chore: update pinned llama.cpp revision

[i386]

Summary

• Update the pinned llama.cpp revision from 8190848bb36c7df4251db4352bd81bc07d0a4385 to e9fa0781f1c25fc4fe8c86be1edc6970661ad6f0.
• Remove the obsolete RPC transport patch; no live mesh-llm source references ggml_backend_rpc or RPC_PROTO_*, and the remaining Skippy queue applies without it.
• Carry the upstream WebGPU test-skip change into the staged-runtime test patch.

Stack

This PR is based on PR #1211 (jd/fix-issue-986-reasoning-effort-on-1194).

Validation

• Regenerated the remaining 13-patch queue with contiguous 0001–0013 numbering.
• Fresh-checkout scripts/prepare-llama.sh pinned replay produced the exact same patched tree; idempotence verified.
• just llama-build — patched CPU/Metal native libraries built successfully.
• just build — full debug product build passed.
• git diff --check — passed.

Summary by CodeRabbit

• New Features

  • Added support for Inkling text, vision, and audio models.
  • Introduced staged execution for prefilling, decoding, verification, sampling, and speculative generation.
  • Added Skippy APIs for model loading, device selection, sessions, state management, activations, tokenization, and chat handling.
  • Added support for multi-part model packages and model slicing.
  • Added improved chat-template options, tool-call parsing, and reasoning controls.

• Bug Fixes

  • Improved validation and error handling for model data, sessions, activations, buffers, and chat inputs.

• Refactor

  • Removed the legacy chat-template interface and updated the public ABI.

[github-actions]

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

[coderabbitai]

📝 Walkthrough

Walkthrough

The change refreshes the llama.cpp revision and patch queue, removes the obsolete RPC transport patch, preserves staged Skippy runtime support, and finalizes chat-template ABI changes.

Changes

Staged Skippy runtime

Layer / File(s)	Summary
Patch queue and upstream alignment third_party/llama.cpp/upstream.txt, third_party/llama.cpp/patches/*	The upstream commit and embedded patch metadata are updated for a 13-patch series. The obsolete RPC transport patch is deleted.
Staged model and runtime execution third_party/llama.cpp/patches/0001-Add-staged-model-graph-and-family-support.patch, third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch, third_party/llama.cpp/patches/0011-Wire-staged-runtime-builds-and-tests.patch	Stage filtering, boundary outputs, session state, KV-page handling, execution lanes, GLM-DSA instrumentation, and backend tests are included.
Chat API and ABI finalization third_party/llama.cpp/patches/0012-Pass-generic-chat-template-kwargs-through-Skippy.patch, third_party/llama.cpp/patches/0013-skippy-remove-legacy-chat-template-ABI.patch	Chat-template kwargs and generation-prompt test selection are added. The legacy skippy_apply_chat_template API is removed, and the ABI patch version increases to 37.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Client
  participant skippy_model
  participant skippy_session
  participant llama_decode
  Client->>skippy_model: Open model and select execution lane
  skippy_model->>skippy_session: Create or reset session
  Client->>skippy_session: Submit staged prefill or decode request
  skippy_session->>llama_decode: Execute filtered graph
  llama_decode-->>skippy_session: Return tokens, activation frame, or state
  skippy_session-->>Client: Return output and runtime signals

Loading

Possibly related PRs

• Mesh-LLM/mesh-llm#1211: Covers the same Skippy chat-template kwargs and legacy ABI changes.
• Mesh-LLM/mesh-llm#1099: Updates the llama.cpp revision and removes an obsolete RPC patch.
• Mesh-LLM/mesh-llm#1104: Reverts a refreshed llama.cpp patch queue and upstream revision.

Suggested reviewers: michaelneale

🚥 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 clearly identifies the primary change: updating the pinned llama.cpp revision.
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.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch agent/update-llama-pin-on-1211

---

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.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In
`@third_party/llama.cpp/patches/0012-Wire-staged-runtime-builds-and-tests.patch`:
- Around line 1232-1234: Update the GLM-DSA contract test flow so WebGPU builds
skip before executing the GLM-DSA contract block. Move the GGML_USE_WEBGPU skip
ahead of that block, or make glm_dsa_backend_config_supported() return false for
WebGPU devices, while preserving existing arch_supported() and backend skip
behavior.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c9ec5fcc-12fd-4112-851c-3e3cd30ad5b0

📥 Commits

Reviewing files that changed from the base of the PR and between b498e45 and 0155776.

