Add native GLM-5.2 sparse execution

[i386]

Overview

Add native GLM-5.2 sparse execution to the Mesh-LLM llama.cpp fork, based directly on the llama.cpp revision pinned by mesh-llm.

The branch is intentionally organized as four reviewable commits:

1. Add GLM-5.2 IndexShare schedule metadata across GGUF conversion, loading, validation, and saving.
2. Implement native Full/Shared GLM-DSA behavior, NORM RoPE, compressed KV storage, and index reuse.
3. Add F16/F32/BF16 and quantized Q8_0, Q5_1, Q5_0, Q4_1, Q4_0, and IQ4_NL GGML_OP_LIGHTNING_INDEXER execution on Metal.
4. Add compact selected-KV sparse decode through Flash Attention, plus correctness and regression tests.

This branch contains no Skippy ABI, staged-runtime hooks, stage sidebands, mesh scheduling, or Skippy-specific configuration. Those integration concerns remain owned by mesh-llm. Native MTP performance work is not claimed by this change.

Additional information

Base revision: 99f3dc32296f825fec94f202da1e9fede1e78cf9

Validation performed on Apple Metal:

• Clean release build completed successfully.
• Focused Metal Lightning Indexer parity passed 108/108 cases across dense and quantized cache types.
• Q4_0 Lightning Indexer at a 4K KV length measured 2,863.90 us on CPU versus 240.48 us on Metal, or 11.91x faster for the isolated operation.
• Quantized Metal matmul regression matrix passed 1,810/1,810.
• Greedy replay passed 256/256 tokens at 2K, 32K, and 128K contexts.
• Downstream two-node GLM-5.2 integration loaded the real Q4_0 kernel on both stages and completed an OpenAI-compatible 128K smoke. Short-context end-to-end throughput remained effectively neutral, so no whole-model speedup is claimed here.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Codex was used to transplant and separate the native GLM changes, remove staged-runtime coupling, and run the documented validation. The submitter remains responsible for reviewing all changes.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: f52dc72d-3595-42eb-8757-3da69785e68a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch glm-5.2

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}