Fix GLM-DSA MTP split-stage load crash

[i386]

Summary

• add a llama patch-queue fix that checks whether a GLM-DSA final split stage has the tensors needed by the native MTP sidecar
• skip optional native MTP draft context creation when those tensors are absent, instead of letting the reused DeepSeek MTP graph dereference null token/head tensors
• keep non-GLM native MTP behavior unchanged

Why

GLM-DSA split serving can load a final stage with output tensors but without the token embedding/head tensors used by the native MTP graph fallback. In that shape, llama_init_from_model(... LLAMA_CONTEXT_TYPE_MTP ...) can crash in ggml_get_rows during graph reservation.

Validation

• LLAMA_WORKDIR=/tmp/mesh-llm-llama-validate.1782193330 scripts/prepare-llama.sh pinned
• MESH_LLM_DYNAMIC_NATIVE_RUNTIME=0 LLAMA_STAGE_BACKEND=metal RUSTC_WRAPPER=/opt/homebrew/bin/sccache just release-build

Summary by CodeRabbit

• Bug Fixes
  • Enhanced model processing validation to prevent incomplete initialization of internal components when required model tensors and dependencies are unavailable, improving overall system robustness.
  • Improved diagnostic messaging to clearly indicate when internal processing components are intentionally skipped due to missing prerequisites, enhancing visibility for troubleshooting.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a5f12b91-a746-4cb0-b024-6ce9dfd7ffe8

📥 Commits

Reviewing files that changed from the base of the PR and between 3e6805d and 43ab2c0.

📒 Files selected for processing (1)

• third_party/llama.cpp/patches/0111-Skip-native-MTP-sidecar-when-split-stage-lacks-tensors.patch

---

📝 Walkthrough

Walkthrough

A new patch file adds skippy_should_create_native_mtp_sidecar to src/skippy.cpp, which checks config->include_output, n_layer_nextn, model architecture, and required MTP block tensor presence before creating the native MTP sidecar. skippy_finish_model_open is updated to use this helper and emit a specific diagnostic when skipping.

Changes

Conditional Native MTP Sidecar Creation

Layer / File(s)	Summary
MTP sidecar guard helper and call site third_party/llama.cpp/patches/0111-Skip-native-MTP-sidecar-when-split-stage-lacks-tensors.patch	Adds skippy_should_create_native_mtp_sidecar gating on config->include_output, n_layer_nextn > 0, architecture type, and required tensor presence in the split-stage layer metadata. Updates skippy_finish_model_open to call this helper; logs a "missing required MTP graph tensors" message and continues without drafts when the helper returns false while the base condition is still satisfied.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'Fix GLM-DSA MTP split-stage load crash' directly addresses the main issue being fixed - a crash when loading final stages in GLM-DSA split serving configurations with native MTP sidecars.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch jd/fix-glm-dsa-mtp-split-load

---

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