GRPO completed two independent refits with finite loss and KL.

Dependency

Requires NVIDIA/Model-Optimizer#2251.

Checklist

• Added topology, grouped-weight, expert, and state-refresh tests.
• Updated ModelOpt quantization documentation.
• Kept optional ModelOpt imports scoped to ModelOpt paths.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.