Add GLM MoE Metal active expert patches

[i386]

Summary

• add a GLM MoE Metal patch queue on top of the merged GLM-DSA llama foundation
• add active-expert Q3 down kernels and Q2 gate/up dispatch controls for GLM-5.2 style expert tensors
• add selector diagnostics and focused backend/perf tests for selected expert paths

Why

This keeps the GLM MoE performance work separate from the core GLM-DSA architecture and IndexShare support. The GLM-DSA foundation has landed in #1034; this PR is the next layer.

Validation

• LLAMA_WORKDIR=$(mktemp -d /tmp/mesh-llm-llama.XXXXXX) scripts/prepare-llama.sh pinned

Summary by CodeRabbit

• Performance

  • Improved Metal acceleration for GLM Mixture-of-Experts models by processing only active experts and skipping zero-weight or inactive slots.
  • Added specialized execution paths for different active-expert counts, potentially reducing dispatch overhead and improving inference efficiency.
  • Updated routing and workload calculations to better reflect active-expert usage.

• Diagnostics

  • Added optional terminal diagnostics for troubleshooting GLM MoE Metal execution.
  • Expanded configuration options for enabling or disabling experimental execution modes.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 3f6215f0-a5d1-49de-8292-6cc6c386b24e

📥 Commits

Reviewing files that changed from the base of the PR and between e92422f and 09ee9ca.

📒 Files selected for processing (12)

• third_party/llama.cpp/patches/0030-ggml-default-GLM-MoE-two-phase-Metal-path.patch
• third_party/llama.cpp/patches/0031-ggml-add-GLM-MoE-Metal-selector-diagnostics.patch
• third_party/llama.cpp/patches/0032-ggml-skip-zero-weight-GLM-MoE-gate-up-slots.patch
• third_party/llama.cpp/patches/0033-tests-cover-full-GLM-MoE-selected-chain.patch
• third_party/llama.cpp/patches/0034-tests-add-GLM-Q2Q3-selected-weight-roofline.patch
• third_party/llama.cpp/patches/0035-ggml-add-active-count-Q3-GLM-MoE-down-kernels.patch
• third_party/llama.cpp/patches/0036-ggml-use-GLM-max-active-policy-for-Q3-down-kernels.patch
• third_party/llama.cpp/patches/0037-ggml-honor-explicit-Q2-gate-up-Metal-variant-flags.patch
• third_party/llama.cpp/patches/0038-ggml-make-GLM-MoE-roofline-honor-active-experts.patch
• third_party/llama.cpp/patches/0039-ggml-avoid-GLM-Q3-fused-tail-under-active-policy.patch
• third_party/llama.cpp/patches/0040-ggml-shrink-GLM-Q2-gate-up-active-dispatch.patch
• third_party/llama.cpp/patches/0041-ggml-skip-inactive-GLM-Q3-down-slots.patch

---

📝 Walkthrough

Walkthrough

GLM MoE Metal execution now defaults to two-phase mode, supports terminal diagnostics and explicit selector flags, skips inactive or zero-weight slots, adds active-count Q3 kernels, applies maximum-active policies to dispatch and roofline paths, and expands backend evaluation and performance coverage.

Changes

GLM MoE Metal execution

Layer / File(s)	Summary
Two-phase diagnostics and selector controls third_party/llama.cpp/patches/0030-...patch, third_party/llama.cpp/patches/0031-...patch	Two-phase execution defaults and override variables are updated; terminal diagnostics, selector precedence, and diagnostic test wiring are added.
Gate-up active-slot dispatch third_party/llama.cpp/patches/0032-...patch, third_party/llama.cpp/patches/0037-...patch, third_party/llama.cpp/patches/0040-...patch	Zero-weight slots are skipped, explicit Q2 variants take precedence, and Q2 dispatch uses active-slot metadata.
Q3 active-slot kernels and dispatch third_party/llama.cpp/patches/0035-...patch, third_party/llama.cpp/patches/0036-...patch, third_party/llama.cpp/patches/0041-...patch	Q3 active2/4/6 pipelines and kernel entrypoints are added, selected through the maximum-active policy, with inactive-slot guards and matching tests.
Active-expert roofline and fusion paths third_party/llama.cpp/patches/0038-...patch, third_party/llama.cpp/patches/0039-...patch	Roofline dispatch and accounting use active experts, while Q3 fused-tail fusion is disabled under the active-expert policy.
Selected-chain and roofline coverage third_party/llama.cpp/patches/0033-...patch, third_party/llama.cpp/patches/0034-...patch	Selected-chain tolerance and variants are expanded, alongside Q2Q3 roofline evaluation and performance cases.

Estimated code review effort: 4 (Complex) | ~60 minutes

Sequence Diagram(s)

sequenceDiagram
  participant GLMActivePolicy
  participant MetalOps
  participant PipelineSelector
  participant MetalKernel
  GLMActivePolicy->>MetalOps: set maximum active experts
  MetalOps->>PipelineSelector: select active-count Q3 pipeline
  PipelineSelector->>MetalKernel: dispatch active2, active4, or active6 kernel
  MetalKernel->>MetalKernel: process only active slots

Loading

Suggested labels: experimental

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch jd/glm-moe-metal-active

---

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.}