GLM-5.1 MoE Port - WIP Workspace State

This repository (SwiftLM) is currently heavily coupled with local instances of its backend dependencies to natively support the baa-ai/GLM-5.1-RAM-270GB-MLX model.

As of April 2026, the local workspace is structured in three active W.I.P repositories that must be maintained together on the glm5.1 branch for the SwiftLM engine to boot correctly.

1. SwiftLM (Root)

• Path: /Users/simba/SwiftLM
• Branch: glm5.1
• Purpose: Inference API, terminal logging, and profiling benchmark orchestrator.
• Status:
  • Forced Package.swift to point to the local ../mlx-swift and ./mlx-swift-lm paths instead of resolving GitHub remotes to allow cross-repo C++ debugging.
  • Updated profile_runner.py to pipe standard output for runtime overcommit monitoring.

2. mlx-swift-lm (Local Sub-Repository)

• Path: /Users/simba/SwiftLM/mlx-swift-lm
• Branch: glm5.1
• Purpose: Defines MLX graph execution (Architecture mapping & Hugging Face Safetensors index bridging).
• Status:
  • Defines GLMMoeDSA architecture and dense MLP SwitchGLU layers.
  • Uses aggressive parameter pruning during sanitize() to gracefully drop self_attn.indexer keys and execute the model using standard Multi Latent Attention (MLA).

3. mlx-swift (C++ / Metal Backend Runtime)

• Path: /Users/simba/SwiftLM/mlx-swift
• Branch: glm5.1
• Purpose: MLX C-Bindings, kernel dispatch, and zero-copy NVMe SSD streaming.
• Status:
  • Unlocked unbounded memory chunk streaming in LoadSSDExpert. Removed the hardcoded [SSDStreamer] Load length exceeds Pinned Buffer capacity condition inside ssd_streamer.mm to allow streaming aggregated massive tensor blocks directly into Metal-configured Unified Memory.

---

Warning for Deployment

Before deploying or returning SwiftLM to production (main):

1. Wait for SharpAI/mlx-swift to successfully merge the glm5.1 branch updates.
2. Wait for SharpAI/mlx-swift-lm to merge the GLM-5.1 specific parameter logic.
3. Update Package.swift here in the root to repoint to the latest GitHub release tags instead of local ../mlx-swift paths.