Megatron GLM evaluation

[pawalt]

Adds a native Megatron evaluation pipeline for GLM-4.7 LoRA checkpoints on Modal.

This PR introduces:

• A new eval.py script that performs Megatron-native evaluation, incorporating critical patches to correctly load LoRA adapters from distributed checkpoints, especially for MoE models. This addresses issues encountered when evaluating LoRA models within the Megatron-Bridge framework.
• A new eval_lora Modal function in modal_train.py to orchestrate the distributed evaluation. This function handles resolving checkpoint paths, ensures a validation dataset is available (by linking to the existing training dataset if validation.jsonl is missing), and launches the torchrun command across the clustered Modal nodes.
• Updated README.md to document the new evaluation functionality and usage instructions.

This enables users to evaluate their GLM-4.7 LoRA checkpoints on Modal using the existing LongMIT dataset, providing a crucial diagnostic tool for their training pipelines.

Checklist

• Example is documented with comments throughout, in a Literate Programming style.
• Example does not require third-party dependencies to be installed locally
• Example follows the style guide
• Example pins its dependencies
  • Example pins container images to a stable tag, not a dynamic tag like latest (e.g., nvcr.io/nvidia/nemo:25.11)
  • Example specifies a python_version for the base image, if it is used (Implicit in nemo:25.11 image)
  • Example pins all dependencies to at least minor version, ~=x.y.z or ==x.y (Dependencies are primarily handled by the base NeMo image)
  • Example dependencies with version < 1 are pinned to patch version, ==0.y.z (N/A, dependencies handled by base image)

---

 

[cursor]

Cursor Agent can help with this pull request. Just @cursor in comments and I'll start working on changes in this branch.
_{Learn more about Cursor Agents}