Add support for Jinja-based tool calling (MiniCPM5 and other recent models).

[vladislavdonchev]

Description

Bumps the llama.cpp submodule to c818263f2, which adds upstream MiniCPM5 chat/tool-call parsing (ggerganov/llama.cpp#24889). Regenerates the full llama.cpp.patches/ set so the cosmocc build still applies cleanly, and wires new upstream sources into the llamafile build.

Branch: llamacpp_c818263f2 (per CONTRIBUTING.md / bump workflow — not committed directly to main)

Submodule

• llama.cpp → c818263f2 (chat: implement minicpm5 parser)

Patch / build reconciliation

• Regenerated all 39 patches via generate-patches after proving a clean build + make check
• llamafile-files/BUILD.mk: drop removed json-partial.cpp; add dflash.cpp and server-stream.cpp
• llamafile/BUILD.mk: link server-stream.cpp.o in LLAMAFILE_SERVER_SUPPORT_OBJS
• New src_models_dflash.cpp.patch: cosmocc/clang fix — move build_arch_graph() after graph<true>/graph<false> template specializations (same pattern as the existing eagle3 patch)
• Refreshed drifted hunks in backend/server patches (CUDA, Vulkan, CPU ops, server models, etc.)

Tests

• Tool-calling integration tests pass --jinja -ngl 0 so Jinja-templated models (including MiniCPM5) parse tool calls correctly

Verification performed locally

• make setup — all patches apply cleanly
• Clean build of o/llamafile/llamafile
• make check — unit tests green
• llamafile:verify-clean round-trip (reset → setup → clean build → check)
• Integration: tests/test_tool_calling.py (3/3) against built binary and rebundled .llamafile with MiniCPM5-1B Q8_0

Host handoff (not run in-session — per bump doc)

• CUDA / ROCm smoke test
• Windows smoke test
• macOS Metal runtime compile + run
• Web UI curl check (GET /, /_app/immutable/...)
• Long-run stability

PR Type

• 🆕 New Feature
• 🚦 Infrastructure

Relevant issues

• Enables MiniCPM5 tool calling via upstream ggerganov/llama.cpp#24889
• Searched mozilla-ai/llamafile issues and open PRs — no duplicate; this bump is novel
• No pre-discussion issue filed: confined routine llama.cpp submodule patch bump (not a new user-facing feature or architectural change per CONTRIBUTING.md §Discuss major changes)

Checklist

• I understand the code I am submitting.
• I have run this code locally and verified the change.
• New and existing tests pass locally, or I have explained why tests were not run.
• Documentation was updated where necessary.
• If I changed code in llama.cpp/, whisper.cpp/, or stable-diffusion.cpp/, I also updated the matching *.patches/ files.
• I have read and followed the contribution guidelines.
• AI Usage:
  • No AI was used.
  • AI was used in an assistive capacity.
  • This PR includes substantial AI-generated content.

AI Usage Information

• AI Model used: Grok (xAI, Research, Code), MiniCPM5 (Action Classifier)
• AI Developer Tool used: opencode an autonomous VM sandbox
• Any other info you'd like to share: AI assisted with the canonical llama.cpp bump workflow (check_patches.sh → reconcile → generate-patches → verify-clean), patch regeneration, BUILD.mk reconciliation for upstream file moves, the cosmocc dflash.cpp compile-order fix, integration-test updates (--jinja), and local verification (build, unit tests, MiniCPM5 tool-calling integration tests). All changes were reviewed against llamafile's patch workflow docs and CONTRIBUTING.md; submodule edits were made only in-place for patch generation, with llama.cpp.patches/ as the committed source of truth.

When answering reviewer questions, please respond yourself rather than pasting reviewer comments into an AI system and posting the reply back unchanged.

• I am an AI Agent filling out this form (check box if true)