Add GLM-5.1 ms-swift LoRA fine-tuning recipe

[William-Gao1]

Adds examples/glm-5.1-msswift — LoRA SFT of GLM-5.1 (744B glm_moe_dsa MoE, 256 routed experts, DeepSeek Sparse Attention) via MS-Swift + Megatron, mirroring the existing glm-4.7-msswift recipe.

Key differences from the GLM-4.7 recipe

• Image: requires megatron-core >= 0.17.1 (...megatron0.17.1-msswift4.3.1-peftstamp). Earlier builds lack the DSA indexer kernel and hard-reject GLM-5.1's interleaved-RoPE + multi-latent-attention combo.
• Compute: 4 nodes × 8 H200 (expert_model_parallel_size=32). The 256 experts shard cleanly only across power-of-2 GPU counts; 24 GPUs can't fit the weights.
• Args: --template glm5_1, --tuner_type lora, --output_dir, --dataloader_num_workers (the 4.x ms-swift arg surface).
• max_length=16384: longer context currently OOMs in the DSA indexer's O(seqlen²) score matrix; context-parallel support for DSA isn't available upstream yet.

Validation

Ran end-to-end on 4×8 H200 at 16k seqlen — completed training and saved a LoRA checkpoint.

[github-actions]

Example Validation Results

❌ examples/glm-5.1-msswift/training/config.py

Status: EXCEPTION | Job: `` | Project: pr-116-examples--glm-5.1-msswift

Error: 400 Client Error: Bad Request for url: https://api.baseten.co/v1/training_projects/8w6zr53/jobs

Cache Summary

▪▪▪▪ There's a new version of truss available, 0.18.13 (you are currently on 
0.16.0)!
▪▪▪▪ To upgrade to the latest version, run: truss upgrade
▪▪▪▪ To disable this check, set `check_for_updates` to false in 
/home/runner/.config/truss/settings.toml
Error fetching cache summary: 404 Client Error: Not Found for url: https://api.baseten.co/v1/training_projects/8w6zr53/cache/summary
Client error: No training jobs with cache enabled found in this project.
ERROR HTTPError: 404 Client Error: Not Found for url: 
https://api.baseten.co/v1/training_projects/8w6zr53/cache/summary

Checkpoints

(no job ID available)