📒 Files selected for processing (3)

• third_party/llama.cpp/patches/0001-Add-RPC-backend-tensor-transport.patch
• third_party/llama.cpp/patches/0012-Wire-staged-runtime-builds-and-tests.patch
• third_party/llama.cpp/upstream.txt

💤 Files with no reviewable changes (1)

• third_party/llama.cpp/patches/0001-Add-RPC-backend-tensor-transport.patch

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (5)

third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch (1)

299-304: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Reject duplicate execution lanes in batch requests.

Both APIs accept the same skippy_session twice. This creates multiple batch rows with one seq_id at the same position. The decode can corrupt that session's KV state.

• third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch#L299-L304: Track and reject duplicate session->seq_id values before frame-batch decode.
• third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch#L760-L765: Apply the same uniqueness check before token-batch decode.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch`
around lines 299 - 304, Reject duplicate execution lanes by tracking each
session->seq_id during request validation and returning
SKIPPY_STATUS_INVALID_ARGUMENT before decoding when a sequence ID repeats. Apply
this uniqueness check in both frame-batch and token-batch validation loops at
third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch lines
299-304 and 760-765.

third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch (1)

1287-1326: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Clear MTP sidecar state after every MTP decode failure.

These paths return after llama_decode fails but retain pending hidden state and sidecar cache data. A later proposal can use stale state.

• third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch#L1287-L1326: Clear the MTP session state before returning a sidecar-sync failure.
• third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch#L1462-L1467: Clear the MTP session state before returning a chained-proposal failure.
• third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch#L1510-L1512: Clear the MTP session state before returning a single-step proposal failure.

Based on learnings, clear sidecar state after MTP failures to prevent stale proposals.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch`
around lines 1287 - 1326, Clear all pending MTP hidden state and sidecar cache
data before returning from each MTP failure path: the sidecar-sync failure at
third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch
lines 1287-1326, the chained-proposal failure at lines 1462-1467, and the
single-step proposal failure at lines 1510-1512. Reuse the existing MTP
session-state reset mechanism so later proposals cannot consume stale state.

Source: Learnings

third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch (1)

793-809: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Reset sampling state after external session-state changes.

Imported state can represent tokens that differ from token_history. The existing sampling chain, grammar sampler, and accepted-token count then retain state from a different sequence.

• third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch#L793-L809: Clear token and signal history, then reset or rebuild sampling state after full or partial state import.
• third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch#L1242-L1248: Reset sampling state after KV-page import changes the resident context.
• third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch#L1496-L1504: Rebuild sampling and grammar state from the restored prefix tokens.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch`
around lines 793 - 809, Update skippy_update_session_state_after_import to clear
token_history and signal_history for imported state, then reset or rebuild the
sampling chain, grammar sampler, and accepted-token count from the restored
prefix. In
third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch
lines 1242-1248, reset sampling state after KV-page import changes the resident
context. In lines 1496-1504, rebuild sampling and grammar state from the
restored prefix tokens.

third_party/llama.cpp/patches/0005-Add-Skippy-model-lifecycle-and-package-support.patch (2)

1581-1582: 🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Use 64-bit file positioning for GGUF data.

On Windows, long is 32-bit. The casts at lines 1582 and 1838 truncate offsets at the 2 GiB boundary. This can corrupt metadata parsing or tensor packaging. Use one checked 64-bit seek helper for both calls.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@third_party/llama.cpp/patches/0005-Add-Skippy-model-lifecycle-and-package-support.patch`
around lines 1581 - 1582, Update skippy_skip and the other seek call around the
GGUF parsing/packaging logic to use a shared checked 64-bit file-positioning
helper instead of casting byte offsets to long. Ensure both seek paths handle
offsets beyond 2 GiB correctly on Windows and preserve failure reporting when
the seek cannot be completed.

---

1540-1545: 🩺 Stability & Availability | 🟠 Major | 🏗️ Heavy lift

Prevent exceptions from crossing the C ABI.

