Update llama.cpp patch queue

[i386]

Summary

• update the pinned llama.cpp upstream from 99f3dc32296f825fec94f202da1e9fede1e78cf9 to e8f19cc0ad70a243c8012bf17b4be601abfc8ea2
• rebase and regenerate the 16-patch Skippy queue against the new upstream
• preserve upstream MiniMax M2 per-layer input tracking while retaining Skippy's filtered layer range
• normalize the duplicate final patch numbering from 0015 to 0015 / 0016

Why

The existing patch queue no longer applied cleanly to current llama.cpp master. Patch 2 conflicted in src/models/minimax-m2.cpp after upstream added layer-input bookkeeping.

Impact

Keeps the branch-local staged llama.cpp runtime current while preserving the existing Skippy ABI and patch-series semantics. The range comparison shows the MiniMax M2 conflict resolution is the only semantic adaptation; the remaining carried commits are unchanged.

Validation

• clean-checkout scripts/prepare-llama.sh pinned
• just build
• cargo test -p mesh-llm --lib (41 passed)

Summary by CodeRabbit

• New Features
  • Added experimental staged inference for splitting model execution across layer ranges.
  • Added model inspection and GGUF slice creation tools.
  • Added chat-template support, reasoning-format handling, grammar improvements, and broader model compatibility.
  • Added token verification, batch sampling, external decoding, and optional MTP draft generation.
  • Added session state, KV-cache, resident-prefix, and recurrent-state save/restore capabilities.
  • Added backend device discovery and model-loading progress events.
• Bug Fixes
  • Improved activation validation, cache handling, detokenization, and grammar error recovery.
  • Correctly reports tensor sizes and supports additional model families.

[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: 0199f6d7-d55a-42ac-8472-e3d9f85dde49

📥 Commits

Reviewing files that changed from the base of the PR and between f13c9dc and 608f236.

📒 Files selected for processing (17)

• third_party/llama.cpp/patches/0001-Add-Skippy-ABI-and-package-writer-foundation.patch
• third_party/llama.cpp/patches/0002-Add-early-staged-model-family-and-chat-support.patch
• third_party/llama.cpp/patches/0003-Add-staged-sampling-checkpoints-and-part-loading.patch
• third_party/llama.cpp/patches/0004-Add-lanes-external-media-and-chat-grammar-support.patch
• third_party/llama.cpp/patches/0005-Add-resident-prefix-cache-and-session-refinements.patch
• third_party/llama.cpp/patches/0006-Expand-staged-execution-across-dense-and-recurrent-f.patch
• third_party/llama.cpp/patches/0007-Expand-staged-execution-across-VL-and-broad-model-fa.patch
• third_party/llama.cpp/patches/0008-Add-external-decode-media-prefill-and-newer-family-s.patch
• third_party/llama.cpp/patches/0009-Add-chat-grammar-device-enumeration-and-runtime-even.patch
• third_party/llama.cpp/patches/0010-Add-MTP-execution-support-and-sampling-cleanup.patch
• third_party/llama.cpp/patches/0011-Pass-reasoning-format-through-stage-chat-templates.patch
• third_party/llama.cpp/patches/0012-Wire-mmap-and-mlock-runtime-load-options.patch
• third_party/llama.cpp/patches/0013-Add-external-MTP-draft-sidecar-attachment.patch
• third_party/llama.cpp/patches/0014-Add-non-frame-native-MTP-decode-ABI.patch
• third_party/llama.cpp/patches/0015-Fix-stage-activation-graph-input-allocation.patch
• third_party/llama.cpp/patches/0016-Recognize-thinking-field-in-chat-auto-parser.patch
• third_party/llama.cpp/upstream.txt

---

📝 Walkthrough

Walkthrough

The patch series extends llama.cpp with Skippy staged execution, activation sidebands, session and KV state transfer, resident-prefix caching, runtime events, device enumeration, native MTP execution, chat-template reasoning formats, and chat auto-parser coverage.

Changes

Skippy staged runtime

Layer / File(s)	Summary
Staged graph and activation plumbing third_party/llama.cpp/patches/0001-*, 0002-*, 0007-*, 0008-*	Adds layer-range graph filtering, activation sidebands, filtered tensor handling, and stage-boundary outputs across supported model families.
Session state, KV pages, and resident prefixes third_party/llama.cpp/patches/0005-*	Adds session state import/export, recurrent-state APIs, native KV-page serialization, resident-prefix lifecycle APIs, and thread-count configuration.
Runtime events, devices, grammar, and cache compaction third_party/llama.cpp/patches/0009-*	Adds runtime model-open events, backend enumeration, min-p sampling, grammar exception handling, lazy trigger trimming, and KV-cache compaction.
Native MTP execution and sampling third_party/llama.cpp/patches/0010-*, 0014-*, 0015-*	Adds MTP graph contexts, next-token validation, model-specific stage filtering, and shared top-k handling.
Chat templates, external MTP, and parser coverage third_party/llama.cpp/patches/0011-*, 0013-*, 0016-*	Passes reasoning formats through chat templates, supports external MTP draft attachment, and recognizes thinking reasoning fields with tests.
Patch-series synchronization third_party/llama.cpp/patches/0003-*, 0004-*, 0006-*, 0012-*, third_party/llama.cpp/upstream.txt	Updates embedded patch metadata and the recorded upstream revision.

Estimated code review effort: 5 (Critical) | ~120 minutes

Possibly related PRs

• Mesh-LLM/mesh-llm#839 — Shares the Skippy ABI and staged-execution patch queue.
• Mesh-LLM/mesh-llm#842 — Connects to the runtime event reporting ABI.
• Mesh-LLM/mesh-llm#858 — Uses the native MTP and batched verification APIs.

Suggested reviewers: ndizazzo, michaelneale

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch jd/refresh-llama-patch-queue

---

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.}

[github-actions]

This pull request is currently a draft. Reviews will not take place until the PR is marked as ready for review.