feat(backends): add CachyLLaMA support

[localai-org-maint-bot]

Summary

• add CachyLLaMA as a drop-in GGUF backend using LocalAI’s llama.cpp gRPC integration
• publish Linux CPU/Vulkan builds for amd64 and arm64 plus Darwin Metal, with backend gallery, importer, dependency-bump, and path-filter wiring
• document SSD-backed prompt cache configuration and expose GGUF/mmproj preferences in the React import form

Testing

• make test-ci-scripts (41 passed)
• parsed backend/index.yaml, backend matrices, and edited workflows with go.yaml.in/yaml/v3
• bash -n on all new shell scripts
• make -n docker-build-cachyllama
• git diff --check
• focused assertions for matrix/image counts, Darwin serialization, and ggml-rpc-server

The importer Go suite and focused React Playwright spec could not be run locally because this environment blocks the required protoc/npm release downloads. The full backend image build is left to CI because repository guidance requires approval before long local builds.

Fixes #10590

[localai-org-maint-bot]

Review pass. The mechanical wiring here is the most complete backend addition I have reviewed against .agents/adding-backends.md, but it is stale in a way that breaks the build.

Blocking: missing disable-tts-task.sh.

backend/cpp/cachyllama/Makefile:46,80 runs only disable-score-task.sh, while backend/cpp/turboquant/Makefile:50-51 and backend/cpp/bonsai/Makefile:44-45 run disable-tts-task.sh as well. That script was introduced by 7b9167eaa (#11392, Qwen3-TTS) merged 2026-08-10, one day after this PR's last CI run.

Since backend/cpp/cachyllama/Makefile:44 deletes the copied patches/ directory, a rebuild against current master compiles SERVER_TASK_TYPE_TTS and the mtmd gen-audio references against a fork that has no TTS patch. The darwin script has the same omission (only two disable-* calls exist in the whole diff, both score-task).

I have not built it, so this is inference rather than a confirmed failure, but the fix is one line per build target and CI needs re-running against current master either way. Every check on this PR is from 2026-08-09, so the green marks describe a tree that no longer exists.

Worth a second look: the shared grpc-server include.

backend/cpp/llama-cpp/grpc-server.cpp:46-54 adds two __has_include pulls of CachyLLaMA-only source into the shared monolithic translation unit. The guards make it inert today, but that file is consumed by llama-cpp, ik-llama-cpp, turboquant and bonsai, so a future upstream name collision would drag a foreign implementation into their builds silently.

Checklist, for the record. Wired: both matrix sections including includeDarwin, all four scripts/lib/backend-filter.mjs touchpoints with tests, bump_deps.yaml with the pin in the Makefile, backend/index.yaml meta plus six image entries with matching tag suffixes and metal:, and importer integration correctly as a drop-in via llama-cpp.go:40,140 with Ginkgo tests and React wiring. The engineNamePreferenceRules omission is correct, since "cachyllama" does not contain the llama-cpp substring and so ranks last, matching turboquant and bonsai.

Two small ones: the comment at pkg/system/capabilities.go:161 ("A Vulkan host has exactly one LLM engine with a Vulkan build") becomes false on merge, and .NOTPARALLEL gains backends/cachyllama when only backends/cachyllama-darwin exists.

Scope, for a maintainer rather than the author: this is the fourth llama.cpp fork backend and adds five image builds per cycle to a matrix already at 458 Linux entries. It answers a real user request (#10590) and upstream is alive, so this is a CI cost question, not a merit one.

Also note DCO: no commit here carries a Signed-off-by.

[localai-org-maint-bot]

The cachyllama backend builds fail because common_json_error is an unknown type in the CachyLLaMA fork — the fork branched from llama.cpp before this type was introduced. The shared grpc-server.cpp uses it in catch clauses (lines 315, 345, 374). A fork-skew probe or compatibility shim is needed, similar to the LLAMA_LOAD_MODE_MMAP probe in prepare.sh.