feat(llama-cpp): expose 12 missing common_params via options[]

[localai-bot]

Summary

A coverage audit of params_parse in backend/cpp/llama-cpp/grpc-server.cpp against upstream llama.cpp (pin 7f3f843c) flagged 12 user-visible common_params knobs that were neither set via the typed ModelOptions proto fields nor reachable through the existing options: array. This PR plumbs them through, following the file's existing patterns.

Newly accepted options: keys

Top-level / batching / IO

• n_ubatch (alias ubatch) — physical batch size; previously force-aliased to n_batch, blocking embedding/rerank workloads needing independent control
• threads_batch (alias n_threads_batch) — main-model batch threads, mirrors the existing draft_threads_batch
• direct_io (alias use_direct_io) — O_DIRECT model loads
• verbosity — llama.cpp log threshold (was commented out at line 479)
• override_tensor (alias tensor_buft_overrides) — per-tensor buffer overrides for the main model, mirrors draft_override_tensor

Embedding / multimodal

• pooling_type (alias pooling) — mean/cls/last/rank/none for embeddings (was only auto-flipped to RANK for rerankers)
• embd_normalize (alias embedding_normalize) — -1/0/1/2/>2; the embedding handler also now reads params_base.embd_normalize instead of a hardcoded 2
• mmproj_use_gpu (alias mmproj_offload) — keep mmproj on CPU
• image_min_tokens / image_max_tokens — per-image vision token budget

Reasoning surface (audit-confirmed gaps)

LocalAI's existing ReasoningConfig.DisableReasoning only feeds the per-request chat_template_kwargs.enable_thinking body field at request time; it does not touch the load-time params.reasoning_format parser or params.enable_reasoning budget. Those remained unwired until this PR.

• reasoning_format — none/auto/deepseek/deepseek-legacy parser for <think> blocks (default: deepseek)
• enable_reasoning (alias reasoning_budget) — -1 unlimited / 0 off / >0 token budget
• prefill_assistant — toggle the trailing-assistant-message prefill in the chat template

Compatibility

All 14 referenced fields/types exist on both the upstream pin (7f3f843c) and on the TheTom/llama-cpp-turboquant fork's common.h, so no LOCALAI_LEGACY_LLAMA_CPP_SPEC guard is needed.

Docs

docs/content/advanced/model-configuration.md gets four new subsections:

• Reasoning Models (load-time options + a note clarifying the per-request enable_thinking distinction)
• Multimodal Backend Options
• Embedding & Reranking Backend Options
• Other Backend Tuning Options

The Speculative Type Values table is also refreshed to show the new dash-separated canonical names from upstream ggml-org/llama.cpp#22964 alongside the underscore aliases LocalAI still accepts.

Test plan

• tests-llama-cpp-grpc CI job builds the backend cleanly (compile gate)
• Spot-check a model config with options: [reasoning_format:deepseek, enable_reasoning:-1] against a DeepSeek-R1 GGUF and confirm reasoning blocks parse correctly
• Spot-check a vision model with options: [mmproj_use_gpu:false] and confirm the projector loads on CPU (memory drops on the GPU)
• Spot-check an embedding model with options: [pooling_type:mean, embd_normalize:0] and confirm the embedding shape/normalization changes
• Smoke options: [n_ubatch:1024] on a rerank workload and confirm we no longer hit the input is too large to process 512-cap