feat(gallery): add Gemma 4 llama.cpp MTP variants; fix gemmable-4-12b-mtp

[localai-bot]

What

Wire up native llama.cpp Multi-Token Prediction (MTP) speculative decoding for the official Gemma 4 models in the gallery.

Google shipped the Gemma 4 MTP drafter heads and llama.cpp merged native support in ggml-org/llama.cpp#23398 (merged 2026-06-07). LocalAI's pinned llama.cpp is well past that merge and the config plumbing (draft_model + core/config/mtp.go) already exists, but no official Gemma 4 entry used it.

Added (dense sizes, unsloth-sourced: target UD-Q4_K_XL + mtp-*.gguf drafter + BF16 mmproj)

• gemma-4-e2b-it-qat-mtp
• gemma-4-e4b-it-qat-mtp
• gemma-4-12b-it-qat-mtp
• gemma-4-31b-it-qat-mtp

These replace the previously commented-out/disabled attempts. Those were disabled because the Janvitos/boxwrench drafter GGUFs declared architecture gemma4_assistant (underscore), which stock llama.cpp rejects. The unsloth drafters use the upstream gemma4-assistant (hyphen) spelling that mtp.go's isDraftOnlyAssistantArch expects, so they load without any backend patch (verified by GGUF header inspection).

The 26B-A4B MoE is intentionally omitted — the upstream PR reports no meaningful MTP speedup for it.

Fixed: gemmable-4-12b-mtp

It loaded the draft-only -mtp GGUF as the main model with no draft_model set (a draft-only head cannot run standalone). Now it loads the target as the model, wires the drafter via draft_model, enables spec_type:draft-mtp, and downloads both files.

Verification

• All sha256 pins taken from the HuggingFace API lfs.oid (reliable content hash even for Xet-backed repos).
• Drafter architectures confirmed as gemma4-assistant (hyphen) via GGUF header range-fetch.
• gallery/index.yaml parses; all new entries resolve model/draft/mmproj/files correctly.

Caveats

• Not runtime smoke-tested (no GPU in the authoring environment). Verification covered arch spelling, sha256, schema, and backend option parsing.
• Community notes flag a llama.cpp MTP regression window around builds b9702/b9717 (good at b9553) worth checking against the current pin before relying on this in production.

🤖 Generated with Claude Code