If a GGUF tensor name contains an oversized layer index, std::stol throws std::out_of_range before skippy_model_info_open can return. Use a checked, non-throwing parser, validate the int32_t range, and return SKIPPY_STATUS_MODEL_ERROR for malformed metadata. The same path is used by skippy_write_gguf_from_parts.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@third_party/llama.cpp/patches/0005-Add-Skippy-model-lifecycle-and-package-support.patch`
around lines 1540 - 1545, Update the tensor-name layer-index parsing in the
helper used by skippy_model_info_open and skippy_write_gguf_from_parts to avoid
throwing std::stol; use checked non-throwing conversion, validate the parsed
value fits int32_t, and return SKIPPY_STATUS_MODEL_ERROR for malformed or
oversized indices without allowing exceptions across the C ABI.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In
`@third_party/llama.cpp/patches/0005-Add-Skippy-model-lifecycle-and-package-support.patch`:
- Around line 1581-1582: Update skippy_skip and the other seek call around the
GGUF parsing/packaging logic to use a shared checked 64-bit file-positioning
helper instead of casting byte offsets to long. Ensure both seek paths handle
offsets beyond 2 GiB correctly on Windows and preserve failure reporting when
the seek cannot be completed.
- Around line 1540-1545: Update the tensor-name layer-index parsing in the
helper used by skippy_model_info_open and skippy_write_gguf_from_parts to avoid
throwing std::stol; use checked non-throwing conversion, validate the parsed
value fits int32_t, and return SKIPPY_STATUS_MODEL_ERROR for malformed or
oversized indices without allowing exceptions across the C ABI.

In
`@third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch`:
- Around line 793-809: Update skippy_update_session_state_after_import to clear
token_history and signal_history for imported state, then reset or rebuild the
sampling chain, grammar sampler, and accepted-token count from the restored
prefix. In
third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch
lines 1242-1248, reset sampling state after KV-page import changes the resident
context. In lines 1496-1504, rebuild sampling and grammar state from the
restored prefix tokens.

In `@third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch`:
- Around line 299-304: Reject duplicate execution lanes by tracking each
session->seq_id during request validation and returning
SKIPPY_STATUS_INVALID_ARGUMENT before decoding when a sequence ID repeats. Apply
this uniqueness check in both frame-batch and token-batch validation loops at
third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch lines
299-304 and 760-765.

In
`@third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch`:
- Around line 1287-1326: Clear all pending MTP hidden state and sidecar cache
data before returning from each MTP failure path: the sidecar-sync failure at
third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch
lines 1287-1326, the chained-proposal failure at lines 1462-1467, and the
single-step proposal failure at lines 1510-1512. Reuse the existing MTP
session-state reset mechanism so later proposals cannot consume stale state.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 45b2ff03-9386-42c5-9a5d-907218069969

📥 Commits

Reviewing files that changed from the base of the PR and between 0155776 and 69a822b.

📒 Files selected for processing (13)

• third_party/llama.cpp/patches/0001-Add-staged-model-graph-and-family-support.patch
• third_party/llama.cpp/patches/0002-Add-GLM-DSA-backend-execution-support.patch
• third_party/llama.cpp/patches/0003-Add-Inkling-model-and-multimodal-support.patch
• third_party/llama.cpp/patches/0004-Add-Skippy-public-ABI-surface.patch
• third_party/llama.cpp/patches/0005-Add-Skippy-model-lifecycle-and-package-support.patch
• third_party/llama.cpp/patches/0006-Add-Skippy-session-and-state-management.patch
• third_party/llama.cpp/patches/0007-Add-Skippy-activation-frame-handling.patch
• third_party/llama.cpp/patches/0008-Add-Skippy-staged-execution-paths.patch
• third_party/llama.cpp/patches/0009-Add-Skippy-sampling-and-speculative-decoding.patch
• third_party/llama.cpp/patches/0010-Add-Skippy-tokenization-and-stage-chat.patch
• third_party/llama.cpp/patches/0011-Wire-staged-runtime-builds-and-tests.patch
• third_party/llama.cpp/patches/0012-Pass-generic-chat-template-kwargs-through-Skippy.patch
• third_party/llama.cpp/patches/0013-skippy-remove-legacy-chat-template-ABI.patch

💤 Files with no reviewable changes (5)

• third_party/llama.cpp/patches/0002-Add-GLM-DSA-backend-execution-support.patch
• third_party/llama.cpp/patches/0010-Add-Skippy-tokenization-and-stage-chat.patch
• third_party/llama.cpp/patches/0003-Add-Inkling-model-and-multimodal-support.patch
• third_party/llama.cpp/patches/0007-Add-Skippy-activation-frame-handling.patch
• third_party/llama.cpp/patches/0001-Add-staged-model-graph-and-family-support.patch